YOLO26 Detection on VisDrone — EdgeFirst Model Zoo
YOLO26 detection models trained on VisDrone2019-DET (10 classes: pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, motor) and validated on its val split (548 images) on real edge hardware with the EdgeFirst Profiler, which runs the pipeline and computes the accuracy on the device itself. Each row in the tables below cites the EdgeFirst Studio validation session (v-XXXX) that produced the measurement.
VisDrone is drone footage: most objects are a few pixels across, so small-object AP (APs) leads every table, followed by the full COCO-style metric set. Every model is published in four input variants that trade time per image for small-object accuracy. For the same family trained on COCO, see EdgeFirst/yolo26-det.
Part of the EdgeFirst Model Zoo. New here? Introducing the EdgeFirst Model Zoo covers what is in the zoo, why every figure links to its validation session, and how to reproduce any of it on your own hardware.
Training experiment: View on EdgeFirst Studio — dataset, training configuration, metrics, and exported artifacts.
Input variants
Each model and size ships in four variants, one folder each. They differ in the input the model sees and in the weights it was trained with, and each one buys small-object AP with time per image:
- 640 is the baseline: one 640×640 pass over the whole frame.
- 1280 shows what a bigger input alone buys: one 1280×1280 pass over the whole frame.
- 1344x768 gives the best result: weights trained on 640×640 tiles run natively on the whole frame at 1344×768, where memory allows.
- tiled (SAHI) slices the frame into 640×640 tiles, runs every tile, and merges the detections. It serves constrained targets, and images larger than VisDrone's ~8 MP, while keeping the benefit of tiled training.
| Variant |
Folder |
Input |
Weights |
Inference |
| 640 |
640/ |
640×640 |
Full-frame |
One pass over the whole frame |
| 1280 |
1280/ |
1280×1280 |
Full-frame |
One pass over the whole frame |
| 1344x768 |
1344x768/ |
1344×768 |
Tiled |
One pass over the whole frame |
| tiled |
tiled/ |
640×640 tiles |
Tiled |
Sliced inference (SAHI): every tile, then merge |
Time per image
Every latency and FPS figure on this card is per image, whatever the image strategy: one 640² pass, one 1280² pass, one 1344×768 pass, or all the 640² tiles of an image plus slicing and merging. The four variants therefore compare directly.
For the tiled variant, time per image is roughly the number of tiles times the time per tile when tiles run one at a time, plus the cost of slicing the frame and merging the detections. Running tiles concurrently helps when the accelerator has spare capacity or when CPU pre- and post-processing can overlap NPU inference; it does not help when one tile already saturates a single NPU, or when memory is the limit (for example the NPU memory pool on the Toradex Verdin i.MX 95). Tiles overlap so that an object cut at a tile border is detected whole in the neighbouring tile; the overlap adds tiles, and therefore time, and produces duplicates that the merge step removes.
Reference accuracy — ONNX FP32
Accuracy ceiling for each variant and size, measured against VisDrone2019-DET val (548 images) with COCO-style metrics. Quantized and compiled artifacts are graded against the reference of the same variant.
| Variant |
Size |
Params |
APs |
AP |
AP50 |
AP75 |
APm |
APl |
AR@100 |
Source |
| 640 |
Nano |
2.7M |
10.32% |
18.38% |
32.52% |
17.55% |
27.10% |
40.61% |
31.53% |
v-1989 |
| 640 |
Small |
10.3M |
13.95% |
22.38% |
38.30% |
22.16% |
32.04% |
46.57% |
36.07% |
v-198f |
| 640 |
Medium |
24.5M |
18.26% |
26.80% |
45.00% |
26.70% |
36.94% |
54.92% |
40.45% |
v-1993 |
| 1280 |
Nano |
2.7M |
20.68% |
29.14% |
47.93% |
29.55% |
39.61% |
50.28% |
43.99% |
v-198b |
| 1280 |
Small |
10.3M |
24.60% |
33.00% |
53.57% |
33.98% |
44.24% |
50.19% |
48.61% |
v-1990 |
| 1280 |
Medium |
24.5M |
28.09% |
36.68% |
58.11% |
38.36% |
48.44% |
54.35% |
52.32% |
v-1994 |
| 1344x768 |
Nano |
2.7M |
21.57% |
30.49% |
49.68% |
31.03% |
42.08% |
51.15% |
45.99% |
v-198e |
| 1344x768 |
Small |
10.3M |
26.50% |
35.23% |
56.16% |
36.58% |
46.78% |
57.02% |
50.61% |
v-1992 |
| 1344x768 |
Medium |
24.5M |
29.38% |
38.18% |
59.95% |
40.30% |
50.12% |
59.02% |
53.27% |
v-1996 |
| tiled |
Nano |
2.7M |
20.35% |
28.77% |
48.33% |
28.89% |
40.09% |
43.56% |
43.74% |
v-198c |
| tiled |
Small |
10.3M |
24.25% |
32.91% |
53.96% |
33.65% |
44.38% |
51.02% |
47.94% |
v-1991 |
| tiled |
Medium |
24.5M |
27.25% |
35.52% |
57.55% |
37.05% |
47.02% |
51.24% |
50.64% |
v-1995 |
On-target validation results
Each row is one EdgeFirst Studio validation session. Click the Source link to inspect the full session — model artifact, dataset version, parameters, per-stage Perfetto trace, and the host hardware description.
Row conventions in the tables below:
- Rows whose
Δ APs cell reads ref are the ONNX FP32 reference runs each quantized or compiled measurement is graded against.
- Δ APs vs FP32 is the small-object AP change against the same variant and size's reference, in percentage points.
- Rows whose
Δ APs cell carries a ⚠ measure more than 10 percentage points of APs below the same variant's float reference: below our expectations for that platform. The numbers are real measurements, reproducible from the linked Studio session, and we are investigating them.
Not supported: … entries, in the short table under each variant's results, are configurations this edition could not produce or run, with the reason: a converter, memory or runtime limitation of that target, not a missing measurement.- Runtime is the inference stack that executed the model; Decode is what the exported graph emits (Smart lifts the operations that quantize worst out of the graph and needs a decoder implementing that cut, the EdgeFirst HAL in EdgeFirst deployments; Logical keeps the head's outputs separate and named;
— means no quantized decode split). Full converter documentation: EdgeFirst model conversion.
- Platform-label suffixes.
(EVK) / (FRDM) / (Phytec) / (Verdin) / (Ezurio) name the development board a session ran on. — latency / — throughput mark the two pipeline configurations of the NXP i.MX 95 Neutron and NXP Ara240 targets: serial inference for the lowest latency, or several inference workers for the highest FPS.
- Time per image (ms) is the end-to-end latency for one image, from the moment the frame is available to the moment its results are ready — preprocess, inference (every tile, for the tiled variant), postprocess and merge. Image acquisition (file load and JPEG decode) and writing results to disk are excluded. Realized FPS is the measured steady-state rate of finished images over the full validation pipeline.
- The full metric set (AP, AP50, AP75, APs, APm, APl and AR at 1, 10 and 100 detections and per object size) for every row is under Full metrics below each table.
640 — 640×640, full-frame weights
| Size |
Platform |
Runtime |
Decode |
APs |
Δ APs vs FP32 (pp) |
AP |
AP50 |
Time per image (ms) |
Realized FPS |
Source |
| Nano |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
10.32% |
ref |
18.38% |
32.52% |
10.00 |
298.9 |
v-1989 |
| Nano |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
10.24% |
-0.08 |
18.31% |
32.47% |
9.02 |
331.8 |
v-19db |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
1.42% |
-8.90 |
6.06% |
14.33% |
328.23 |
4.6 |
v-1c8b |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Combined |
1.42% |
-8.90 |
6.06% |
14.32% |
110.79 |
9.5 |
v-1aa4 |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Logical |
1.42% |
-8.90 |
6.06% |
14.33% |
113.43 |
9.5 |
v-1aac |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Smart |
8.75% |
-1.57 |
16.26% |
29.37% |
120.64 |
9.4 |
v-1ab0 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
9.42% |
-0.90 |
17.36% |
31.35% |
57.96 |
31.9 |
v-17b9 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
9.41% |
-0.91 |
17.36% |
31.35% |
104.64 |
40.4 |
v-1a39 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
9.23% |
-1.09 |
16.98% |
30.44% |
46.57 |
31.6 |
v-1bbd |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
1.52% |
-8.80 |
6.41% |
15.10% |
42.23 |
28.9 |
v-1bbb |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
1.51% |
-8.81 |
6.41% |
15.06% |
42.85 |
28.3 |
v-1bb9 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
9.23% |
-1.09 |
16.98% |
30.44% |
93.59 |
50.9 |
v-1bbe |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
1.52% |
-8.80 |
6.41% |
15.10% |
82.45 |
56.0 |
v-1bbc |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
1.50% |
-8.82 |
6.41% |
15.06% |
82.73 |
55.5 |
v-1bba |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
9.23% |
-1.09 |
16.98% |
30.44% |
89.28 |
53.3 |
v-1aa9 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
9.42% |
-0.90 |
17.36% |
31.35% |
66.70 |
30.6 |
v-17f1 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
9.42% |
-0.90 |
17.36% |
31.35% |
108.77 |
36.4 |
v-1a3d |
| Nano |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
7.44% |
-2.88 |
15.46% |
29.35% |
25.29 |
47.4 |
v-1656 |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
10.37% |
+0.05 |
18.53% |
32.91% |
36.10 |
177.9 |
v-1942 |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
9.29% |
-1.03 |
17.22% |
31.00% |
34.33 |
183.0 |
v-1943 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
1.88% |
-8.44 |
7.72% |
17.46% |
12.40 |
448.9 |
v-16ec |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
9.75% |
-0.57 |
17.52% |
31.53% |
12.02 |
351.3 |
v-16ed |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
10.14% |
-0.18 |
18.18% |
32.33% |
14.79 |
227.6 |
v-16ee |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
9.33% |
-0.99 |
17.08% |
30.61% |
11.84 |
362.6 |
v-1d45 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
1.47% |
-8.85 |
6.83% |
15.71% |
12.20 |
465.2 |
v-1d47 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
10.18% |
-0.14 |
18.22% |
32.32% |
18.44 |
221.7 |
v-1d48 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
10.15% |
-0.17 |
18.19% |
32.35% |
13.18 |
221.0 |
v-1d4a |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
10.18% |
-0.14 |
18.21% |
32.30% |
14.66 |
217.7 |
v-1d49 |
| Small |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
13.95% |
ref |
22.38% |
38.30% |
12.06 |
238.6 |
v-198f |
| Small |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
13.84% |
-0.11 |
22.29% |
38.25% |
9.52 |
264.3 |
v-19e0 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
2.65% |
-11.30 |
9.34% |
20.81% |
448.72 |
4.2 |
v-1c98 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Combined |
2.65% |
-11.30 |
9.34% |
20.80% |
214.92 |
4.8 |
v-1b12 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Logical |
2.65% |
-11.30 |
9.34% |
20.81% |
214.87 |
4.8 |
v-1b1d |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Smart |
11.87% |
-2.08 |
20.25% |
35.47% |
221.45 |
4.8 |
v-1b22 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
12.80% |
-1.15 |
20.97% |
36.58% |
87.78 |
15.6 |
v-17d8 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
12.80% |
-1.15 |
20.98% |
36.58% |
171.55 |
30.0 |
v-1a41 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
12.38% |
-1.57 |
20.54% |
36.00% |
83.63 |
14.0 |
v-1be5 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
2.71% |
-11.24 |
9.33% |
20.84% |
79.33 |
13.7 |
v-1be2 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
2.65% |
-11.30 |
9.26% |
20.73% |
80.17 |
13.5 |
v-1bdf |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
12.38% |
-1.57 |
20.54% |
36.00% |
141.66 |
31.9 |
v-1be7 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
2.71% |
-11.24 |
9.33% |
20.84% |
135.22 |
31.6 |
v-1be4 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
2.65% |
-11.30 |
9.27% |
20.73% |
132.50 |
32.1 |
v-1be1 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
12.38% |
-1.57 |
20.54% |
36.00% |
140.12 |
32.1 |
v-18c4 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
12.80% |
-1.15 |
20.97% |
36.58% |
93.52 |
15.6 |
v-1801 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
12.80% |
-1.15 |
20.98% |
36.58% |
188.65 |
27.8 |
v-1a48 |
| Small |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
5.93% |
-8.02 |
14.33% |
27.78% |
20.77 |
83.4 |
v-1642 |
| Small |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
5.93% |
-8.02 |
14.33% |
27.78% |
18.06 |
139.4 |
v-1986 |
| Small |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
12.13% |
-1.82 |
20.33% |
36.06% |
53.06 |
20.2 |
v-165d |
| Small |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
14.19% |
+0.24 |
22.52% |
38.57% |
42.95 |
128.3 |
v-15e4 |
| Small |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
12.02% |
-1.93 |
20.06% |
36.34% |
37.73 |
170.7 |
v-15e5 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
3.27% |
-10.68 |
10.52% |
22.77% |
13.32 |
330.2 |
v-173f |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
13.27% |
-0.68 |
21.39% |
37.34% |
15.55 |
324.7 |
v-1741 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
13.84% |
-0.11 |
22.28% |
38.18% |
17.99 |
156.0 |
v-1742 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
12.65% |
-1.30 |
20.86% |
36.56% |
13.13 |
332.2 |
v-1d89 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
3.28% |
-10.67 |
10.08% |
22.20% |
10.37 |
350.8 |
v-1d8b |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
13.97% |
+0.02 |
22.33% |
38.20% |
18.73 |
153.5 |
v-1d8c |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
13.84% |
-0.11 |
22.28% |
38.16% |
16.15 |
150.6 |
v-1d8e |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
14.01% |
+0.06 |
22.39% |
38.26% |
18.27 |
150.1 |
v-1d8d |
| Medium |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
18.26% |
ref |
26.80% |
45.00% |
20.21 |
148.8 |
v-1993 |
| Medium |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
18.05% |
-0.21 |
26.67% |
44.83% |
13.23 |
188.3 |
v-19e4 |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
2.13% |
-16.13 |
8.89% |
20.02% |
628.00 |
2.7 |
v-1cae |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Combined |
2.13% |
-16.13 |
8.89% |
20.02% |
367.98 |
2.8 |
v-1c10 |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Logical |
2.13% |
-16.13 |
8.89% |
20.02% |
367.37 |
2.8 |
v-1c18 |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Smart |
14.77% |
-3.49 |
23.18% |
39.72% |
372.27 |
2.8 |
v-1c21 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
16.63% |
-1.63 |
25.21% |
43.23% |
146.76 |
8.1 |
v-17ef |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
16.63% |
-1.63 |
25.21% |
43.23% |
378.49 |
11.6 |
v-1a4d |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
15.76% |
-2.50 |
24.40% |
42.13% |
145.69 |
7.4 |
v-1c2d |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
2.20% |
-16.06 |
9.22% |
20.91% |
142.10 |
7.3 |
v-1c28 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
2.21% |
-16.05 |
9.23% |
20.93% |
143.08 |
7.3 |
v-1c23 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
15.76% |
-2.50 |
24.40% |
42.13% |
361.36 |
11.6 |
v-1c2e |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
2.18% |
-16.08 |
9.12% |
20.74% |
354.61 |
11.4 |
v-1c2a |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
2.21% |
-16.05 |
9.23% |
20.93% |
353.74 |
11.5 |
v-1c26 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
15.76% |
-2.50 |
24.40% |
42.13% |
361.25 |
11.6 |
v-18d7 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
16.63% |
-1.63 |
25.21% |
43.23% |
149.19 |
8.2 |
v-180e |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
16.63% |
-1.63 |
25.21% |
43.23% |
375.85 |
11.7 |
v-1a58 |
| Medium |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
6.57% |
-11.69 ⚠ |
16.00% |
30.90% |
35.10 |
38.8 |
v-164f |
| Medium |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
6.57% |
-11.69 ⚠ |
16.00% |
30.90% |
31.06 |
46.3 |
v-1999 |
| Medium |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
14.27% |
-3.99 |
23.09% |
41.25% |
94.50 |
11.0 |
v-1667 |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
18.56% |
+0.30 |
27.03% |
45.37% |
66.95 |
70.5 |
v-15f7 |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
14.86% |
-3.40 |
23.40% |
41.49% |
47.26 |
104.9 |
v-15f9 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
3.40% |
-14.86 |
10.49% |
23.46% |
13.15 |
208.8 |
v-17a2 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
17.50% |
-0.76 |
26.04% |
44.13% |
15.31 |
206.8 |
v-17a4 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
18.23% |
-0.03 |
26.79% |
45.03% |
28.37 |
91.2 |
v-17a5 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
16.92% |
-1.34 |
25.39% |
43.48% |
16.32 |
213.5 |
v-1dbc |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
2.39% |
-15.87 |
9.32% |
21.93% |
13.26 |
217.1 |
v-1dbd |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
18.21% |
-0.05 |
26.79% |
44.96% |
29.54 |
89.4 |
v-1dbe |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
18.19% |
-0.07 |
26.76% |
45.02% |
27.86 |
89.8 |
v-1dc0 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
18.22% |
-0.04 |
26.78% |
44.96% |
27.67 |
88.0 |
v-1dbf |
| Sizes |
Platform |
Runtime |
Decode |
Status |
| Nano |
NXP Ara240 (FRDM) |
Ara DVM |
Smart |
Not supported: ara2-converter 2.7.3: the DVM's detections do not match ground truth (0.0 AP), reproducible on re-conversion (DE-3020) |
Full metrics — 640
| Size |
Platform |
Runtime |
Decode |
Precision |
APs |
AP |
AP50 |
AP75 |
APm |
APl |
AR@1 |
AR@10 |
AR@100 |
ARs |
ARm |
ARl |
| Nano |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
FP32 |
10.32% |
18.38% |
32.52% |
17.55% |
27.10% |
40.61% |
8.82% |
23.54% |
31.53% |
22.09% |
43.66% |
59.66% |
| Nano |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
FP16 |
10.24% |
18.31% |
32.47% |
17.46% |
27.06% |
40.82% |
8.80% |
23.44% |
31.40% |
21.99% |
43.47% |
59.76% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
INT8 |
1.42% |
6.06% |
14.33% |
4.24% |
9.25% |
29.05% |
4.59% |
11.03% |
15.35% |
6.05% |
27.51% |
47.22% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Combined |
INT8 |
1.42% |
6.06% |
14.32% |
4.24% |
9.23% |
29.03% |
4.56% |
10.97% |
15.25% |
5.98% |
27.37% |
47.08% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Logical |
INT8 |
1.42% |
6.06% |
14.33% |
4.24% |
9.25% |
29.05% |
4.59% |
11.03% |
15.35% |
6.05% |
27.51% |
47.22% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Smart |
INT8 |
8.75% |
16.26% |
29.37% |
15.67% |
24.37% |
37.19% |
8.00% |
21.50% |
28.97% |
19.45% |
41.50% |
54.64% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
INT8 |
9.42% |
17.36% |
31.35% |
16.75% |
25.96% |
42.19% |
8.37% |
22.69% |
30.43% |
20.73% |
43.40% |
61.04% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
INT8 |
9.41% |
17.36% |
31.35% |
16.75% |
25.96% |
42.19% |
8.37% |
22.69% |
30.43% |
20.73% |
43.40% |
61.04% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
INT8 |
9.23% |
16.98% |
30.44% |
16.51% |
25.50% |
39.89% |
8.16% |
21.90% |
29.62% |
20.36% |
42.01% |
57.63% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
INT8 |
1.52% |
6.41% |
15.10% |
4.51% |
9.94% |
30.93% |
4.60% |
11.33% |
15.68% |
6.42% |
28.05% |
52.12% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
INT8 |
1.51% |
6.41% |
15.06% |
4.49% |
9.90% |
31.10% |
4.61% |
11.32% |
15.63% |
6.35% |
27.98% |
51.14% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
INT8 |
9.23% |
16.98% |
30.44% |
16.51% |
25.50% |
39.89% |
8.16% |
21.90% |
29.62% |
20.36% |
42.01% |
57.63% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
INT8 |
1.52% |
6.41% |
15.10% |
4.51% |
9.94% |
30.93% |
4.60% |
11.33% |
15.68% |
6.42% |
28.05% |
52.12% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
INT8 |
1.50% |
6.41% |
15.06% |
4.49% |
9.90% |
31.10% |
4.61% |
11.32% |
15.63% |
6.35% |
27.98% |
51.14% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
INT8 |
9.23% |
16.98% |
30.44% |
16.51% |
25.50% |
39.89% |
8.16% |
21.90% |
29.62% |
20.36% |
42.01% |
57.63% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
INT8 |
9.42% |
17.36% |
31.35% |
16.75% |
25.96% |
42.19% |
8.37% |
22.69% |
30.43% |
20.73% |
43.40% |
61.04% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
INT8 |
9.42% |
17.36% |
31.35% |
16.75% |
25.96% |
42.19% |
8.37% |
22.69% |
30.43% |
20.73% |
43.40% |
61.04% |
| Nano |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
INT8 |
7.44% |
15.46% |
29.35% |
14.15% |
24.16% |
38.97% |
7.98% |
20.83% |
27.74% |
17.62% |
41.70% |
56.30% |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
FP16 |
10.37% |
18.53% |
32.91% |
17.79% |
27.32% |
43.87% |
8.79% |
23.87% |
31.82% |
22.24% |
44.38% |
63.84% |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
INT8 |
9.29% |
17.22% |
31.00% |
16.46% |
25.66% |
41.03% |
8.28% |
22.63% |
30.35% |
20.71% |
43.30% |
58.82% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
INT8 |
1.88% |
7.72% |
17.46% |
5.71% |
12.68% |
33.66% |
5.33% |
12.89% |
17.61% |
7.97% |
30.26% |
52.10% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
INT8 |
9.75% |
17.52% |
31.53% |
16.52% |
25.98% |
39.67% |
8.60% |
22.67% |
30.41% |
21.16% |
42.44% |
56.47% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
INT16 |
10.14% |
18.18% |
32.33% |
17.39% |
26.89% |
41.15% |
8.86% |
23.43% |
31.28% |
21.69% |
43.55% |
59.32% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
INT8 |
9.33% |
17.08% |
30.61% |
16.51% |
25.32% |
39.81% |
8.03% |
22.06% |
29.57% |
20.21% |
41.76% |
60.13% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
INT8 |
1.47% |
6.83% |
15.71% |
5.45% |
11.23% |
30.59% |
4.68% |
11.91% |
16.44% |
6.73% |
29.22% |
52.66% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
INT16 |
10.18% |
18.22% |
32.32% |
17.46% |
27.01% |
40.46% |
8.86% |
23.41% |
31.31% |
21.78% |
43.60% |
59.04% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
INT16 |
10.15% |
18.19% |
32.35% |
17.40% |
26.91% |
40.97% |
8.87% |
23.43% |
31.29% |
21.68% |
43.60% |
59.31% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
INT16 |
10.18% |
18.21% |
32.30% |
17.45% |
27.02% |
40.47% |
8.86% |
23.43% |
31.32% |
21.76% |
43.69% |
59.08% |
| Small |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
FP32 |
13.95% |
22.38% |
38.30% |
22.16% |
32.04% |
46.57% |
9.69% |
27.33% |
36.07% |
27.42% |
47.96% |
65.29% |
| Small |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
FP16 |
13.84% |
22.29% |
38.25% |
22.09% |
32.01% |
46.21% |
9.63% |
27.22% |
35.91% |
27.24% |
47.84% |
65.12% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
INT8 |
2.65% |
9.34% |
20.81% |
7.33% |
15.95% |
38.04% |
5.74% |
14.88% |
20.43% |
10.47% |
34.43% |
54.40% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Combined |
INT8 |
2.65% |
9.34% |
20.80% |
7.33% |
15.95% |
38.05% |
5.73% |
14.88% |
20.44% |
10.47% |
34.44% |
54.40% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Logical |
INT8 |
2.65% |
9.34% |
20.81% |
7.33% |
15.95% |
38.04% |
5.74% |
14.88% |
20.43% |
10.47% |
34.43% |
54.40% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Smart |
INT8 |
11.87% |
20.25% |
35.47% |
19.87% |
30.01% |
43.32% |
9.08% |
25.45% |
34.14% |
25.47% |
46.16% |
60.23% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
INT8 |
12.80% |
20.97% |
36.58% |
20.56% |
30.31% |
48.30% |
9.25% |
26.39% |
35.29% |
26.55% |
47.52% |
65.26% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
INT8 |
12.80% |
20.98% |
36.58% |
20.56% |
30.31% |
48.30% |
9.25% |
26.39% |
35.29% |
26.55% |
47.52% |
65.26% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
INT8 |
12.38% |
20.54% |
36.00% |
19.91% |
29.67% |
45.07% |
9.00% |
26.02% |
34.79% |
26.24% |
46.40% |
63.29% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
INT8 |
2.71% |
9.33% |
20.84% |
7.17% |
15.70% |
38.88% |
5.75% |
15.23% |
20.88% |
10.76% |
35.01% |
56.42% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
INT8 |
2.65% |
9.26% |
20.73% |
7.15% |
15.62% |
38.97% |
5.74% |
15.15% |
20.73% |
10.59% |
34.83% |
56.62% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
INT8 |
12.38% |
20.54% |
36.00% |
19.91% |
29.67% |
45.07% |
9.00% |
26.02% |
34.79% |
26.24% |
46.40% |
63.29% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
INT8 |
2.71% |
9.33% |
20.84% |
7.17% |
15.70% |
38.89% |
5.75% |
15.23% |
20.88% |
10.76% |
35.01% |
56.42% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
INT8 |
2.65% |
9.27% |
20.73% |
7.15% |
15.62% |
38.97% |
5.74% |
15.15% |
20.73% |
10.59% |
34.83% |
56.62% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
INT8 |
12.38% |
20.54% |
36.00% |
19.91% |
29.67% |
45.07% |
9.00% |
26.02% |
34.79% |
26.24% |
46.40% |
63.29% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
INT8 |
12.80% |
20.97% |
36.58% |
20.56% |
30.31% |
48.30% |
9.25% |
26.39% |
35.29% |
26.55% |
47.52% |
65.26% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
INT8 |
12.80% |
20.98% |
36.58% |
20.56% |
30.31% |
48.30% |
9.25% |
26.39% |
35.29% |
26.55% |
47.52% |
65.26% |
| Small |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
INT16 |
5.93% |
14.33% |
27.78% |
13.31% |
23.78% |
41.87% |
7.47% |
19.79% |
26.98% |
17.79% |
39.27% |
56.34% |
| Small |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
INT16 |
5.93% |
14.33% |
27.78% |
13.31% |
23.77% |
41.87% |
7.47% |
19.79% |
26.98% |
17.79% |
39.27% |
56.34% |
| Small |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
INT8 |
12.13% |
20.33% |
36.06% |
20.10% |
30.02% |
43.21% |
9.05% |
25.30% |
33.66% |
25.67% |
45.18% |
58.95% |
| Small |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
FP16 |
14.19% |
22.52% |
38.57% |
22.35% |
32.18% |
48.72% |
9.75% |
27.54% |
36.33% |
27.52% |
48.74% |
64.92% |
| Small |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
INT8 |
12.02% |
20.06% |
36.34% |
19.42% |
30.11% |
49.92% |
8.99% |
25.62% |
34.29% |
25.26% |
46.71% |
62.74% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
INT8 |
3.27% |
10.52% |
22.77% |
8.57% |
17.90% |
39.78% |
6.26% |
16.18% |
21.94% |
11.72% |
36.58% |
55.59% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
INT8 |
13.27% |
21.39% |
37.34% |
21.13% |
30.81% |
46.63% |
9.36% |
26.57% |
35.16% |
26.49% |
47.01% |
64.14% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
INT16 |
13.84% |
22.28% |
38.18% |
21.99% |
31.99% |
46.10% |
9.67% |
27.02% |
35.73% |
27.10% |
47.56% |
62.25% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
INT8 |
12.65% |
20.86% |
36.56% |
20.40% |
30.40% |
44.81% |
9.02% |
26.00% |
34.63% |
26.14% |
46.71% |
60.83% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
INT8 |
3.28% |
10.08% |
22.20% |
7.72% |
16.98% |
40.42% |
5.92% |
15.82% |
21.51% |
11.62% |
35.78% |
55.20% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
INT16 |
13.97% |
22.33% |
38.20% |
22.00% |
32.04% |
46.89% |
9.71% |
27.18% |
35.88% |
27.22% |
47.78% |
63.83% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
INT16 |
13.84% |
22.28% |
38.16% |
21.99% |
32.00% |
46.12% |
9.67% |
27.02% |
35.73% |
27.09% |
47.58% |
62.25% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
INT16 |
14.01% |
22.39% |
38.26% |
22.12% |
32.08% |
46.82% |
9.72% |
27.23% |
35.93% |
27.28% |
47.81% |
63.47% |
| Medium |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
FP32 |
18.26% |
26.80% |
45.00% |
26.70% |
36.94% |
54.92% |
10.98% |
30.88% |
40.45% |
32.29% |
51.78% |
68.85% |
| Medium |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
FP16 |
18.05% |
26.67% |
44.83% |
26.58% |
36.85% |
54.84% |
10.93% |
30.77% |
40.28% |
32.04% |
51.67% |
69.60% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
INT8 |
2.13% |
8.89% |
20.02% |
6.82% |
14.90% |
35.29% |
5.64% |
14.00% |
18.82% |
8.41% |
32.95% |
51.14% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Combined |
INT8 |
2.13% |
8.89% |
20.02% |
6.82% |
14.90% |
35.24% |
5.65% |
13.98% |
18.79% |
8.40% |
32.89% |
51.06% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Logical |
INT8 |
2.13% |
8.89% |
20.02% |
6.82% |
14.90% |
35.29% |
5.64% |
14.00% |
18.82% |
8.41% |
32.95% |
51.14% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Smart |
INT8 |
14.77% |
23.18% |
39.72% |
23.25% |
33.12% |
44.10% |
9.93% |
27.56% |
36.44% |
28.10% |
47.87% |
58.58% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
INT8 |
16.63% |
25.21% |
43.23% |
25.00% |
35.59% |
55.52% |
10.48% |
29.88% |
39.40% |
30.94% |
51.23% |
73.50% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
INT8 |
16.63% |
25.21% |
43.23% |
25.00% |
35.59% |
55.52% |
10.48% |
29.88% |
39.40% |
30.94% |
51.23% |
73.50% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
INT8 |
15.76% |
24.40% |
42.13% |
24.34% |
34.58% |
51.39% |
10.07% |
28.73% |
38.13% |
29.83% |
49.55% |
67.42% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
INT8 |
2.20% |
9.22% |
20.91% |
7.05% |
15.43% |
42.14% |
5.58% |
14.68% |
19.71% |
9.26% |
33.83% |
59.31% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
INT8 |
2.21% |
9.23% |
20.93% |
7.04% |
15.43% |
42.04% |
5.59% |
14.68% |
19.68% |
9.26% |
33.78% |
59.18% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
INT8 |
15.76% |
24.40% |
42.13% |
24.34% |
34.58% |
51.40% |
10.07% |
28.73% |
38.13% |
29.83% |
49.55% |
67.42% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
INT8 |
2.18% |
9.12% |
20.74% |
6.95% |
15.35% |
41.44% |
5.57% |
14.66% |
19.68% |
9.25% |
33.80% |
59.19% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
INT8 |
2.21% |
9.23% |
20.93% |
7.04% |
15.43% |
42.04% |
5.59% |
14.68% |
19.68% |
9.26% |
33.78% |
59.18% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
INT8 |
15.76% |
24.40% |
42.13% |
24.34% |
34.58% |
51.39% |
10.07% |
28.73% |
38.13% |
29.83% |
49.55% |
67.42% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
INT8 |
16.63% |
25.21% |
43.23% |
25.00% |
35.59% |
55.52% |
10.48% |
29.88% |
39.40% |
30.94% |
51.23% |
73.50% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
INT8 |
16.63% |
25.21% |
43.23% |
25.00% |
35.59% |
55.52% |
10.48% |
29.88% |
39.40% |
30.94% |
51.23% |
73.50% |
| Medium |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
INT16 |
6.57% |
16.00% |
30.90% |
14.87% |
26.03% |
48.40% |
8.12% |
20.99% |
28.56% |
19.02% |
42.18% |
60.68% |
| Medium |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
INT16 |
6.57% |
16.00% |
30.90% |
14.87% |
26.03% |
48.40% |
8.12% |
20.99% |
28.56% |
19.02% |
42.18% |
60.68% |
| Medium |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
INT8 |
14.27% |
23.09% |
41.25% |
22.37% |
33.26% |
52.34% |
10.04% |
27.83% |
36.67% |
28.01% |
49.02% |
65.89% |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
FP16 |
18.56% |
27.03% |
45.37% |
26.83% |
37.58% |
57.27% |
10.95% |
31.22% |
41.07% |
32.65% |
53.02% |
74.46% |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
INT8 |
14.86% |
23.40% |
41.49% |
22.91% |
34.25% |
49.60% |
9.88% |
28.82% |
38.62% |
30.11% |
50.63% |
67.59% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
INT8 |
3.40% |
10.49% |
23.46% |
8.06% |
16.79% |
46.25% |
6.41% |
16.46% |
22.41% |
12.07% |
36.84% |
58.63% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
INT8 |
17.50% |
26.04% |
44.13% |
26.13% |
36.33% |
54.33% |
10.60% |
30.11% |
39.71% |
31.39% |
51.14% |
68.59% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
INT16 |
18.23% |
26.79% |
45.03% |
26.72% |
36.93% |
53.53% |
10.86% |
30.78% |
40.42% |
32.24% |
51.71% |
68.08% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
INT8 |
16.92% |
25.39% |
43.48% |
25.35% |
35.66% |
51.59% |
10.65% |
29.64% |
39.00% |
30.64% |
50.51% |
66.03% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
INT8 |
2.39% |
9.32% |
21.93% |
6.65% |
15.67% |
41.02% |
5.83% |
14.93% |
19.95% |
9.69% |
34.18% |
54.80% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
INT16 |
18.21% |
26.79% |
44.96% |
26.71% |
36.94% |
53.80% |
10.82% |
30.79% |
40.37% |
32.13% |
51.70% |
68.07% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
INT16 |
18.19% |
26.76% |
45.02% |
26.72% |
36.94% |
53.33% |
10.86% |
30.74% |
40.38% |
32.19% |
51.71% |
67.92% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
INT16 |
18.22% |
26.78% |
44.96% |
26.64% |
36.91% |
53.81% |
10.83% |
30.78% |
40.36% |
32.13% |
51.68% |
68.08% |
1280 — 1280×1280, full-frame weights
| Size |
Platform |
Runtime |
Decode |
APs |
Δ APs vs FP32 (pp) |
AP |
AP50 |
Time per image (ms) |
Realized FPS |
Source |
| Nano |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
20.68% |
ref |
29.14% |
47.93% |
22.28 |
136.0 |
v-198b |
| Nano |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
20.59% |
-0.09 |
29.01% |
47.78% |
16.31 |
177.5 |
v-19dc |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
5.00% |
-15.68 |
12.70% |
26.96% |
1739.41 |
1.1 |
v-1c8f |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Combined |
5.00% |
-15.68 |
12.70% |
26.96% |
858.65 |
1.2 |
v-1ab5 |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Logical |
5.00% |
-15.68 |
12.70% |
26.96% |
858.50 |
1.2 |
v-1abe |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Smart |
17.67% |
-3.01 |
25.74% |
43.81% |
871.15 |
1.2 |
v-1ac7 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
18.62% |
-2.06 |
27.10% |
45.53% |
305.45 |
3.9 |
v-17bb |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
18.62% |
-2.06 |
27.10% |
45.53% |
423.78 |
11.9 |
v-1a3a |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
18.71% |
-1.97 |
26.90% |
45.44% |
323.15 |
3.4 |
v-1bc9 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
5.40% |
-15.28 |
13.90% |
28.08% |
322.63 |
3.3 |
v-1bc3 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
5.40% |
-15.28 |
13.90% |
28.08% |
315.99 |
3.3 |
v-1bbf |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
18.71% |
-1.97 |
26.90% |
45.44% |
358.51 |
12.6 |
v-1bcb |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
5.40% |
-15.28 |
13.90% |
28.08% |
336.96 |
12.7 |
v-1bc7 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
5.40% |
-15.28 |
13.90% |
28.08% |
339.05 |
12.7 |
v-1bc2 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
18.71% |
-1.97 |
26.90% |
45.44% |
353.74 |
12.9 |
v-1aaa |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
18.62% |
-2.06 |
27.10% |
45.53% |
313.08 |
3.8 |
v-17f5 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
18.62% |
-2.06 |
27.10% |
45.53% |
452.47 |
11.2 |
v-1a3f |
| Nano |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
17.00% |
-3.68 |
25.77% |
44.82% |
80.21 |
13.9 |
v-1659 |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
20.61% |
-0.07 |
29.09% |
47.99% |
77.77 |
70.7 |
v-1944 |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
18.87% |
-1.81 |
26.93% |
45.37% |
69.06 |
88.6 |
v-1946 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
6.49% |
-14.19 |
15.65% |
30.31% |
21.77 |
128.2 |
v-16fe |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
19.91% |
-0.77 |
28.39% |
47.07% |
26.44 |
128.2 |
v-1700 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
20.54% |
-0.14 |
29.15% |
48.05% |
52.45 |
51.2 |
v-1701 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
19.33% |
-1.35 |
27.47% |
46.11% |
26.40 |
128.6 |
v-1d4b |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
5.49% |
-15.19 |
13.99% |
28.55% |
20.88 |
127.7 |
v-1d4c |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
20.60% |
-0.08 |
29.20% |
48.02% |
52.53 |
51.1 |
v-1d4d |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
20.59% |
-0.09 |
29.17% |
48.06% |
48.84 |
49.6 |
v-1d4f |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
20.62% |
-0.06 |
29.21% |
48.03% |
48.70 |
49.2 |
v-1d4e |
| Small |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
24.60% |
ref |
33.00% |
53.57% |
42.83 |
101.6 |
v-1990 |
| Small |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
24.25% |
-0.35 |
32.78% |
53.36% |
22.52 |
142.3 |
v-19e1 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
7.14% |
-17.46 |
15.47% |
31.46% |
2616.42 |
0.6 |
v-1c9c |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Combined |
7.14% |
-17.46 |
15.47% |
31.46% |
1705.04 |
0.6 |
v-1b26 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Logical |
7.14% |
-17.46 |
15.47% |
31.46% |
1704.79 |
0.6 |
v-1b41 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Smart |
21.78% |
-2.82 |
29.65% |
49.86% |
1716.34 |
0.6 |
v-1b5a |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
22.52% |
-2.08 |
30.69% |
51.00% |
592.72 |
1.8 |
v-17dc |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
22.52% |
-2.08 |
30.69% |
51.00% |
734.23 |
6.2 |
v-1a42 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
22.42% |
-2.18 |
30.42% |
50.73% |
675.72 |
1.5 |
v-1bfb |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
7.49% |
-17.11 |
16.62% |
32.92% |
675.30 |
1.5 |
v-1bf2 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
7.49% |
-17.11 |
16.62% |
32.92% |
667.53 |
1.5 |
v-1bea |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
22.42% |
-2.18 |
30.42% |
50.73% |
664.83 |
6.4 |
v-1c01 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
7.49% |
-17.11 |
16.62% |
32.92% |
651.21 |
6.3 |
v-1bf8 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
7.49% |
-17.11 |
16.62% |
32.92% |
655.38 |
6.3 |
v-1bf1 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
22.42% |
-2.18 |
30.42% |
50.73% |
661.71 |
6.4 |
v-18c5 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
22.52% |
-2.08 |
30.69% |
51.00% |
602.78 |
1.8 |
v-1803 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
22.52% |
-2.08 |
30.69% |
51.00% |
770.58 |
6.0 |
v-1a49 |
| Small |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
21.36% |
-3.24 |
30.05% |
50.21% |
160.17 |
6.8 |
v-165e |
| Small |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
24.33% |
-0.27 |
32.85% |
53.46% |
120.63 |
39.6 |
v-15e7 |
| Small |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
22.08% |
-2.52 |
29.27% |
50.34% |
86.37 |
59.9 |
v-15ea |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
7.73% |
-16.87 |
17.33% |
33.55% |
33.80 |
72.0 |
v-178b |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
23.88% |
-0.72 |
32.16% |
52.82% |
39.78 |
72.5 |
v-178c |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
24.61% |
+0.01 |
33.02% |
53.56% |
72.27 |
33.8 |
v-178f |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
22.19% |
-2.41 |
30.53% |
50.88% |
39.39 |
73.5 |
v-1d8f |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
7.31% |
-17.29 |
16.86% |
33.84% |
34.26 |
73.7 |
v-1d90 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
24.63% |
+0.03 |
33.01% |
53.56% |
72.28 |
33.7 |
v-1d91 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
24.66% |
+0.06 |
33.04% |
53.56% |
67.95 |
33.6 |
v-1d93 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
24.64% |
+0.04 |
33.02% |
53.55% |
68.62 |
33.1 |
v-1d92 |
| Medium |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
28.09% |
ref |
36.68% |
58.11% |
84.50 |
53.3 |
v-1994 |
| Medium |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
27.75% |
-0.34 |
36.41% |
57.95% |
56.47 |
86.3 |
v-19e5 |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
5.61% |
-22.48 |
13.51% |
30.02% |
3368.34 |
0.4 |
v-1cb6 |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Combined |
5.61% |
-22.48 |
13.51% |
30.02% |
2385.55 |
0.4 |
v-1c2c |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Logical |
5.61% |
-22.48 |
13.51% |
30.02% |
2385.82 |
0.4 |
v-1c3e |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Smart |
24.96% |
-3.13 |
33.54% |
54.46% |
2400.53 |
0.4 |
v-1c4e |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
26.37% |
-1.72 |
34.85% |
56.55% |
839.14 |
1.3 |
v-1832 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
26.37% |
-1.72 |
34.86% |
56.55% |
1565.95 |
2.7 |
v-1a50 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
26.36% |
-1.73 |
34.79% |
56.57% |
932.26 |
1.1 |
v-1c3d |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
6.12% |
-21.97 |
14.40% |
31.62% |
920.56 |
1.1 |
v-1c37 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
6.12% |
-21.97 |
14.40% |
31.62% |
910.47 |
1.1 |
v-1c2f |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
26.36% |
-1.73 |
34.79% |
56.57% |
1525.02 |
2.7 |
v-1c42 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
6.12% |
-21.97 |
14.40% |
31.62% |
1508.99 |
2.7 |
v-1c3b |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
6.12% |
-21.97 |
14.40% |
31.62% |
1507.63 |
2.7 |
v-1c36 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
26.36% |
-1.73 |
34.79% |
56.57% |
1525.79 |
2.7 |
v-18d9 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
26.37% |
-1.72 |
34.85% |
56.55% |
845.90 |
1.3 |
v-1833 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
26.37% |
-1.72 |
34.85% |
56.55% |
1588.45 |
2.7 |
v-1a5b |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
27.80% |
-0.29 |
36.49% |
58.01% |
208.92 |
20.8 |
v-15fb |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
25.30% |
-2.79 |
33.10% |
55.17% |
130.03 |
34.9 |
v-15fe |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
8.02% |
-20.07 |
17.51% |
34.73% |
50.68 |
45.6 |
v-17d0 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
27.27% |
-0.82 |
35.77% |
57.42% |
56.06 |
45.6 |
v-17d1 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
28.01% |
-0.08 |
36.67% |
57.97% |
116.89 |
19.2 |
v-17d2 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
25.49% |
-2.60 |
33.71% |
55.12% |
55.05 |
47.4 |
v-1dc1 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
9.13% |
-18.96 |
18.51% |
37.66% |
49.71 |
47.1 |
v-1dc2 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
28.04% |
-0.05 |
36.64% |
58.00% |
118.64 |
18.9 |
v-1dc3 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
27.97% |
-0.12 |
36.64% |
57.97% |
112.78 |
19.1 |
v-1dc5 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
28.02% |
-0.07 |
36.62% |
58.03% |
114.89 |
18.7 |
v-1dc4 |
| Sizes |
Platform |
Runtime |
Decode |
Status |
| Nano, Small, Medium |
NXP Ara240 (FRDM) |
Ara DVM |
Smart |
Not supported: ara2-converter 2.7.3: ARA-240 area attention spans at most 1024 positions; 1280x1280 spans 1600 (DE-2996) |
| Medium |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
Not supported: Hailo-8L: insufficient on-chip memory for this model and input size |
Full metrics — 1280
| Size |
Platform |
Runtime |
Decode |
Precision |
APs |
AP |
AP50 |
AP75 |
APm |
APl |
AR@1 |
AR@10 |
AR@100 |
ARs |
ARm |
ARl |
| Nano |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
FP32 |
20.68% |
29.14% |
47.93% |
29.55% |
39.61% |
50.28% |
11.64% |
33.24% |
43.99% |
36.18% |
54.96% |
68.11% |
| Nano |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
FP16 |
20.59% |
29.01% |
47.78% |
29.38% |
39.46% |
50.38% |
11.63% |
33.09% |
43.84% |
36.09% |
54.75% |
68.03% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
INT8 |
5.00% |
12.70% |
26.96% |
10.77% |
22.19% |
40.08% |
7.11% |
19.48% |
27.29% |
17.50% |
41.37% |
55.60% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Combined |
INT8 |
5.00% |
12.70% |
26.96% |
10.77% |
22.19% |
40.08% |
7.11% |
19.48% |
27.29% |
17.50% |
41.37% |
55.60% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Logical |
INT8 |
5.00% |
12.70% |
26.96% |
10.77% |
22.19% |
40.08% |
7.11% |
19.48% |
27.29% |
17.50% |
41.37% |
55.60% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Smart |
INT8 |
17.67% |
25.74% |
43.81% |
25.95% |
36.39% |
44.84% |
10.51% |
31.04% |
41.23% |
33.24% |
52.39% |
61.47% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
INT8 |
18.62% |
27.10% |
45.53% |
27.54% |
37.56% |
46.91% |
10.83% |
31.69% |
42.13% |
34.34% |
53.35% |
61.52% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
INT8 |
18.62% |
27.10% |
45.53% |
27.54% |
37.56% |
46.91% |
10.83% |
31.69% |
42.13% |
34.34% |
53.35% |
61.52% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
INT8 |
18.71% |
26.90% |
45.44% |
27.21% |
37.13% |
45.18% |
10.83% |
31.66% |
42.06% |
34.41% |
53.10% |
60.61% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
INT8 |
5.40% |
13.90% |
28.08% |
12.33% |
23.01% |
41.35% |
7.67% |
20.17% |
28.08% |
18.43% |
42.18% |
54.45% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
INT8 |
5.40% |
13.90% |
28.08% |
12.33% |
23.01% |
41.35% |
7.67% |
20.17% |
28.08% |
18.43% |
42.18% |
54.45% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
INT8 |
18.71% |
26.90% |
45.44% |
27.21% |
37.13% |
45.18% |
10.83% |
31.66% |
42.06% |
34.41% |
53.10% |
60.61% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
INT8 |
5.40% |
13.90% |
28.08% |
12.33% |
23.01% |
41.35% |
7.67% |
20.17% |
28.08% |
18.43% |
42.18% |
54.45% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
INT8 |
5.40% |
13.90% |
28.08% |
12.33% |
23.01% |
41.35% |
7.67% |
20.17% |
28.08% |
18.43% |
42.18% |
54.45% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
INT8 |
18.71% |
26.90% |
45.44% |
27.21% |
37.13% |
45.18% |
10.83% |
31.66% |
42.06% |
34.41% |
53.10% |
60.61% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
INT8 |
18.62% |
27.10% |
45.53% |
27.54% |
37.56% |
46.91% |
10.83% |
31.69% |
42.13% |
34.34% |
53.35% |
61.52% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
INT8 |
18.62% |
27.10% |
45.53% |
27.54% |
37.56% |
46.91% |
10.83% |
31.69% |
42.13% |
34.34% |
53.35% |
61.52% |
| Nano |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
INT8 |
17.00% |
25.77% |
44.82% |
25.50% |
36.56% |
49.24% |
10.71% |
30.22% |
40.21% |
31.33% |
52.53% |
67.85% |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
FP16 |
20.61% |
29.09% |
47.99% |
29.47% |
39.58% |
50.90% |
11.60% |
33.14% |
43.86% |
36.12% |
54.79% |
68.21% |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
INT8 |
18.87% |
26.93% |
45.37% |
27.52% |
37.19% |
47.90% |
10.68% |
31.57% |
41.95% |
34.31% |
52.96% |
63.86% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
INT8 |
6.49% |
15.65% |
30.31% |
14.38% |
25.29% |
44.93% |
8.22% |
21.65% |
30.08% |
20.59% |
44.12% |
63.13% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
INT8 |
19.91% |
28.39% |
47.07% |
28.96% |
38.65% |
51.05% |
11.33% |
32.65% |
43.36% |
35.40% |
54.52% |
66.69% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
INT16 |
20.54% |
29.15% |
48.05% |
29.59% |
39.76% |
50.81% |
11.62% |
33.33% |
44.04% |
36.19% |
55.14% |
64.87% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
INT8 |
19.33% |
27.47% |
46.11% |
28.01% |
37.59% |
47.03% |
11.13% |
31.88% |
42.45% |
34.92% |
53.32% |
64.95% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
INT8 |
5.49% |
13.99% |
28.55% |
11.99% |
22.77% |
44.00% |
7.61% |
20.13% |
28.01% |
18.81% |
41.77% |
58.62% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
INT16 |
20.60% |
29.20% |
48.02% |
29.68% |
39.91% |
50.96% |
11.66% |
33.41% |
44.15% |
36.28% |
55.19% |
65.20% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
INT16 |
20.59% |
29.17% |
48.06% |
29.58% |
39.75% |
50.90% |
11.62% |
33.37% |
44.08% |
36.26% |
55.13% |
65.67% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
INT16 |
20.62% |
29.21% |
48.03% |
29.65% |
39.89% |
50.88% |
11.67% |
33.44% |
44.17% |
36.32% |
55.21% |
65.16% |
| Small |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
FP32 |
24.60% |
33.00% |
53.57% |
33.98% |
44.24% |
50.19% |
12.62% |
36.93% |
48.61% |
41.08% |
59.74% |
69.02% |
| Small |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
FP16 |
24.25% |
32.78% |
53.36% |
33.62% |
44.16% |
50.25% |
12.60% |
36.72% |
48.29% |
40.59% |
59.65% |
68.91% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
INT8 |
7.14% |
15.47% |
31.46% |
13.58% |
25.86% |
41.45% |
8.41% |
22.66% |
31.68% |
21.92% |
46.82% |
59.94% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Combined |
INT8 |
7.14% |
15.47% |
31.46% |
13.58% |
25.86% |
41.45% |
8.41% |
22.66% |
31.68% |
21.92% |
46.82% |
59.94% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Logical |
INT8 |
7.14% |
15.47% |
31.46% |
13.58% |
25.86% |
41.45% |
8.41% |
22.66% |
31.68% |
21.92% |
46.82% |
59.94% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Smart |
INT8 |
21.78% |
29.65% |
49.86% |
30.34% |
40.41% |
44.08% |
11.62% |
34.68% |
45.92% |
38.40% |
57.36% |
64.35% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
INT8 |
22.52% |
30.69% |
51.00% |
31.47% |
41.43% |
47.05% |
12.04% |
35.54% |
47.09% |
39.80% |
58.31% |
66.96% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
INT8 |
22.52% |
30.69% |
51.00% |
31.47% |
41.43% |
47.05% |
12.04% |
35.54% |
47.09% |
39.80% |
58.31% |
66.96% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
INT8 |
22.42% |
30.42% |
50.73% |
31.11% |
40.83% |
45.50% |
11.98% |
35.23% |
46.81% |
39.43% |
57.87% |
65.68% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
INT8 |
7.49% |
16.62% |
32.92% |
15.19% |
26.79% |
41.23% |
8.67% |
23.28% |
32.31% |
22.49% |
47.31% |
58.33% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
INT8 |
7.49% |
16.62% |
32.92% |
15.19% |
26.79% |
41.23% |
8.67% |
23.28% |
32.31% |
22.49% |
47.31% |
58.33% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
INT8 |
22.42% |
30.42% |
50.73% |
31.12% |
40.83% |
45.50% |
11.98% |
35.23% |
46.81% |
39.43% |
57.87% |
65.68% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
INT8 |
7.49% |
16.62% |
32.92% |
15.19% |
26.79% |
41.23% |
8.67% |
23.28% |
32.31% |
22.49% |
47.31% |
58.33% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
INT8 |
7.49% |
16.62% |
32.92% |
15.19% |
26.79% |
41.23% |
8.67% |
23.28% |
32.31% |
22.49% |
47.31% |
58.33% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
INT8 |
22.42% |
30.42% |
50.73% |
31.11% |
40.83% |
45.50% |
11.98% |
35.23% |
46.81% |
39.43% |
57.87% |
65.68% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
INT8 |
22.52% |
30.69% |
51.00% |
31.47% |
41.43% |
47.05% |
12.04% |
35.54% |
47.09% |
39.80% |
58.31% |
66.96% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
INT8 |
22.52% |
30.69% |
51.00% |
31.47% |
41.43% |
47.05% |
12.04% |
35.54% |
47.09% |
39.80% |
58.31% |
66.96% |
| Small |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
INT8 |
21.36% |
30.05% |
50.21% |
30.96% |
41.88% |
51.24% |
12.27% |
34.20% |
45.50% |
37.79% |
56.90% |
66.81% |
| Small |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
FP16 |
24.33% |
32.85% |
53.46% |
33.77% |
44.39% |
50.46% |
12.61% |
36.81% |
48.44% |
40.74% |
60.02% |
68.92% |
| Small |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
INT8 |
22.08% |
29.27% |
50.34% |
29.53% |
40.11% |
48.24% |
11.74% |
34.64% |
46.38% |
39.35% |
57.12% |
64.90% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
INT8 |
7.73% |
17.33% |
33.55% |
16.00% |
28.42% |
44.32% |
8.85% |
23.78% |
33.05% |
23.43% |
47.72% |
60.84% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
INT8 |
23.88% |
32.16% |
52.82% |
33.12% |
43.15% |
49.50% |
12.42% |
36.29% |
47.82% |
40.61% |
58.33% |
67.64% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
INT16 |
24.61% |
33.02% |
53.56% |
33.90% |
44.22% |
49.99% |
12.73% |
36.94% |
48.73% |
41.33% |
59.95% |
65.38% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
INT8 |
22.19% |
30.53% |
50.88% |
31.32% |
41.47% |
48.08% |
11.99% |
35.34% |
46.77% |
39.41% |
57.76% |
63.43% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
INT8 |
7.31% |
16.86% |
33.84% |
14.59% |
27.82% |
43.28% |
8.51% |
23.28% |
31.93% |
22.09% |
46.59% |
57.75% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
INT16 |
24.63% |
33.01% |
53.56% |
33.87% |
44.21% |
50.19% |
12.71% |
36.80% |
48.61% |
41.20% |
59.77% |
66.02% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
INT16 |
24.66% |
33.04% |
53.56% |
33.89% |
44.23% |
49.97% |
12.73% |
36.96% |
48.75% |
41.35% |
59.93% |
65.38% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
INT16 |
24.64% |
33.02% |
53.55% |
33.91% |
44.21% |
50.18% |
12.72% |
36.84% |
48.61% |
41.22% |
59.73% |
65.89% |
| Medium |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
FP32 |
28.09% |
36.68% |
58.11% |
38.36% |
48.44% |
54.35% |
13.66% |
40.08% |
52.32% |
44.65% |
64.05% |
72.68% |
| Medium |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
FP16 |
27.75% |
36.41% |
57.95% |
38.07% |
48.31% |
54.15% |
13.61% |
39.82% |
52.00% |
44.32% |
63.76% |
71.83% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
INT8 |
5.61% |
13.51% |
30.02% |
10.32% |
24.38% |
41.34% |
7.44% |
21.16% |
29.54% |
19.03% |
45.76% |
56.54% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Combined |
INT8 |
5.61% |
13.51% |
30.02% |
10.32% |
24.38% |
41.34% |
7.44% |
21.16% |
29.54% |
19.03% |
45.76% |
56.54% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Logical |
INT8 |
5.61% |
13.51% |
30.02% |
10.32% |
24.38% |
41.34% |
7.44% |
21.16% |
29.54% |
19.03% |
45.76% |
56.54% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Smart |
INT8 |
24.96% |
33.54% |
54.46% |
35.38% |
45.78% |
47.40% |
12.56% |
37.63% |
49.48% |
41.54% |
61.59% |
67.40% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
INT8 |
26.37% |
34.85% |
56.55% |
36.68% |
46.69% |
53.85% |
13.17% |
38.89% |
51.02% |
43.57% |
62.71% |
69.92% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
INT8 |
26.37% |
34.86% |
56.55% |
36.69% |
46.69% |
53.85% |
13.17% |
38.89% |
51.02% |
43.57% |
62.71% |
69.92% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
INT8 |
26.36% |
34.79% |
56.57% |
36.52% |
46.45% |
52.98% |
13.03% |
38.69% |
50.68% |
43.32% |
62.15% |
69.64% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
INT8 |
6.12% |
14.40% |
31.62% |
11.15% |
24.44% |
48.50% |
7.89% |
21.80% |
30.35% |
20.05% |
46.36% |
61.58% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
INT8 |
6.12% |
14.40% |
31.62% |
11.15% |
24.44% |
48.50% |
7.89% |
21.80% |
30.35% |
20.05% |
46.36% |
61.58% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
INT8 |
26.36% |
34.79% |
56.57% |
36.52% |
46.45% |
52.97% |
13.03% |
38.69% |
50.68% |
43.32% |
62.15% |
69.64% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
INT8 |
6.12% |
14.40% |
31.62% |
11.15% |
24.44% |
48.50% |
7.89% |
21.80% |
30.35% |
20.05% |
46.36% |
61.58% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
INT8 |
6.12% |
14.40% |
31.62% |
11.15% |
24.44% |
48.50% |
7.89% |
21.80% |
30.35% |
20.05% |
46.36% |
61.58% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
INT8 |
26.36% |
34.79% |
56.57% |
36.52% |
46.44% |
52.99% |
13.03% |
38.69% |
50.68% |
43.32% |
62.15% |
69.64% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
INT8 |
26.37% |
34.85% |
56.55% |
36.68% |
46.69% |
53.85% |
13.17% |
38.89% |
51.02% |
43.57% |
62.71% |
69.92% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
INT8 |
26.37% |
34.85% |
56.55% |
36.68% |
46.69% |
53.84% |
13.17% |
38.89% |
51.02% |
43.57% |
62.71% |
69.92% |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
FP16 |
27.80% |
36.49% |
58.01% |
38.07% |
48.53% |
54.89% |
13.66% |
39.84% |
52.03% |
44.26% |
63.91% |
71.92% |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
INT8 |
25.30% |
33.10% |
55.17% |
34.53% |
44.62% |
52.43% |
12.61% |
37.78% |
49.98% |
42.79% |
61.47% |
65.38% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
INT8 |
8.02% |
17.51% |
34.73% |
15.75% |
29.89% |
50.54% |
9.10% |
24.57% |
34.31% |
24.35% |
50.15% |
64.79% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
INT8 |
27.27% |
35.77% |
57.42% |
37.21% |
47.44% |
55.37% |
13.37% |
39.13% |
51.29% |
43.71% |
62.86% |
71.50% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
INT16 |
28.01% |
36.67% |
57.97% |
38.58% |
48.61% |
54.75% |
13.69% |
40.14% |
52.40% |
44.53% |
64.37% |
72.32% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
INT8 |
25.49% |
33.71% |
55.12% |
35.21% |
45.01% |
50.11% |
12.86% |
38.46% |
50.71% |
42.73% |
63.03% |
71.43% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
INT8 |
9.13% |
18.51% |
37.66% |
16.34% |
30.84% |
47.05% |
9.21% |
25.18% |
34.47% |
24.14% |
50.73% |
65.32% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
INT16 |
28.04% |
36.64% |
58.00% |
38.46% |
48.53% |
55.02% |
13.69% |
40.03% |
52.36% |
44.53% |
64.32% |
72.47% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
INT16 |
27.97% |
36.64% |
57.97% |
38.56% |
48.57% |
54.89% |
13.70% |
40.12% |
52.36% |
44.47% |
64.34% |
72.37% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
INT16 |
28.02% |
36.62% |
58.03% |
38.44% |
48.51% |
54.93% |
13.69% |
40.04% |
52.35% |
44.50% |
64.33% |
72.44% |
1344x768 — 1344×768, tiled weights
| Size |
Platform |
Runtime |
Decode |
APs |
Δ APs vs FP32 (pp) |
AP |
AP50 |
Time per image (ms) |
Realized FPS |
Source |
| Nano |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
21.57% |
ref |
30.49% |
49.68% |
17.55 |
175.7 |
v-198e |
| Nano |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
21.39% |
-0.18 |
30.35% |
49.56% |
13.30 |
234.9 |
v-19df |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
5.98% |
-15.59 |
13.47% |
26.01% |
980.04 |
1.8 |
v-1c95 |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Combined |
5.98% |
-15.59 |
13.47% |
26.01% |
416.59 |
2.5 |
v-1af3 |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Logical |
5.98% |
-15.59 |
13.47% |
26.01% |
416.52 |
2.5 |
v-1af9 |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Smart |
15.10% |
-6.47 |
22.70% |
38.29% |
429.15 |
2.5 |
v-1b05 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
19.94% |
-1.63 |
28.67% |
47.72% |
167.43 |
8.2 |
v-17d4 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
19.94% |
-1.63 |
28.67% |
47.72% |
281.99 |
19.5 |
v-1a40 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
19.92% |
-1.65 |
28.64% |
47.82% |
152.63 |
7.7 |
v-1bdc |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
7.71% |
-13.86 |
17.12% |
32.60% |
146.16 |
7.5 |
v-1bda |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
7.76% |
-13.81 |
17.15% |
32.64% |
149.99 |
7.2 |
v-1bd7 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
19.92% |
-1.65 |
28.64% |
47.82% |
205.42 |
23.1 |
v-1bdd |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
7.71% |
-13.86 |
17.12% |
32.60% |
188.74 |
23.5 |
v-1bdb |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
7.76% |
-13.81 |
17.15% |
32.64% |
188.55 |
23.5 |
v-1bd8 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
19.92% |
-1.65 |
28.64% |
47.82% |
202.50 |
23.6 |
v-1aaf |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
19.94% |
-1.63 |
28.67% |
47.72% |
182.86 |
8.1 |
v-17fe |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
19.94% |
-1.63 |
28.67% |
47.72% |
295.15 |
18.0 |
v-1a47 |
| Nano |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
11.61% |
-9.96 |
21.05% |
37.72% |
35.40 |
52.1 |
v-17b6 |
| Nano |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
11.61% |
-9.96 |
21.05% |
37.72% |
31.28 |
112.2 |
v-1983 |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
21.38% |
-0.19 |
30.33% |
49.60% |
68.49 |
101.2 |
v-1949 |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
19.61% |
-1.96 |
28.24% |
47.02% |
63.15 |
115.5 |
v-194a |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
8.48% |
-13.09 |
18.17% |
34.08% |
18.22 |
160.2 |
v-1725 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
20.87% |
-0.70 |
29.53% |
48.92% |
22.55 |
202.0 |
v-1727 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
21.08% |
-0.49 |
29.92% |
49.06% |
35.95 |
80.2 |
v-1728 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
20.34% |
-1.23 |
28.97% |
48.25% |
20.97 |
207.9 |
v-1d65 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
7.98% |
-13.59 |
17.59% |
33.27% |
20.16 |
209.2 |
v-1d66 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
21.23% |
-0.34 |
30.00% |
49.13% |
36.72 |
79.7 |
v-1d67 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
21.09% |
-0.48 |
29.93% |
49.06% |
31.85 |
81.1 |
v-1d69 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
21.22% |
-0.35 |
29.99% |
49.12% |
32.70 |
80.7 |
v-1d68 |
| Small |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
26.50% |
ref |
35.23% |
56.16% |
23.78 |
138.2 |
v-1992 |
| Small |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
26.24% |
-0.26 |
35.01% |
56.09% |
15.54 |
189.9 |
v-19e3 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
9.32% |
-17.18 |
19.32% |
37.50% |
1442.72 |
1.2 |
v-1cab |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Combined |
9.32% |
-17.18 |
19.32% |
37.49% |
819.69 |
1.2 |
v-1be6 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Logical |
9.32% |
-17.18 |
19.32% |
37.49% |
819.47 |
1.2 |
v-1bf5 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Smart |
23.34% |
-3.16 |
32.22% |
52.52% |
829.58 |
1.2 |
v-1c02 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
23.74% |
-2.76 |
32.25% |
53.02% |
294.07 |
4.0 |
v-17e8 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
23.74% |
-2.76 |
32.25% |
53.02% |
435.16 |
11.3 |
v-1a45 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
23.80% |
-2.70 |
32.12% |
53.08% |
320.58 |
3.3 |
v-1c20 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
9.54% |
-16.96 |
19.40% |
37.59% |
319.01 |
3.3 |
v-1c1b |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
9.54% |
-16.96 |
19.42% |
37.62% |
314.72 |
3.3 |
v-1c16 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
23.80% |
-2.70 |
32.12% |
53.08% |
379.80 |
11.5 |
v-1c22 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
9.54% |
-16.96 |
19.40% |
37.58% |
370.75 |
11.3 |
v-1c1d |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
9.54% |
-16.96 |
19.42% |
37.62% |
370.42 |
11.4 |
v-1c19 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
23.80% |
-2.70 |
32.12% |
53.08% |
380.28 |
11.5 |
v-18d5 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
23.74% |
-2.76 |
32.25% |
53.02% |
298.61 |
3.9 |
v-180b |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
23.74% |
-2.76 |
32.25% |
53.02% |
456.30 |
10.9 |
v-1a4e |
| Small |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
12.84% |
-13.66 ⚠ |
23.06% |
41.50% |
59.17 |
24.8 |
v-1649 |
| Small |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
12.84% |
-13.66 ⚠ |
23.06% |
41.50% |
50.14 |
30.9 |
v-198d |
| Small |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
26.24% |
-0.26 |
34.97% |
55.98% |
89.14 |
60.4 |
v-15f2 |
| Small |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
24.13% |
-2.37 |
30.13% |
51.77% |
67.78 |
87.1 |
v-15f4 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
10.05% |
-16.45 |
20.44% |
38.19% |
24.55 |
107.2 |
v-179b |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
25.47% |
-1.03 |
34.07% |
55.29% |
27.70 |
107.8 |
v-179c |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
26.57% |
+0.07 |
35.40% |
56.39% |
50.90 |
50.7 |
v-179d |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
24.41% |
-2.09 |
33.30% |
54.32% |
27.59 |
109.1 |
v-1da3 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
9.24% |
-17.26 |
19.73% |
37.61% |
24.37 |
108.5 |
v-1da4 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
26.61% |
+0.11 |
35.41% |
56.42% |
50.72 |
50.6 |
v-1da5 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
26.55% |
+0.05 |
35.38% |
56.38% |
47.82 |
49.7 |
v-1dac |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
26.59% |
+0.09 |
35.39% |
56.44% |
48.01 |
49.6 |
v-1de9 |
| Medium |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
29.38% |
ref |
38.18% |
59.95% |
56.90 |
80.5 |
v-1996 |
| Medium |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
29.07% |
-0.31 |
37.91% |
59.81% |
33.06 |
125.5 |
v-19e8 |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
5.89% |
-23.49 |
12.62% |
28.77% |
1924.12 |
0.8 |
v-1cc2 |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Combined |
5.87% |
-23.51 |
12.56% |
28.65% |
1246.82 |
0.8 |
v-1c74 |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Logical |
5.91% |
-23.47 |
12.61% |
28.75% |
1246.78 |
0.8 |
v-1c76 |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Smart |
25.00% |
-4.38 |
33.96% |
55.33% |
1258.01 |
0.8 |
v-1c7b |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
26.42% |
-2.96 |
35.58% |
57.79% |
446.56 |
2.5 |
v-1816 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
26.42% |
-2.96 |
35.59% |
57.80% |
972.31 |
4.3 |
v-1a5d |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
26.53% |
-2.85 |
35.49% |
57.97% |
470.13 |
2.2 |
v-1c61 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
6.18% |
-23.20 |
13.33% |
29.94% |
474.16 |
2.2 |
v-1c5c |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
6.17% |
-23.21 |
13.32% |
29.90% |
471.89 |
2.2 |
v-1c58 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
26.55% |
-2.83 |
35.49% |
57.97% |
943.74 |
4.3 |
v-1c64 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
6.19% |
-23.19 |
13.34% |
30.00% |
936.17 |
4.3 |
v-1c5f |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
6.19% |
-23.19 |
13.35% |
30.00% |
937.68 |
4.3 |
v-1c5b |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
26.54% |
-2.84 |
35.49% |
57.97% |
943.95 |
4.3 |
v-18e3 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
26.40% |
-2.98 |
35.58% |
57.79% |
451.50 |
2.5 |
v-181f |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
26.40% |
-2.98 |
35.58% |
57.79% |
980.99 |
4.3 |
v-1a65 |
| Medium |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
11.55% |
-17.83 ⚠ |
20.59% |
39.98% |
129.76 |
9.1 |
v-1665 |
| Medium |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
11.55% |
-17.83 ⚠ |
20.58% |
40.04% |
121.94 |
9.8 |
v-19a3 |
| Medium |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
25.45% |
-3.93 |
34.95% |
57.27% |
242.25 |
4.3 |
v-167b |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
29.02% |
-0.36 |
37.91% |
59.84% |
146.64 |
32.1 |
v-1608 |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
23.73% |
-5.65 |
33.03% |
55.51% |
95.25 |
51.7 |
v-160b |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
9.06% |
-20.32 |
19.17% |
37.86% |
34.02 |
72.0 |
v-17ea |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
28.33% |
-1.05 |
37.21% |
59.31% |
38.66 |
70.5 |
v-17eb |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
29.37% |
-0.01 |
38.20% |
59.99% |
81.56 |
28.3 |
v-17ec |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
27.77% |
-1.61 |
36.43% |
58.54% |
38.22 |
74.0 |
v-1de4 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
7.71% |
-21.67 |
16.72% |
35.43% |
34.38 |
73.1 |
v-1de5 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
29.41% |
+0.03 |
38.20% |
59.91% |
82.92 |
28.0 |
v-1de6 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
29.42% |
+0.04 |
38.23% |
60.00% |
78.37 |
28.1 |
v-1de8 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
29.41% |
+0.03 |
38.24% |
59.91% |
80.14 |
27.6 |
v-1de7 |
| Sizes |
Platform |
Runtime |
Decode |
Status |
| Nano, Small |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
Not supported: Hailo-8L: insufficient on-chip memory for this model and input size |
Full metrics — 1344x768
| Size |
Platform |
Runtime |
Decode |
Precision |
APs |
AP |
AP50 |
AP75 |
APm |
APl |
AR@1 |
AR@10 |
AR@100 |
ARs |
ARm |
ARl |
| Nano |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
FP32 |
21.57% |
30.49% |
49.68% |
31.03% |
42.08% |
51.15% |
11.95% |
34.85% |
45.99% |
37.24% |
58.60% |
69.90% |
| Nano |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
FP16 |
21.39% |
30.35% |
49.56% |
30.84% |
42.09% |
51.17% |
11.93% |
34.70% |
45.82% |
37.08% |
58.51% |
70.03% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
INT8 |
5.98% |
13.47% |
26.01% |
12.59% |
22.90% |
30.93% |
7.44% |
20.34% |
28.23% |
19.07% |
41.23% |
44.63% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Combined |
INT8 |
5.98% |
13.47% |
26.01% |
12.59% |
22.90% |
30.93% |
7.44% |
20.34% |
28.23% |
19.07% |
41.23% |
44.63% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Logical |
INT8 |
5.98% |
13.47% |
26.01% |
12.59% |
22.90% |
30.93% |
7.44% |
20.34% |
28.23% |
19.07% |
41.23% |
44.63% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Smart |
INT8 |
15.10% |
22.70% |
38.29% |
23.26% |
32.91% |
32.83% |
9.93% |
28.74% |
38.30% |
29.91% |
49.95% |
48.66% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
INT8 |
19.94% |
28.67% |
47.72% |
29.30% |
39.93% |
49.28% |
11.62% |
33.18% |
43.91% |
35.45% |
56.12% |
67.38% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
INT8 |
19.94% |
28.67% |
47.72% |
29.30% |
39.93% |
49.28% |
11.62% |
33.18% |
43.91% |
35.45% |
56.12% |
67.38% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
INT8 |
19.92% |
28.64% |
47.82% |
29.21% |
40.01% |
46.23% |
11.52% |
33.11% |
43.83% |
35.34% |
56.10% |
62.47% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
INT8 |
7.71% |
17.12% |
32.60% |
16.27% |
28.40% |
42.82% |
8.75% |
23.59% |
32.47% |
22.61% |
47.28% |
58.23% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
INT8 |
7.76% |
17.15% |
32.64% |
16.32% |
28.37% |
42.74% |
8.75% |
23.65% |
32.49% |
22.59% |
47.38% |
58.19% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
INT8 |
19.92% |
28.64% |
47.82% |
29.21% |
40.01% |
46.23% |
11.52% |
33.11% |
43.83% |
35.34% |
56.10% |
62.47% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
INT8 |
7.71% |
17.12% |
32.60% |
16.27% |
28.40% |
42.82% |
8.75% |
23.59% |
32.47% |
22.61% |
47.28% |
58.23% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
INT8 |
7.76% |
17.15% |
32.64% |
16.32% |
28.37% |
42.74% |
8.75% |
23.65% |
32.49% |
22.59% |
47.38% |
58.19% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
INT8 |
19.92% |
28.64% |
47.82% |
29.21% |
40.01% |
46.23% |
11.52% |
33.11% |
43.83% |
35.34% |
56.10% |
62.47% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
INT8 |
19.94% |
28.67% |
47.72% |
29.30% |
39.93% |
49.28% |
11.62% |
33.18% |
43.91% |
35.45% |
56.12% |
67.38% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
INT8 |
19.94% |
28.67% |
47.72% |
29.30% |
39.92% |
49.28% |
11.62% |
33.18% |
43.91% |
35.45% |
56.12% |
67.38% |
| Nano |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
INT16 |
11.61% |
21.05% |
37.72% |
21.03% |
33.35% |
42.70% |
9.86% |
26.24% |
36.35% |
28.49% |
46.95% |
56.79% |
| Nano |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
INT16 |
11.61% |
21.05% |
37.72% |
21.03% |
33.34% |
42.70% |
9.86% |
26.24% |
36.35% |
28.49% |
46.95% |
56.79% |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
FP16 |
21.38% |
30.33% |
49.60% |
30.75% |
42.00% |
50.82% |
11.95% |
34.71% |
45.82% |
37.08% |
58.48% |
69.82% |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
INT8 |
19.61% |
28.24% |
47.02% |
29.01% |
39.61% |
48.38% |
11.32% |
32.99% |
43.83% |
35.41% |
56.25% |
69.18% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
INT8 |
8.48% |
18.17% |
34.08% |
17.46% |
29.47% |
46.72% |
9.17% |
24.78% |
33.94% |
23.98% |
48.91% |
65.52% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
INT8 |
20.87% |
29.53% |
48.92% |
29.96% |
40.89% |
49.59% |
11.79% |
33.95% |
44.76% |
36.38% |
56.85% |
67.29% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
INT16 |
21.08% |
29.92% |
49.06% |
30.45% |
41.48% |
48.58% |
11.91% |
34.43% |
45.34% |
36.79% |
57.58% |
67.07% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
INT8 |
20.34% |
28.97% |
48.25% |
29.32% |
40.45% |
48.91% |
11.65% |
33.50% |
44.30% |
35.89% |
56.81% |
66.89% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
INT8 |
7.98% |
17.59% |
33.27% |
16.50% |
29.15% |
45.58% |
8.76% |
24.10% |
33.21% |
23.51% |
48.12% |
63.21% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
INT16 |
21.23% |
30.00% |
49.13% |
30.64% |
41.40% |
48.76% |
11.92% |
34.38% |
45.33% |
36.93% |
57.26% |
67.07% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
INT16 |
21.09% |
29.93% |
49.06% |
30.48% |
41.48% |
48.59% |
11.92% |
34.45% |
45.36% |
36.81% |
57.59% |
67.08% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
INT16 |
21.22% |
29.99% |
49.12% |
30.63% |
41.36% |
48.91% |
11.92% |
34.37% |
45.32% |
36.93% |
57.20% |
67.18% |
| Small |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
FP32 |
26.50% |
35.23% |
56.16% |
36.58% |
46.78% |
57.02% |
13.28% |
38.53% |
50.61% |
42.82% |
62.20% |
70.89% |
| Small |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
FP16 |
26.24% |
35.01% |
56.09% |
36.47% |
46.60% |
56.21% |
13.17% |
38.30% |
50.28% |
42.50% |
61.89% |
70.08% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
INT8 |
9.32% |
19.32% |
37.50% |
17.94% |
31.26% |
44.96% |
9.24% |
25.47% |
34.65% |
24.39% |
50.57% |
62.47% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Combined |
INT8 |
9.32% |
19.32% |
37.49% |
17.94% |
31.26% |
44.96% |
9.24% |
25.47% |
34.65% |
24.39% |
50.57% |
62.47% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Logical |
INT8 |
9.32% |
19.32% |
37.49% |
17.94% |
31.25% |
44.96% |
9.24% |
25.47% |
34.65% |
24.39% |
50.57% |
62.47% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Smart |
INT8 |
23.34% |
32.22% |
52.52% |
33.89% |
44.29% |
48.88% |
12.27% |
36.23% |
47.81% |
39.62% |
60.18% |
66.77% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
INT8 |
23.74% |
32.25% |
53.02% |
33.48% |
43.71% |
52.29% |
12.34% |
36.59% |
48.53% |
40.64% |
60.26% |
66.65% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
INT8 |
23.74% |
32.25% |
53.02% |
33.48% |
43.71% |
52.29% |
12.34% |
36.59% |
48.53% |
40.64% |
60.26% |
66.65% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
INT8 |
23.80% |
32.12% |
53.08% |
33.60% |
43.51% |
50.20% |
12.30% |
36.39% |
48.33% |
40.57% |
59.91% |
66.14% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
INT8 |
9.54% |
19.40% |
37.59% |
17.87% |
30.75% |
47.21% |
9.45% |
25.87% |
35.31% |
25.31% |
50.81% |
62.10% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
INT8 |
9.54% |
19.42% |
37.62% |
17.91% |
30.79% |
47.24% |
9.46% |
25.87% |
35.31% |
25.29% |
50.81% |
62.11% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
INT8 |
23.80% |
32.12% |
53.08% |
33.60% |
43.51% |
50.20% |
12.30% |
36.39% |
48.33% |
40.57% |
59.91% |
66.14% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
INT8 |
9.54% |
19.40% |
37.58% |
17.87% |
30.74% |
47.21% |
9.45% |
25.87% |
35.31% |
25.31% |
50.81% |
62.10% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
INT8 |
9.54% |
19.42% |
37.62% |
17.91% |
30.79% |
47.24% |
9.46% |
25.87% |
35.31% |
25.29% |
50.81% |
62.11% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
INT8 |
23.80% |
32.12% |
53.08% |
33.60% |
43.51% |
50.20% |
12.30% |
36.39% |
48.33% |
40.57% |
59.91% |
66.14% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
INT8 |
23.74% |
32.25% |
53.02% |
33.48% |
43.71% |
52.29% |
12.34% |
36.59% |
48.53% |
40.64% |
60.26% |
66.65% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
INT8 |
23.74% |
32.25% |
53.02% |
33.48% |
43.71% |
52.29% |
12.34% |
36.59% |
48.53% |
40.64% |
60.26% |
66.65% |
| Small |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
INT16 |
12.84% |
23.06% |
41.50% |
22.89% |
35.92% |
49.04% |
10.41% |
28.40% |
39.16% |
30.75% |
51.24% |
62.90% |
| Small |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
INT16 |
12.84% |
23.06% |
41.50% |
22.89% |
35.92% |
49.04% |
10.41% |
28.40% |
39.16% |
30.75% |
51.24% |
62.90% |
| Small |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
FP16 |
26.24% |
34.97% |
55.98% |
36.30% |
46.43% |
56.47% |
13.14% |
38.34% |
50.25% |
42.51% |
61.77% |
69.85% |
| Small |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
INT8 |
24.13% |
30.13% |
51.77% |
30.68% |
39.88% |
53.51% |
11.68% |
35.49% |
47.38% |
40.67% |
58.01% |
64.59% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
INT8 |
10.05% |
20.44% |
38.19% |
19.38% |
32.30% |
51.79% |
9.78% |
26.59% |
36.29% |
26.61% |
51.28% |
65.70% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
INT8 |
25.47% |
34.07% |
55.29% |
35.45% |
45.32% |
54.66% |
12.90% |
37.67% |
49.52% |
41.99% |
60.91% |
69.45% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
INT16 |
26.57% |
35.40% |
56.39% |
36.53% |
47.06% |
55.01% |
13.21% |
38.84% |
50.96% |
43.08% |
62.61% |
69.88% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
INT8 |
24.41% |
33.30% |
54.32% |
34.44% |
44.97% |
55.57% |
12.63% |
37.19% |
49.14% |
41.15% |
61.23% |
70.62% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
INT8 |
9.24% |
19.73% |
37.61% |
18.67% |
32.21% |
51.78% |
9.45% |
25.96% |
35.46% |
25.55% |
50.77% |
64.14% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
INT16 |
26.61% |
35.41% |
56.42% |
36.56% |
47.06% |
54.81% |
13.24% |
38.94% |
51.02% |
43.19% |
62.67% |
69.84% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
INT16 |
26.55% |
35.38% |
56.38% |
36.50% |
47.03% |
55.01% |
13.22% |
38.82% |
50.94% |
43.06% |
62.59% |
69.86% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
INT16 |
26.59% |
35.39% |
56.44% |
36.54% |
47.05% |
54.77% |
13.24% |
38.92% |
51.01% |
43.18% |
62.67% |
69.84% |
| Medium |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
FP32 |
29.38% |
38.18% |
59.95% |
40.30% |
50.12% |
59.02% |
14.40% |
40.84% |
53.27% |
45.37% |
64.99% |
73.62% |
| Medium |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
FP16 |
29.07% |
37.91% |
59.81% |
39.88% |
49.99% |
58.86% |
14.31% |
40.55% |
52.90% |
44.92% |
64.74% |
73.55% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
INT8 |
5.89% |
12.62% |
28.77% |
9.08% |
23.62% |
45.53% |
7.45% |
21.36% |
30.48% |
19.79% |
46.65% |
60.28% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Combined |
INT8 |
5.87% |
12.56% |
28.65% |
9.01% |
23.60% |
45.52% |
7.45% |
21.35% |
30.46% |
19.78% |
46.58% |
60.28% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Logical |
INT8 |
5.91% |
12.61% |
28.75% |
9.05% |
23.63% |
45.53% |
7.45% |
21.36% |
30.48% |
19.79% |
46.65% |
60.28% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Smart |
INT8 |
25.00% |
33.96% |
55.33% |
35.17% |
45.86% |
52.08% |
12.99% |
37.76% |
49.73% |
41.79% |
61.32% |
69.41% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
INT8 |
26.42% |
35.58% |
57.79% |
36.92% |
47.71% |
56.83% |
13.51% |
39.25% |
51.90% |
44.37% |
63.01% |
72.11% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
INT8 |
26.42% |
35.59% |
57.80% |
36.92% |
47.71% |
56.83% |
13.51% |
39.25% |
51.90% |
44.37% |
63.01% |
72.11% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
INT8 |
26.53% |
35.49% |
57.97% |
36.76% |
47.44% |
56.69% |
13.44% |
39.16% |
51.73% |
44.09% |
63.05% |
72.59% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
INT8 |
6.18% |
13.33% |
29.94% |
9.69% |
25.04% |
50.03% |
7.92% |
22.38% |
31.94% |
21.10% |
48.79% |
62.60% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
INT8 |
6.17% |
13.32% |
29.90% |
9.70% |
25.06% |
50.04% |
7.92% |
22.38% |
31.94% |
21.10% |
48.78% |
62.74% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
INT8 |
26.55% |
35.49% |
57.97% |
36.76% |
47.44% |
56.70% |
13.44% |
39.16% |
51.73% |
44.09% |
63.05% |
72.59% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
INT8 |
6.19% |
13.34% |
30.00% |
9.72% |
25.06% |
50.02% |
7.92% |
22.38% |
31.94% |
21.10% |
48.79% |
62.60% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
INT8 |
6.19% |
13.35% |
30.00% |
9.72% |
25.06% |
50.04% |
7.92% |
22.38% |
31.94% |
21.10% |
48.78% |
62.74% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
INT8 |
26.54% |
35.49% |
57.97% |
36.76% |
47.45% |
56.69% |
13.44% |
39.16% |
51.73% |
44.09% |
63.05% |
72.59% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
INT8 |
26.40% |
35.58% |
57.79% |
36.91% |
47.71% |
56.83% |
13.51% |
39.25% |
51.90% |
44.37% |
63.01% |
72.11% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
INT8 |
26.40% |
35.58% |
57.79% |
36.92% |
47.72% |
56.83% |
13.51% |
39.25% |
51.90% |
44.37% |
63.01% |
72.11% |
| Medium |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
INT16 |
11.55% |
20.59% |
39.98% |
18.87% |
36.07% |
51.14% |
9.63% |
28.54% |
40.66% |
31.63% |
54.00% |
64.64% |
| Medium |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
INT16 |
11.55% |
20.58% |
40.04% |
18.82% |
36.05% |
51.14% |
9.63% |
28.54% |
40.66% |
31.63% |
54.00% |
64.64% |
| Medium |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
INT8 |
25.45% |
34.95% |
57.27% |
36.35% |
48.26% |
56.38% |
13.66% |
37.94% |
49.30% |
40.39% |
62.83% |
72.42% |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
FP16 |
29.02% |
37.91% |
59.84% |
40.02% |
50.02% |
58.67% |
14.29% |
40.63% |
52.91% |
44.98% |
64.67% |
74.10% |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
INT8 |
23.73% |
33.03% |
55.51% |
34.64% |
45.29% |
53.45% |
12.84% |
37.09% |
49.12% |
41.06% |
61.06% |
70.56% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
INT8 |
9.06% |
19.17% |
37.86% |
17.31% |
31.86% |
52.75% |
9.79% |
25.98% |
35.60% |
25.25% |
51.62% |
65.68% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
INT8 |
28.33% |
37.21% |
59.31% |
39.18% |
49.15% |
57.99% |
14.02% |
39.97% |
52.27% |
44.68% |
63.46% |
74.24% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
INT16 |
29.37% |
38.20% |
59.99% |
40.12% |
50.39% |
58.35% |
14.33% |
40.88% |
53.27% |
45.26% |
65.20% |
74.13% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
INT8 |
27.77% |
36.43% |
58.54% |
38.40% |
48.20% |
56.06% |
13.95% |
39.61% |
51.79% |
43.92% |
63.52% |
72.95% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
INT8 |
7.71% |
16.72% |
35.43% |
13.53% |
28.98% |
49.06% |
9.06% |
24.21% |
33.18% |
22.39% |
49.70% |
64.29% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
INT16 |
29.41% |
38.20% |
59.91% |
40.16% |
50.34% |
58.68% |
14.40% |
40.93% |
53.35% |
45.35% |
65.26% |
74.63% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
INT16 |
29.42% |
38.23% |
60.00% |
40.14% |
50.39% |
58.36% |
14.34% |
40.89% |
53.30% |
45.31% |
65.18% |
74.13% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
INT16 |
29.41% |
38.24% |
59.91% |
40.21% |
50.38% |
58.68% |
14.40% |
40.96% |
53.40% |
45.40% |
65.31% |
74.63% |
tiled — 640×640 tiles, tiled weights
| Size |
Platform |
Runtime |
Decode |
APs |
Δ APs vs FP32 (pp) |
AP |
AP50 |
Time per image (ms) |
Realized FPS |
Source |
| Nano |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
20.35% |
ref |
28.77% |
48.33% |
37.05 |
55.5 |
v-198c |
| Nano |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
20.29% |
-0.06 |
28.71% |
48.25% |
32.25 |
69.1 |
v-19dd |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
9.71% |
-10.64 |
17.63% |
35.95% |
1674.14 |
0.9 |
v-1c92 |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Combined |
9.71% |
-10.64 |
17.63% |
35.95% |
565.30 |
1.9 |
v-1ad5 |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Logical |
9.71% |
-10.64 |
17.63% |
35.95% |
565.43 |
1.9 |
v-1ade |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Smart |
15.70% |
-4.65 |
23.07% |
40.08% |
596.54 |
1.8 |
v-1aea |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
19.04% |
-1.31 |
27.41% |
46.85% |
214.41 |
6.5 |
v-17ce |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
19.04% |
-1.31 |
27.41% |
46.85% |
317.31 |
10.5 |
v-1a3e |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
19.22% |
-1.13 |
27.45% |
46.84% |
220.86 |
6.0 |
v-1bd4 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
11.95% |
-8.40 |
21.00% |
42.27% |
211.86 |
5.5 |
v-1bd0 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
11.84% |
-8.51 |
20.96% |
42.20% |
216.47 |
5.3 |
v-1bcc |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
19.22% |
-1.13 |
27.45% |
46.84% |
325.80 |
10.4 |
v-1bd6 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
11.95% |
-8.40 |
21.00% |
42.27% |
316.18 |
11.4 |
v-1bd2 |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
11.84% |
-8.51 |
20.96% |
42.21% |
315.00 |
11.6 |
v-1bcd |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
19.22% |
-1.13 |
27.45% |
46.84% |
310.82 |
10.6 |
v-1aad |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
19.04% |
-1.31 |
27.41% |
46.85% |
236.74 |
6.5 |
v-17fa |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
19.04% |
-1.31 |
27.41% |
46.85% |
339.11 |
9.6 |
v-1a46 |
| Nano |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
16.43% |
-3.92 |
25.13% |
45.44% |
79.68 |
22.4 |
v-173e |
| Nano |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
16.43% |
-3.92 |
25.14% |
45.43% |
66.43 |
56.9 |
v-197f |
| Nano |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
17.63% |
-2.72 |
26.19% |
46.59% |
129.10 |
9.1 |
v-165b |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
20.24% |
-0.11 |
28.67% |
48.25% |
128.02 |
45.8 |
v-1947 |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
18.17% |
-2.18 |
26.28% |
45.73% |
119.61 |
52.7 |
v-1948 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
12.56% |
-7.79 |
21.37% |
42.92% |
36.63 |
104.5 |
v-1708 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
19.39% |
-0.96 |
27.56% |
47.14% |
49.42 |
88.1 |
v-170a |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
19.98% |
-0.37 |
28.43% |
47.93% |
82.20 |
43.8 |
v-170b |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
18.94% |
-1.41 |
26.97% |
46.74% |
50.20 |
90.7 |
v-1d50 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
11.88% |
-8.47 |
20.94% |
42.28% |
37.40 |
103.7 |
v-1d51 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
20.08% |
-0.27 |
28.53% |
47.98% |
92.42 |
43.4 |
v-1d53 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
19.99% |
-0.36 |
28.44% |
47.93% |
68.42 |
43.1 |
v-1d55 |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
20.10% |
-0.25 |
28.55% |
48.00% |
70.79 |
42.6 |
v-1d54 |
| Small |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
24.25% |
ref |
32.91% |
53.96% |
48.80 |
46.1 |
v-1991 |
| Small |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
24.19% |
-0.06 |
32.87% |
53.92% |
38.23 |
58.6 |
v-19e2 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
11.29% |
-12.96 |
20.74% |
43.52% |
2296.56 |
0.8 |
v-1ca6 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Combined |
11.29% |
-12.96 |
20.74% |
43.52% |
1091.45 |
0.9 |
v-1bb3 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Logical |
11.29% |
-12.96 |
20.73% |
43.52% |
1092.98 |
0.9 |
v-1bc4 |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Smart |
21.00% |
-3.25 |
29.17% |
50.38% |
1123.50 |
0.9 |
v-1bd3 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
22.15% |
-2.10 |
30.36% |
51.72% |
395.40 |
3.0 |
v-17e2 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
22.15% |
-2.10 |
30.36% |
51.72% |
636.26 |
6.1 |
v-1a43 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
22.47% |
-1.78 |
30.46% |
51.85% |
425.28 |
2.7 |
v-1c12 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
12.40% |
-11.85 |
21.80% |
45.46% |
414.93 |
2.6 |
v-1c0a |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
12.34% |
-11.91 |
21.77% |
45.40% |
420.73 |
2.5 |
v-1c05 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
22.47% |
-1.78 |
30.45% |
51.85% |
638.24 |
6.1 |
v-1c15 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
12.40% |
-11.85 |
21.79% |
45.46% |
649.65 |
5.9 |
v-1c0e |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
12.34% |
-11.91 |
21.76% |
45.40% |
629.63 |
6.2 |
v-1c08 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
22.47% |
-1.78 |
30.46% |
51.85% |
633.19 |
6.1 |
v-18d1 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
22.15% |
-2.10 |
30.36% |
51.72% |
412.73 |
3.0 |
v-1807 |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
22.15% |
-2.10 |
30.36% |
51.72% |
644.39 |
6.0 |
v-1a4a |
| Small |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
17.95% |
-6.30 |
26.61% |
49.80% |
103.48 |
16.2 |
v-1646 |
| Small |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
17.93% |
-6.32 |
26.62% |
49.81% |
82.31 |
27.0 |
v-1987 |
| Small |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
21.50% |
-2.75 |
30.17% |
52.28% |
272.43 |
3.9 |
v-1662 |
| Small |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
24.25% |
+0.00 |
32.82% |
53.89% |
175.63 |
27.9 |
v-15eb |
| Small |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
21.16% |
-3.09 |
26.61% |
48.14% |
137.20 |
37.9 |
v-15ec |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
15.02% |
-9.23 |
24.80% |
48.55% |
51.45 |
67.5 |
v-1769 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
23.50% |
-0.75 |
31.83% |
53.26% |
59.65 |
69.1 |
v-176b |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
24.27% |
+0.02 |
32.97% |
53.95% |
89.24 |
30.5 |
v-176c |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
22.25% |
-2.00 |
30.50% |
51.78% |
57.30 |
69.4 |
v-1d94 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
13.38% |
-10.87 |
23.11% |
46.62% |
55.15 |
67.3 |
v-1d95 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
24.25% |
+0.00 |
32.95% |
53.89% |
87.52 |
30.2 |
v-1d96 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
24.28% |
+0.03 |
32.97% |
53.95% |
85.87 |
29.4 |
v-1d98 |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
24.25% |
+0.00 |
32.94% |
53.87% |
81.29 |
29.5 |
v-1d97 |
| Medium |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
27.25% |
ref |
35.52% |
57.55% |
97.37 |
32.0 |
v-1995 |
| Medium |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
27.24% |
-0.01 |
35.47% |
57.50% |
58.03 |
42.2 |
v-19e7 |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
11.57% |
-15.68 |
20.39% |
44.52% |
3210.35 |
0.5 |
v-1cbe |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Combined |
11.57% |
-15.68 |
20.37% |
44.51% |
1878.50 |
0.5 |
v-1c5e |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Logical |
11.57% |
-15.68 |
20.39% |
44.52% |
1879.21 |
0.5 |
v-1c69 |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Smart |
23.45% |
-3.80 |
31.02% |
52.76% |
1906.86 |
0.5 |
v-1c6f |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
25.08% |
-2.17 |
32.32% |
55.02% |
703.03 |
1.6 |
v-17f3 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
25.08% |
-2.17 |
32.34% |
55.04% |
1794.21 |
2.2 |
v-1a55 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
25.11% |
-2.14 |
32.49% |
55.32% |
742.67 |
1.4 |
v-1c52 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
12.40% |
-14.85 |
21.41% |
46.72% |
728.69 |
1.4 |
v-1c4a |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
12.43% |
-14.82 |
21.38% |
46.67% |
735.66 |
1.4 |
v-1c43 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
25.11% |
-2.14 |
32.50% |
55.34% |
1766.27 |
2.2 |
v-1c55 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
12.38% |
-14.87 |
21.39% |
46.67% |
1801.98 |
2.2 |
v-1c4d |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
12.42% |
-14.83 |
21.40% |
46.67% |
1777.08 |
2.2 |
v-1c47 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
25.11% |
-2.14 |
32.49% |
55.34% |
1771.22 |
2.2 |
v-18de |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
25.08% |
-2.17 |
32.34% |
55.04% |
701.65 |
1.6 |
v-1811 |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
25.08% |
-2.17 |
32.34% |
55.04% |
1766.17 |
2.2 |
v-1a60 |
| Medium |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
19.51% |
-7.74 |
27.52% |
51.95% |
184.34 |
7.5 |
v-1651 |
| Medium |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
19.43% |
-7.82 |
27.41% |
51.85% |
159.73 |
9.0 |
v-199c |
| Medium |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
23.28% |
-3.97 |
31.08% |
53.93% |
483.07 |
2.2 |
v-1675 |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
27.29% |
+0.04 |
35.42% |
57.49% |
303.11 |
14.6 |
v-1602 |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
19.99% |
-7.26 |
28.94% |
52.34% |
201.89 |
22.3 |
v-1603 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
15.96% |
-11.29 |
25.86% |
50.98% |
64.80 |
41.2 |
v-17c4 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
26.43% |
-0.82 |
34.68% |
57.17% |
79.90 |
40.9 |
v-17c5 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
27.19% |
-0.06 |
35.53% |
57.63% |
140.49 |
17.8 |
v-17c6 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
25.97% |
-1.28 |
33.54% |
56.13% |
80.46 |
41.1 |
v-1dd0 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
13.72% |
-13.53 |
22.87% |
48.77% |
68.19 |
41.7 |
v-1dd1 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
27.28% |
+0.03 |
35.51% |
57.51% |
145.35 |
17.5 |
v-1dd2 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
27.20% |
-0.05 |
35.54% |
57.58% |
135.37 |
17.3 |
v-1dd4 |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
27.28% |
+0.03 |
35.54% |
57.50% |
135.20 |
17.1 |
v-1dd3 |
Full metrics — tiled
| Size |
Platform |
Runtime |
Decode |
Precision |
APs |
AP |
AP50 |
AP75 |
APm |
APl |
AR@1 |
AR@10 |
AR@100 |
ARs |
ARm |
ARl |
| Nano |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
FP32 |
20.35% |
28.77% |
48.33% |
28.89% |
40.09% |
43.56% |
11.34% |
33.86% |
43.74% |
35.17% |
55.99% |
56.33% |
| Nano |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
FP16 |
20.29% |
28.71% |
48.25% |
28.82% |
39.98% |
43.41% |
11.33% |
33.82% |
43.70% |
35.17% |
55.86% |
56.22% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
INT8 |
9.71% |
17.63% |
35.95% |
15.57% |
28.98% |
29.05% |
8.31% |
23.36% |
29.88% |
20.13% |
43.89% |
38.94% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Combined |
INT8 |
9.71% |
17.63% |
35.95% |
15.56% |
28.98% |
29.05% |
8.31% |
23.41% |
29.90% |
20.17% |
43.91% |
38.94% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Logical |
INT8 |
9.71% |
17.63% |
35.95% |
15.57% |
28.97% |
29.05% |
8.31% |
23.36% |
29.88% |
20.13% |
43.89% |
38.94% |
| Nano |
NXP i.MX 8M Plus + VeriSilicon NPU (EVK) |
LiteRT |
Smart |
INT8 |
15.70% |
23.07% |
40.08% |
22.82% |
33.57% |
30.08% |
9.85% |
29.03% |
37.85% |
29.65% |
49.64% |
40.53% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
INT8 |
19.04% |
27.41% |
46.85% |
27.65% |
38.87% |
43.42% |
11.07% |
32.57% |
42.09% |
33.70% |
54.21% |
56.26% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
INT8 |
19.04% |
27.41% |
46.85% |
27.65% |
38.87% |
43.42% |
11.07% |
32.57% |
42.09% |
33.70% |
54.20% |
56.26% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
INT8 |
19.22% |
27.45% |
46.84% |
27.58% |
38.71% |
44.17% |
11.02% |
32.73% |
42.20% |
33.84% |
54.31% |
58.26% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
INT8 |
11.95% |
21.00% |
42.27% |
18.67% |
33.34% |
42.83% |
9.33% |
26.35% |
33.24% |
23.24% |
48.00% |
56.99% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
INT8 |
11.84% |
20.96% |
42.20% |
18.65% |
33.38% |
42.89% |
9.32% |
26.28% |
33.15% |
23.05% |
47.99% |
57.01% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
INT8 |
19.22% |
27.45% |
46.84% |
27.58% |
38.71% |
44.17% |
11.02% |
32.73% |
42.20% |
33.84% |
54.31% |
58.26% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
INT8 |
11.95% |
21.00% |
42.27% |
18.67% |
33.34% |
42.83% |
9.33% |
26.35% |
33.24% |
23.24% |
48.00% |
56.99% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
INT8 |
11.84% |
20.96% |
42.21% |
18.65% |
33.38% |
42.89% |
9.32% |
26.28% |
33.15% |
23.05% |
47.99% |
57.02% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
INT8 |
19.22% |
27.45% |
46.84% |
27.57% |
38.70% |
44.17% |
11.02% |
32.73% |
42.20% |
33.84% |
54.31% |
58.26% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
INT8 |
19.04% |
27.41% |
46.85% |
27.65% |
38.87% |
43.42% |
11.07% |
32.57% |
42.09% |
33.70% |
54.21% |
56.26% |
| Nano |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
INT8 |
19.04% |
27.41% |
46.85% |
27.65% |
38.87% |
43.42% |
11.07% |
32.57% |
42.09% |
33.70% |
54.21% |
56.26% |
| Nano |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
INT16 |
16.43% |
25.13% |
45.44% |
24.23% |
36.98% |
42.96% |
10.57% |
30.25% |
38.66% |
29.50% |
52.02% |
57.38% |
| Nano |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
INT16 |
16.43% |
25.14% |
45.43% |
24.27% |
37.00% |
43.02% |
10.58% |
30.26% |
38.66% |
29.49% |
52.07% |
57.40% |
| Nano |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
INT8 |
17.63% |
26.19% |
46.59% |
25.38% |
37.91% |
42.80% |
10.82% |
31.69% |
40.76% |
31.82% |
54.20% |
55.07% |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
FP16 |
20.24% |
28.67% |
48.25% |
28.63% |
39.99% |
43.72% |
11.36% |
33.80% |
43.69% |
35.17% |
55.88% |
56.85% |
| Nano |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
INT8 |
18.17% |
26.28% |
45.73% |
26.22% |
37.29% |
43.81% |
10.69% |
31.85% |
41.34% |
33.18% |
53.21% |
58.18% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
INT8 |
12.56% |
21.37% |
42.92% |
19.28% |
33.65% |
39.22% |
9.48% |
26.76% |
34.01% |
23.84% |
48.73% |
53.07% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
INT8 |
19.39% |
27.56% |
47.14% |
27.37% |
38.76% |
42.70% |
11.01% |
32.82% |
42.47% |
34.17% |
54.44% |
54.71% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
INT16 |
19.98% |
28.43% |
47.93% |
28.47% |
39.95% |
45.13% |
11.47% |
33.61% |
43.34% |
34.82% |
55.55% |
61.11% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
INT8 |
18.94% |
26.97% |
46.74% |
26.87% |
37.92% |
41.52% |
10.99% |
32.25% |
41.80% |
33.49% |
53.82% |
55.12% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
INT8 |
11.88% |
20.94% |
42.28% |
18.33% |
33.00% |
40.07% |
9.58% |
26.23% |
33.35% |
23.15% |
48.00% |
52.61% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
INT16 |
20.08% |
28.53% |
47.98% |
28.52% |
40.07% |
46.00% |
11.41% |
33.58% |
43.41% |
34.89% |
55.57% |
60.86% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
INT16 |
19.99% |
28.44% |
47.93% |
28.53% |
39.97% |
45.14% |
11.47% |
33.64% |
43.38% |
34.84% |
55.61% |
61.12% |
| Nano |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
INT16 |
20.10% |
28.55% |
48.00% |
28.56% |
40.06% |
46.03% |
11.42% |
33.62% |
43.45% |
34.94% |
55.59% |
60.89% |
| Small |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
FP32 |
24.25% |
32.91% |
53.96% |
33.65% |
44.38% |
51.02% |
12.62% |
37.11% |
47.94% |
40.09% |
59.36% |
63.65% |
| Small |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
FP16 |
24.19% |
32.87% |
53.92% |
33.58% |
44.31% |
51.08% |
12.62% |
37.04% |
47.87% |
40.02% |
59.25% |
63.70% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
INT8 |
11.29% |
20.74% |
43.52% |
17.89% |
33.93% |
40.57% |
9.58% |
26.11% |
32.92% |
22.50% |
48.58% |
54.85% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Combined |
INT8 |
11.29% |
20.74% |
43.52% |
17.89% |
33.93% |
40.57% |
9.58% |
26.10% |
32.90% |
22.47% |
48.58% |
54.85% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Logical |
INT8 |
11.29% |
20.73% |
43.52% |
17.89% |
33.90% |
40.57% |
9.58% |
26.11% |
32.92% |
22.50% |
48.58% |
54.85% |
| Small |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Smart |
INT8 |
21.00% |
29.17% |
50.38% |
28.96% |
40.30% |
43.11% |
11.66% |
34.38% |
44.62% |
36.84% |
56.37% |
57.63% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
INT8 |
22.15% |
30.36% |
51.72% |
31.03% |
41.78% |
46.21% |
11.81% |
35.03% |
45.57% |
37.73% |
57.29% |
59.15% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
INT8 |
22.15% |
30.36% |
51.72% |
31.03% |
41.77% |
46.21% |
11.81% |
35.03% |
45.57% |
37.73% |
57.29% |
59.15% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
INT8 |
22.47% |
30.46% |
51.85% |
31.06% |
41.57% |
45.48% |
11.87% |
35.25% |
45.74% |
38.14% |
57.36% |
57.10% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
INT8 |
12.40% |
21.80% |
45.46% |
18.60% |
34.84% |
42.16% |
9.84% |
26.90% |
33.80% |
23.48% |
49.45% |
55.08% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
INT8 |
12.34% |
21.77% |
45.40% |
18.55% |
34.88% |
42.02% |
9.82% |
26.87% |
33.75% |
23.45% |
49.35% |
54.98% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
INT8 |
22.47% |
30.45% |
51.85% |
31.06% |
41.56% |
45.48% |
11.87% |
35.25% |
45.74% |
38.14% |
57.36% |
57.10% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
INT8 |
12.40% |
21.79% |
45.46% |
18.60% |
34.83% |
42.17% |
9.84% |
26.90% |
33.80% |
23.48% |
49.45% |
55.08% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
INT8 |
12.34% |
21.76% |
45.40% |
18.54% |
34.88% |
42.01% |
9.82% |
26.87% |
33.75% |
23.45% |
49.35% |
54.98% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
INT8 |
22.47% |
30.46% |
51.85% |
31.06% |
41.57% |
45.49% |
11.87% |
35.25% |
45.74% |
38.14% |
57.36% |
57.10% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
INT8 |
22.15% |
30.36% |
51.72% |
31.03% |
41.78% |
46.21% |
11.81% |
35.03% |
45.57% |
37.73% |
57.29% |
59.15% |
| Small |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
INT8 |
22.15% |
30.36% |
51.72% |
31.03% |
41.78% |
46.21% |
11.81% |
35.03% |
45.57% |
37.73% |
57.29% |
59.15% |
| Small |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
INT16 |
17.95% |
26.61% |
49.80% |
24.81% |
38.84% |
42.05% |
10.85% |
31.23% |
39.98% |
30.80% |
53.90% |
54.76% |
| Small |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
INT16 |
17.93% |
26.62% |
49.81% |
24.85% |
38.84% |
43.61% |
10.85% |
31.24% |
39.99% |
30.81% |
53.91% |
57.76% |
| Small |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
INT8 |
21.50% |
30.17% |
52.28% |
30.12% |
41.91% |
48.68% |
11.82% |
34.91% |
44.83% |
36.61% |
56.89% |
63.38% |
| Small |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
FP16 |
24.25% |
32.82% |
53.89% |
33.55% |
44.17% |
51.05% |
12.64% |
36.97% |
47.78% |
40.00% |
59.02% |
63.56% |
| Small |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
INT8 |
21.16% |
26.61% |
48.14% |
26.16% |
35.76% |
44.74% |
10.86% |
33.73% |
44.02% |
37.95% |
53.45% |
55.58% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
INT8 |
15.02% |
24.80% |
48.55% |
22.20% |
37.97% |
46.86% |
10.58% |
29.39% |
37.22% |
27.13% |
52.27% |
58.66% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
INT8 |
23.50% |
31.83% |
53.26% |
32.47% |
42.93% |
48.73% |
12.33% |
36.30% |
47.00% |
39.44% |
58.25% |
61.37% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
INT16 |
24.27% |
32.97% |
53.95% |
33.82% |
44.45% |
49.95% |
12.65% |
37.10% |
48.06% |
40.26% |
59.44% |
63.01% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
INT8 |
22.25% |
30.50% |
51.78% |
31.07% |
42.11% |
48.61% |
12.14% |
35.60% |
46.08% |
38.18% |
58.00% |
63.22% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
INT8 |
13.38% |
23.11% |
46.62% |
20.29% |
36.81% |
47.01% |
10.21% |
28.29% |
35.72% |
25.39% |
51.53% |
59.87% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
INT16 |
24.25% |
32.95% |
53.89% |
33.71% |
44.53% |
49.91% |
12.61% |
37.08% |
48.15% |
40.35% |
59.59% |
62.58% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
INT16 |
24.28% |
32.97% |
53.95% |
33.79% |
44.45% |
49.93% |
12.65% |
37.10% |
48.06% |
40.26% |
59.44% |
63.00% |
| Small |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
INT16 |
24.25% |
32.94% |
53.87% |
33.72% |
44.53% |
49.89% |
12.61% |
37.08% |
48.15% |
40.35% |
59.59% |
62.56% |
| Medium |
ONNX FP32 (CUDA · RTX 3090) |
ONNX |
— |
FP32 |
27.25% |
35.52% |
57.55% |
37.05% |
47.02% |
51.24% |
13.58% |
39.36% |
50.64% |
43.41% |
61.56% |
66.30% |
| Medium |
ONNX FP16 (CUDA · RTX 3090) |
ONNX |
— |
FP16 |
27.24% |
35.47% |
57.50% |
37.02% |
47.04% |
51.14% |
13.58% |
39.33% |
50.56% |
43.35% |
61.57% |
66.25% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency |
LiteRT |
Logical |
INT8 |
11.57% |
20.39% |
44.52% |
16.69% |
34.02% |
44.88% |
9.59% |
26.27% |
32.95% |
22.57% |
48.81% |
57.84% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Combined |
INT8 |
11.57% |
20.37% |
44.51% |
16.68% |
34.00% |
44.88% |
9.59% |
26.27% |
32.93% |
22.57% |
48.76% |
57.84% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Logical |
INT8 |
11.57% |
20.39% |
44.52% |
16.71% |
34.02% |
44.88% |
9.59% |
26.27% |
32.95% |
22.57% |
48.81% |
57.84% |
| Medium |
NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) |
LiteRT |
Smart |
INT8 |
23.45% |
31.02% |
52.76% |
31.60% |
42.20% |
46.40% |
12.38% |
36.31% |
46.92% |
39.59% |
58.40% |
60.06% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency |
LiteRT |
Smart |
INT8 |
25.08% |
32.32% |
55.02% |
32.71% |
43.15% |
43.93% |
12.92% |
37.45% |
48.41% |
41.55% |
59.47% |
59.57% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput |
LiteRT |
Smart |
INT8 |
25.08% |
32.34% |
55.04% |
32.73% |
43.15% |
43.92% |
12.92% |
37.45% |
48.41% |
41.55% |
59.47% |
59.57% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Smart |
INT8 |
25.11% |
32.49% |
55.32% |
32.92% |
43.55% |
45.29% |
12.80% |
37.56% |
48.53% |
41.49% |
59.66% |
62.14% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Logical |
INT8 |
12.40% |
21.41% |
46.72% |
17.04% |
35.51% |
43.25% |
9.91% |
27.28% |
34.12% |
23.73% |
50.44% |
59.53% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency |
LiteRT |
Combined |
INT8 |
12.43% |
21.38% |
46.67% |
16.97% |
35.45% |
42.91% |
9.94% |
27.27% |
34.13% |
23.76% |
50.39% |
59.35% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Smart |
INT8 |
25.11% |
32.50% |
55.34% |
32.96% |
43.55% |
45.28% |
12.80% |
37.56% |
48.53% |
41.49% |
59.66% |
62.14% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Logical |
INT8 |
12.38% |
21.39% |
46.67% |
17.02% |
35.48% |
43.21% |
9.90% |
27.27% |
34.11% |
23.72% |
50.42% |
59.50% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput |
LiteRT |
Combined |
INT8 |
12.42% |
21.40% |
46.67% |
16.99% |
35.46% |
42.91% |
9.94% |
27.27% |
34.13% |
23.76% |
50.39% |
59.35% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput |
LiteRT |
Smart |
INT8 |
25.11% |
32.49% |
55.34% |
32.93% |
43.56% |
45.27% |
12.80% |
37.56% |
48.53% |
41.49% |
59.66% |
62.14% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency |
LiteRT |
Smart |
INT8 |
25.08% |
32.34% |
55.04% |
32.73% |
43.15% |
43.93% |
12.92% |
37.45% |
48.41% |
41.55% |
59.47% |
59.57% |
| Medium |
NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput |
LiteRT |
Smart |
INT8 |
25.08% |
32.34% |
55.04% |
32.74% |
43.16% |
43.93% |
12.92% |
37.45% |
48.41% |
41.55% |
59.47% |
59.57% |
| Medium |
NXP Ara240 (FRDM) — latency |
Ara DVM |
Smart |
INT16 |
19.51% |
27.52% |
51.95% |
26.10% |
40.09% |
42.19% |
11.57% |
33.08% |
42.42% |
34.17% |
55.91% |
58.02% |
| Medium |
NXP Ara240 (FRDM) — throughput |
Ara DVM |
Smart |
INT16 |
19.43% |
27.41% |
51.85% |
26.01% |
40.00% |
42.37% |
11.53% |
33.02% |
42.38% |
34.15% |
55.81% |
58.10% |
| Medium |
Raspberry Pi 5 + Hailo-8L NPU |
Hailo HEF |
— |
INT8 |
23.28% |
31.08% |
53.93% |
31.06% |
42.59% |
46.94% |
12.50% |
35.49% |
45.58% |
37.87% |
57.80% |
62.16% |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT FP16) |
TensorRT |
— |
FP16 |
27.29% |
35.42% |
57.49% |
36.93% |
46.83% |
50.22% |
13.59% |
39.29% |
50.50% |
43.31% |
61.35% |
65.25% |
| Medium |
NVIDIA Jetson Orin Nano (TensorRT INT8) |
TensorRT |
— |
INT8 |
19.99% |
28.94% |
52.34% |
28.22% |
41.84% |
49.15% |
12.08% |
34.39% |
43.86% |
34.93% |
58.06% |
64.15% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset |
ONNX · QNN EP |
Logical |
INT8 |
15.96% |
25.86% |
50.98% |
23.21% |
39.94% |
47.75% |
11.18% |
30.38% |
38.14% |
28.07% |
53.56% |
62.94% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset |
ONNX · QNN EP |
Smart |
INT8 |
26.43% |
34.68% |
57.17% |
35.78% |
46.26% |
50.08% |
13.29% |
38.61% |
49.77% |
42.58% |
60.79% |
65.08% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset |
ONNX · QNN EP |
Smart |
INT16 |
27.19% |
35.53% |
57.63% |
37.04% |
47.01% |
49.99% |
13.66% |
39.30% |
50.52% |
43.20% |
61.51% |
66.04% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Smart |
INT8 |
25.97% |
33.54% |
56.13% |
34.34% |
44.99% |
48.07% |
13.18% |
37.86% |
48.73% |
41.76% |
59.76% |
62.35% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard) |
ONNX · QNN EP |
Logical |
INT8 |
13.72% |
22.87% |
48.77% |
18.92% |
37.09% |
45.42% |
10.17% |
28.02% |
35.19% |
24.92% |
51.20% |
60.10% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Smart |
INT16 |
27.28% |
35.51% |
57.51% |
37.07% |
47.05% |
50.20% |
13.71% |
39.39% |
50.63% |
43.35% |
61.67% |
66.28% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) |
ONNX · QNN EP |
Logical |
INT16 |
27.20% |
35.54% |
57.58% |
37.08% |
47.01% |
50.12% |
13.68% |
39.33% |
50.54% |
43.24% |
61.53% |
66.09% |
| Medium |
Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE) |
ONNX · QNN EP |
Logical |
INT16 |
27.28% |
35.54% |
57.50% |
37.10% |
47.08% |
50.36% |
13.72% |
39.39% |
50.64% |
43.35% |
61.67% |
66.34% |
Validation pipeline
- EdgeFirst Profiler runs on the target hardware, executes the full inference pipeline (capture → preprocess → inference → postprocess, with slicing and merging for the tiled variant), and scores the predictions against ground truth on the device.
- EdgeFirst Studio receives the metrics, predictions, charts and trace, and publishes them as the
v-XXXX validation session every row above cites.
- EdgeFirst HAL (open source) provides the hardware-accelerated preprocessing and post-decoding primitives used at both validation and deployment time.
Validation runs decode a JPEG per image and evaluate at a 0.001 confidence threshold to capture every detection for mAP. A camera pipeline at a deployment threshold of 0.25–0.75 carries a lighter host and postprocess load.
Downloads
Every file below is the artifact a validation session in the tables above measured. Files are organized as {variant}/{target folder}/ and named visdrone-yolo26{size}-det-{precision}[-smart|-logical]-{variant}{extension}. Each model file embeds the EdgeFirst edgefirst.json metadata (training session, dataset version, calibration artifact, converter chain), so a single file is sufficient for deployment.
640/
ONNX FP32 — Any platform with ONNX Runtime.
| Size |
File |
Variant |
Download |
| Nano |
640/onnx/visdrone-yolo26n-det-fp16-640.onnx |
FP16 |
Download |
| Nano |
640/onnx/visdrone-yolo26n-det-fp32-640.onnx |
FP32 |
Download |
| Small |
640/onnx/visdrone-yolo26s-det-fp16-640.onnx |
FP16 |
Download |
| Small |
640/onnx/visdrone-yolo26s-det-fp32-640.onnx |
FP32 |
Download |
| Medium |
640/onnx/visdrone-yolo26m-det-fp16-640.onnx |
FP16 |
Download |
| Medium |
640/onnx/visdrone-yolo26m-det-fp32-640.onnx |
FP32 |
Download |
TFLite INT8 — CPU or NPU via runtime delegate (NXP i.MX 8M Plus VX Delegate).
| Size |
File |
Variant |
Download |
| Nano |
640/tflite/visdrone-yolo26n-det-int8-640.tflite |
Logical · INT8 |
Download |
| Nano |
640/tflite/visdrone-yolo26n-det-int8-smart-640.tflite |
Smart · INT8 |
Download |
| Small |
640/tflite/visdrone-yolo26s-det-int8-640.tflite |
Logical · INT8 |
Download |
| Small |
640/tflite/visdrone-yolo26s-det-int8-smart-640.tflite |
Smart · INT8 |
Download |
| Medium |
640/tflite/visdrone-yolo26m-det-int8-640.tflite |
Logical · INT8 |
Download |
| Medium |
640/tflite/visdrone-yolo26m-det-int8-smart-640.tflite |
Smart · INT8 |
Download |
NXP i.MX 95 (eIQ Neutron) — eIQ Neutron NPU optimized.
| Size |
File |
Variant |
Download |
| Nano |
640/imx95/visdrone-yolo26n-det-int8-640.imx95.tflite |
Logical · INT8 |
Download |
| Nano |
640/imx95/visdrone-yolo26n-det-int8-smart-640.imx95.tflite |
Smart · INT8 |
Download |
| Small |
640/imx95/visdrone-yolo26s-det-int8-640.imx95.tflite |
Logical · INT8 |
Download |
| Small |
640/imx95/visdrone-yolo26s-det-int8-smart-640.imx95.tflite |
Smart · INT8 |
Download |
| Medium |
640/imx95/visdrone-yolo26m-det-int8-640.imx95.tflite |
Logical · INT8 |
Download |
| Medium |
640/imx95/visdrone-yolo26m-det-int8-smart-640.imx95.tflite |
Smart · INT8 |
Download |
NXP Ara240 — Ara240 DNPU compiled model.
| Size |
File |
Variant |
Download |
| Small |
640/ara240/visdrone-yolo26s-det-int16-640.dvm |
Smart · INT16 |
Download |
| Medium |
640/ara240/visdrone-yolo26m-det-int16-640.dvm |
Smart · INT16 |
Download |
RPi5 + Hailo-8L (13 TOPS) — Hailo-8L 13 TOPS accelerator.
| Size |
File |
Variant |
Download |
| Nano |
640/hailo/visdrone-yolo26n-det-int8-640.hailo8l.hef |
INT8 |
Download |
| Small |
640/hailo/visdrone-yolo26s-det-int8-640.hailo8l.hef |
INT8 |
Download |
| Medium |
640/hailo/visdrone-yolo26m-det-int8-640.hailo8l.hef |
INT8 |
Download |
NVIDIA Jetson (TensorRT) — Jetson FP16 and INT8 engines.
| Size |
File |
Variant |
Download |
| Nano |
640/jetson/visdrone-yolo26n-det-fp16-640.engine |
FP16 |
Download |
| Nano |
640/jetson/visdrone-yolo26n-det-int8-640.engine |
INT8 |
Download |
| Small |
640/jetson/visdrone-yolo26s-det-fp16-640.engine |
FP16 |
Download |
| Small |
640/jetson/visdrone-yolo26s-det-int8-640.engine |
INT8 |
Download |
| Medium |
640/jetson/visdrone-yolo26m-det-fp16-640.engine |
FP16 |
Download |
| Medium |
640/jetson/visdrone-yolo26m-det-int8-640.engine |
INT8 |
Download |
Qualcomm Dragonwing IQ-9075 (Hexagon v73) — Offline-compiled Hexagon context binary, run through the ONNX Runtime QNN execution provider. Loads only on the Dragonwing IQ-9075 (Hexagon v73).
| Size |
File |
Variant |
Download |
| Nano |
640/qnn/visdrone-yolo26n-det-w8a16-logical-in8-640.iq9075.qnn.onnx |
Logical · INT16 · Standard |
Download |
| Nano |
640/qnn/visdrone-yolo26n-det-w8a16-logical-tfe-seqmse-in8-640.iq9075.qnn.onnx |
Logical · INT16 · TFE + SeqMSE |
Download |
| Nano |
640/qnn/visdrone-yolo26n-det-w8a16-smart-in8-640.iq9075.qnn.onnx |
Smart · INT16 · Standard |
Download |
| Nano |
640/qnn/visdrone-yolo26n-det-w8a16-smart-tfe-seqmse-in8-640.iq9075.qnn.onnx |
Smart · INT16 · TFE + SeqMSE |
Download |
| Nano |
640/qnn/visdrone-yolo26n-det-w8a8-logical-640.iq9075.qnn.onnx |
Logical · INT8 · Standard |
Download |
| Nano |
640/qnn/visdrone-yolo26n-det-w8a8-logical-tfe-seqmse-640.iq9075.qnn.onnx |
Logical · INT8 · TFE + SeqMSE |
Download |
| Nano |
640/qnn/visdrone-yolo26n-det-w8a8-smart-640.iq9075.qnn.onnx |
Smart · INT8 · Standard |
Download |
| Nano |
640/qnn/visdrone-yolo26n-det-w8a8-smart-tfe-seqmse-640.iq9075.qnn.onnx |
Smart · INT8 · TFE + SeqMSE |
Download |
| Small |
640/qnn/visdrone-yolo26s-det-w8a16-logical-in8-640.iq9075.qnn.onnx |
Logical · INT16 · Standard |
Download |
| Small |
640/qnn/visdrone-yolo26s-det-w8a16-logical-tfe-seqmse-in8-640.iq9075.qnn.onnx |
Logical · INT16 · TFE + SeqMSE |
Download |
| Small |
640/qnn/visdrone-yolo26s-det-w8a16-smart-in8-640.iq9075.qnn.onnx |
Smart · INT16 · Standard |
Download |
| Small |
640/qnn/visdrone-yolo26s-det-w8a16-smart-tfe-seqmse-in8-640.iq9075.qnn.onnx |
Smart · INT16 · TFE + SeqMSE |
Download |
| Small |
640/qnn/visdrone-yolo26s-det-w8a8-logical-640.iq9075.qnn.onnx |
Logical · INT8 · Standard |
Download |
| Small |
640/qnn/visdrone-yolo26s-det-w8a8-logical-tfe-seqmse-640.iq9075.qnn.onnx |
Logical · INT8 · TFE + SeqMSE |
Download |
| Small |
640/qnn/visdrone-yolo26s-det-w8a8-smart-640.iq9075.qnn.onnx |
Smart · INT8 · Standard |
Download |
| Small |
640/qnn/visdrone-yolo26s-det-w8a8-smart-tfe-seqmse-640.iq9075.qnn.onnx |
Smart · INT8 · TFE + SeqMSE |
Download |
| Medium |
640/qnn/visdrone-yolo26m-det-w8a16-logical-in8-640.iq9075.qnn.onnx |
Logical · INT16 · Standard |
Download |
| Medium |
640/qnn/visdrone-yolo26m-det-w8a16-logical-tfe-seqmse-in8-640.iq9075.qnn.onnx |
Logical · INT16 · TFE + SeqMSE |
Download |
| Medium |
640/qnn/visdrone-yolo26m-det-w8a16-smart-in8-640.iq9075.qnn.onnx |
Smart · INT16 · Standard |
Download |
| Medium |
640/qnn/visdrone-yolo26m-det-w8a16-smart-tfe-seqmse-in8-640.iq9075.qnn.onnx |
Smart · INT16 · TFE + SeqMSE |
Download |
| Medium |
640/qnn/visdrone-yolo26m-det-w8a8-logical-640.iq9075.qnn.onnx |
Logical · INT8 · Standard |
Download |
| Medium |
640/qnn/visdrone-yolo26m-det-w8a8-logical-tfe-seqmse-640.iq9075.qnn.onnx |
Logical · INT8 · TFE + SeqMSE |
Download |
| Medium |
640/qnn/visdrone-yolo26m-det-w8a8-smart-640.iq9075.qnn.onnx |
Smart · INT8 · Standard |
Download |
| Medium |
640/qnn/visdrone-yolo26m-det-w8a8-smart-tfe-seqmse-640.iq9075.qnn.onnx |
Smart · INT8 · TFE + SeqMSE |
Download |
1280/
ONNX FP32 — Any platform with ONNX Runtime.
| Size |
File |
Variant |
Download |
| Nano |
1280/onnx/visdrone-yolo26n-det-fp16-1280.onnx |
FP16 |
Download |
| Nano |
1280/onnx/visdrone-yolo26n-det-fp32-1280.onnx |
FP32 |
Download |
| Small |
1280/onnx/visdrone-yolo26s-det-fp16-1280.onnx |
FP16 |
Download |
| Small |
1280/onnx/visdrone-yolo26s-det-fp32-1280.onnx |
FP32 |
Download |
| Medium |
1280/onnx/visdrone-yolo26m-det-fp16-1280.onnx |
FP16 |
Download |
| Medium |
1280/onnx/visdrone-yolo26m-det-fp32-1280.onnx |
FP32 |
Download |
TFLite INT8 — CPU or NPU via runtime delegate (NXP i.MX 8M Plus VX Delegate).
| Size |
File |
Variant |
Download |
| Nano |
1280/tflite/visdrone-yolo26n-det-int8-1280.tflite |
Logical · INT8 |
Download |
| Nano |
1280/tflite/visdrone-yolo26n-det-int8-smart-1280.tflite |
Smart · INT8 |
Download |
| Small |
1280/tflite/visdrone-yolo26s-det-int8-1280.tflite |
Logical · INT8 |
Download |
| Small |
1280/tflite/visdrone-yolo26s-det-int8-smart-1280.tflite |
Smart · INT8 |
Download |
| Medium |
1280/tflite/visdrone-yolo26m-det-int8-1280.tflite |
Logical · INT8 |
Download |
| Medium |
1280/tflite/visdrone-yolo26m-det-int8-smart-1280.tflite |
Smart · INT8 |
Download |
NXP i.MX 95 (eIQ Neutron) — eIQ Neutron NPU optimized.
| Size |
File |
Variant |
Download |
| Nano |
1280/imx95/visdrone-yolo26n-det-int8-1280.imx95.tflite |
Logical · INT8 |
Download |
| Nano |
1280/imx95/visdrone-yolo26n-det-int8-smart-1280.imx95.tflite |
Smart · INT8 |
Download |
| Small |
1280/imx95/visdrone-yolo26s-det-int8-1280.imx95.tflite |
Logical · INT8 |
Download |
| Small |
1280/imx95/visdrone-yolo26s-det-int8-smart-1280.imx95.tflite |
Smart · INT8 |
Download |
| Medium |
1280/imx95/visdrone-yolo26m-det-int8-1280.imx95.tflite |
Logical · INT8 |
Download |
| Medium |
1280/imx95/visdrone-yolo26m-det-int8-smart-1280.imx95.tflite |
Smart · INT8 |
Download |
RPi5 + Hailo-8L (13 TOPS) — Hailo-8L 13 TOPS accelerator.
| Size |
File |
Variant |
Download |
| Nano |
1280/hailo/visdrone-yolo26n-det-int8-1280.hailo8l.hef |
INT8 |
Download |
| Small |
1280/hailo/visdrone-yolo26s-det-int8-1280.hailo8l.hef |
INT8 |
Download |
NVIDIA Jetson (TensorRT) — Jetson FP16 and INT8 engines.
| Size |
File |
Variant |
Download |
| Nano |
1280/jetson/visdrone-yolo26n-det-fp16-1280.engine |
FP16 |
Download |
| Nano |
1280/jetson/visdrone-yolo26n-det-int8-1280.engine |
INT8 |
Download |
| Small |
1280/jetson/visdrone-yolo26s-det-fp16-1280.engine |
FP16 |
Download |
| Small |
1280/jetson/visdrone-yolo26s-det-int8-1280.engine |
INT8 |
Download |
| Medium |
1280/jetson/visdrone-yolo26m-det-fp16-1280.engine |
FP16 |
Download |
| Medium |
1280/jetson/visdrone-yolo26m-det-int8-1280.engine |
INT8 |
Download |
Qualcomm Dragonwing IQ-9075 (Hexagon v73) — Offline-compiled Hexagon context binary, run through the ONNX Runtime QNN execution provider. Loads only on the Dragonwing IQ-9075 (Hexagon v73).
| Size |
File |
Variant |
Download |
| Nano |
1280/qnn/visdrone-yolo26n-det-w8a16-logical-in8-1280.iq9075.qnn.onnx |
Logical · INT16 · Standard |
Download |
| Nano |
1280/qnn/visdrone-yolo26n-det-w8a16-logical-tfe-seqmse-in8-1280.iq9075.qnn.onnx |
Logical · INT16 · TFE + SeqMSE |
Download |
| Nano |
1280/qnn/visdrone-yolo26n-det-w8a16-smart-in8-1280.iq9075.qnn.onnx |
Smart · INT16 · Standard |
Download |
| Nano |
1280/qnn/visdrone-yolo26n-det-w8a16-smart-tfe-seqmse-in8-1280.iq9075.qnn.onnx |
Smart · INT16 · TFE + SeqMSE |
Download |
| Nano |
1280/qnn/visdrone-yolo26n-det-w8a8-logical-1280.iq9075.qnn.onnx |
Logical · INT8 · Standard |
Download |
| Nano |
1280/qnn/visdrone-yolo26n-det-w8a8-logical-tfe-seqmse-1280.iq9075.qnn.onnx |
Logical · INT8 · TFE + SeqMSE |
Download |
| Nano |
1280/qnn/visdrone-yolo26n-det-w8a8-smart-1280.iq9075.qnn.onnx |
Smart · INT8 · Standard |
Download |
| Nano |
1280/qnn/visdrone-yolo26n-det-w8a8-smart-tfe-seqmse-1280.iq9075.qnn.onnx |
Smart · INT8 · TFE + SeqMSE |
Download |
| Small |
1280/qnn/visdrone-yolo26s-det-w8a16-logical-in8-1280.iq9075.qnn.onnx |
Logical · INT16 · Standard |
Download |
| Small |
1280/qnn/visdrone-yolo26s-det-w8a16-logical-tfe-seqmse-in8-1280.iq9075.qnn.onnx |
Logical · INT16 · TFE + SeqMSE |
Download |
| Small |
1280/qnn/visdrone-yolo26s-det-w8a16-smart-in8-1280.iq9075.qnn.onnx |
Smart · INT16 · Standard |
Download |
| Small |
1280/qnn/visdrone-yolo26s-det-w8a16-smart-tfe-seqmse-in8-1280.iq9075.qnn.onnx |
Smart · INT16 · TFE + SeqMSE |
Download |
| Small |
1280/qnn/visdrone-yolo26s-det-w8a8-logical-1280.iq9075.qnn.onnx |
Logical · INT8 · Standard |
Download |
| Small |
1280/qnn/visdrone-yolo26s-det-w8a8-logical-tfe-seqmse-1280.iq9075.qnn.onnx |
Logical · INT8 · TFE + SeqMSE |
Download |
| Small |
1280/qnn/visdrone-yolo26s-det-w8a8-smart-1280.iq9075.qnn.onnx |
Smart · INT8 · Standard |
Download |
| Small |
1280/qnn/visdrone-yolo26s-det-w8a8-smart-tfe-seqmse-1280.iq9075.qnn.onnx |
Smart · INT8 · TFE + SeqMSE |
Download |
| Medium |
1280/qnn/visdrone-yolo26m-det-w8a16-logical-in8-1280.iq9075.qnn.onnx |
Logical · INT16 · Standard |
Download |
| Medium |
1280/qnn/visdrone-yolo26m-det-w8a16-logical-tfe-seqmse-in8-1280.iq9075.qnn.onnx |
Logical · INT16 · TFE + SeqMSE |
Download |
| Medium |
1280/qnn/visdrone-yolo26m-det-w8a16-smart-in8-1280.iq9075.qnn.onnx |
Smart · INT16 · Standard |
Download |
| Medium |
1280/qnn/visdrone-yolo26m-det-w8a16-smart-tfe-seqmse-in8-1280.iq9075.qnn.onnx |
Smart · INT16 · TFE + SeqMSE |
Download |
| Medium |
1280/qnn/visdrone-yolo26m-det-w8a8-logical-1280.iq9075.qnn.onnx |
Logical · INT8 · Standard |
Download |
| Medium |
1280/qnn/visdrone-yolo26m-det-w8a8-logical-tfe-seqmse-1280.iq9075.qnn.onnx |
Logical · INT8 · TFE + SeqMSE |
Download |
| Medium |
1280/qnn/visdrone-yolo26m-det-w8a8-smart-1280.iq9075.qnn.onnx |
Smart · INT8 · Standard |
Download |
| Medium |
1280/qnn/visdrone-yolo26m-det-w8a8-smart-tfe-seqmse-1280.iq9075.qnn.onnx |
Smart · INT8 · TFE + SeqMSE |
Download |
1344x768/
ONNX FP32 — Any platform with ONNX Runtime.
| Size |
File |
Variant |
Download |
| Nano |
1344x768/onnx/visdrone-yolo26n-det-fp16-1344x768.onnx |
FP16 |
Download |
| Nano |
1344x768/onnx/visdrone-yolo26n-det-fp32-1344x768.onnx |
FP32 |
Download |
| Small |
1344x768/onnx/visdrone-yolo26s-det-fp16-1344x768.onnx |
FP16 |
Download |
| Small |
1344x768/onnx/visdrone-yolo26s-det-fp32-1344x768.onnx |
FP32 |
Download |
| Medium |
1344x768/onnx/visdrone-yolo26m-det-fp16-1344x768.onnx |
FP16 |
Download |
| Medium |
1344x768/onnx/visdrone-yolo26m-det-fp32-1344x768.onnx |
FP32 |
Download |
TFLite INT8 — CPU or NPU via runtime delegate (NXP i.MX 8M Plus VX Delegate).
| Size |
File |
Variant |
Download |
| Nano |
1344x768/tflite/visdrone-yolo26n-det-int8-1344x768.tflite |
Logical · INT8 |
Download |
| Nano |
1344x768/tflite/visdrone-yolo26n-det-int8-smart-1344x768.tflite |
Smart · INT8 |
Download |
| Small |
1344x768/tflite/visdrone-yolo26s-det-int8-1344x768.tflite |
Logical · INT8 |
Download |
| Small |
1344x768/tflite/visdrone-yolo26s-det-int8-smart-1344x768.tflite |
Smart · INT8 |
Download |
| Medium |
1344x768/tflite/visdrone-yolo26m-det-int8-1344x768.tflite |
Logical · INT8 |
Download |
| Medium |
1344x768/tflite/visdrone-yolo26m-det-int8-smart-1344x768.tflite |
Smart · INT8 |
Download |
NXP i.MX 95 (eIQ Neutron) — eIQ Neutron NPU optimized.
| Size |
File |
Variant |
Download |
| Nano |
1344x768/imx95/visdrone-yolo26n-det-int8-1344x768.imx95.tflite |
Logical · INT8 |
Download |
| Nano |
1344x768/imx95/visdrone-yolo26n-det-int8-smart-1344x768.imx95.tflite |
Smart · INT8 |
Download |
| Small |
1344x768/imx95/visdrone-yolo26s-det-int8-1344x768.imx95.tflite |
Logical · INT8 |
Download |
| Small |
1344x768/imx95/visdrone-yolo26s-det-int8-smart-1344x768.imx95.tflite |
Smart · INT8 |
Download |
| Medium |
1344x768/imx95/visdrone-yolo26m-det-int8-1344x768.imx95.tflite |
Logical · INT8 |
Download |
| Medium |
1344x768/imx95/visdrone-yolo26m-det-int8-smart-1344x768.imx95.tflite |
Smart · INT8 |
Download |
NXP Ara240 — Ara240 DNPU compiled model.
| Size |
File |
Variant |
Download |
| Nano |
1344x768/ara240/visdrone-yolo26n-det-int16-1344x768.dvm |
Smart · INT16 |
Download |
| Small |
1344x768/ara240/visdrone-yolo26s-det-int16-1344x768.dvm |
Smart · INT16 |
Download |
| Medium |
1344x768/ara240/visdrone-yolo26m-det-int16-1344x768.dvm |
Smart · INT16 |
Download |
RPi5 + Hailo-8L (13 TOPS) — Hailo-8L 13 TOPS accelerator.
| Size |
File |
Variant |
Download |
| Medium |
1344x768/hailo/visdrone-yolo26m-det-int8-1344x768.hailo8l.hef |
INT8 |
Download |
NVIDIA Jetson (TensorRT) — Jetson FP16 and INT8 engines.
| Size |
File |
Variant |
Download |
| Nano |
1344x768/jetson/visdrone-yolo26n-det-fp16-1344x768.engine |
FP16 |
Download |
| Nano |
1344x768/jetson/visdrone-yolo26n-det-int8-1344x768.engine |
INT8 |
Download |
| Small |
1344x768/jetson/visdrone-yolo26s-det-fp16-1344x768.engine |
FP16 |
Download |
| Small |
1344x768/jetson/visdrone-yolo26s-det-int8-1344x768.engine |
INT8 |
Download |
| Medium |
1344x768/jetson/visdrone-yolo26m-det-fp16-1344x768.engine |
FP16 |
Download |
| Medium |
1344x768/jetson/visdrone-yolo26m-det-int8-1344x768.engine |
INT8 |
Download |
Qualcomm Dragonwing IQ-9075 (Hexagon v73) — Offline-compiled Hexagon context binary, run through the ONNX Runtime QNN execution provider. Loads only on the Dragonwing IQ-9075 (Hexagon v73).
| Size |
File |
Variant |
Download |
| Nano |
1344x768/qnn/visdrone-yolo26n-det-w8a16-logical-in8-1344x768.iq9075.qnn.onnx |
Logical · INT16 · Standard |
Download |
| Nano |
1344x768/qnn/visdrone-yolo26n-det-w8a16-logical-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx |
Logical · INT16 · TFE + SeqMSE |
Download |
| Nano |
1344x768/qnn/visdrone-yolo26n-det-w8a16-smart-in8-1344x768.iq9075.qnn.onnx |
Smart · INT16 · Standard |
Download |
| Nano |
1344x768/qnn/visdrone-yolo26n-det-w8a16-smart-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx |
Smart · INT16 · TFE + SeqMSE |
Download |
| Nano |
1344x768/qnn/visdrone-yolo26n-det-w8a8-logical-1344x768.iq9075.qnn.onnx |
Logical · INT8 · Standard |
Download |
| Nano |
1344x768/qnn/visdrone-yolo26n-det-w8a8-logical-tfe-seqmse-1344x768.iq9075.qnn.onnx |
Logical · INT8 · TFE + SeqMSE |
Download |
| Nano |
1344x768/qnn/visdrone-yolo26n-det-w8a8-smart-1344x768.iq9075.qnn.onnx |
Smart · INT8 · Standard |
Download |
| Nano |
1344x768/qnn/visdrone-yolo26n-det-w8a8-smart-tfe-seqmse-1344x768.iq9075.qnn.onnx |
Smart · INT8 · TFE + SeqMSE |
Download |
| Small |
1344x768/qnn/visdrone-yolo26s-det-w8a16-logical-in8-1344x768.iq9075.qnn.onnx |
Logical · INT16 · Standard |
Download |
| Small |
1344x768/qnn/visdrone-yolo26s-det-w8a16-logical-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx |
Logical · INT16 · TFE + SeqMSE |
Download |
| Small |
1344x768/qnn/visdrone-yolo26s-det-w8a16-smart-in8-1344x768.iq9075.qnn.onnx |
Smart · INT16 · Standard |
Download |
| Small |
1344x768/qnn/visdrone-yolo26s-det-w8a16-smart-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx |
Smart · INT16 · TFE + SeqMSE |
Download |
| Small |
1344x768/qnn/visdrone-yolo26s-det-w8a8-logical-1344x768.iq9075.qnn.onnx |
Logical · INT8 · Standard |
Download |
| Small |
1344x768/qnn/visdrone-yolo26s-det-w8a8-logical-tfe-seqmse-1344x768.iq9075.qnn.onnx |
Logical · INT8 · TFE + SeqMSE |
Download |
| Small |
1344x768/qnn/visdrone-yolo26s-det-w8a8-smart-1344x768.iq9075.qnn.onnx |
Smart · INT8 · Standard |
Download |
| Small |
1344x768/qnn/visdrone-yolo26s-det-w8a8-smart-tfe-seqmse-1344x768.iq9075.qnn.onnx |
Smart · INT8 · TFE + SeqMSE |
Download |
| Medium |
1344x768/qnn/visdrone-yolo26m-det-w8a16-logical-in8-1344x768.iq9075.qnn.onnx |
Logical · INT16 · Standard |
Download |
| Medium |
1344x768/qnn/visdrone-yolo26m-det-w8a16-logical-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx |
Logical · INT16 · TFE + SeqMSE |
Download |
| Medium |
1344x768/qnn/visdrone-yolo26m-det-w8a16-smart-in8-1344x768.iq9075.qnn.onnx |
Smart · INT16 · Standard |
Download |
| Medium |
1344x768/qnn/visdrone-yolo26m-det-w8a16-smart-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx |
Smart · INT16 · TFE + SeqMSE |
Download |
| Medium |
1344x768/qnn/visdrone-yolo26m-det-w8a8-logical-1344x768.iq9075.qnn.onnx |
Logical · INT8 · Standard |
Download |
| Medium |
1344x768/qnn/visdrone-yolo26m-det-w8a8-logical-tfe-seqmse-1344x768.iq9075.qnn.onnx |
Logical · INT8 · TFE + SeqMSE |
Download |
| Medium |
1344x768/qnn/visdrone-yolo26m-det-w8a8-smart-1344x768.iq9075.qnn.onnx |
Smart · INT8 · Standard |
Download |
| Medium |
1344x768/qnn/visdrone-yolo26m-det-w8a8-smart-tfe-seqmse-1344x768.iq9075.qnn.onnx |
Smart · INT8 · TFE + SeqMSE |
Download |
tiled/
ONNX FP32 — Any platform with ONNX Runtime.
| Size |
File |
Variant |
Download |
| Nano |
tiled/onnx/visdrone-yolo26n-det-fp16-tiled.onnx |
FP16 |
Download |
| Nano |
tiled/onnx/visdrone-yolo26n-det-fp32-tiled.onnx |
FP32 |
Download |
| Small |
tiled/onnx/visdrone-yolo26s-det-fp16-tiled.onnx |
FP16 |
Download |
| Small |
tiled/onnx/visdrone-yolo26s-det-fp32-tiled.onnx |
FP32 |
Download |
| Medium |
tiled/onnx/visdrone-yolo26m-det-fp16-tiled.onnx |
FP16 |
Download |
| Medium |
tiled/onnx/visdrone-yolo26m-det-fp32-tiled.onnx |
FP32 |
Download |
TFLite INT8 — CPU or NPU via runtime delegate (NXP i.MX 8M Plus VX Delegate).
| Size |
File |
Variant |
Download |
| Nano |
tiled/tflite/visdrone-yolo26n-det-int8-smart-tiled.tflite |
Smart · INT8 |
Download |
| Nano |
tiled/tflite/visdrone-yolo26n-det-int8-tiled.tflite |
Logical · INT8 |
Download |
| Small |
tiled/tflite/visdrone-yolo26s-det-int8-smart-tiled.tflite |
Smart · INT8 |
Download |
| Small |
tiled/tflite/visdrone-yolo26s-det-int8-tiled.tflite |
Logical · INT8 |
Download |
| Medium |
tiled/tflite/visdrone-yolo26m-det-int8-smart-tiled.tflite |
Smart · INT8 |
Download |
| Medium |
tiled/tflite/visdrone-yolo26m-det-int8-tiled.tflite |
Logical · INT8 |
Download |
NXP i.MX 95 (eIQ Neutron) — eIQ Neutron NPU optimized.
| Size |
File |
Variant |
Download |
| Nano |
tiled/imx95/visdrone-yolo26n-det-int8-smart-tiled.imx95.tflite |
Smart · INT8 |
Download |
| Nano |
tiled/imx95/visdrone-yolo26n-det-int8-tiled.imx95.tflite |
Logical · INT8 |
Download |
| Small |
tiled/imx95/visdrone-yolo26s-det-int8-smart-tiled.imx95.tflite |
Smart · INT8 |
Download |
| Small |
tiled/imx95/visdrone-yolo26s-det-int8-tiled.imx95.tflite |
Logical · INT8 |
Download |
| Medium |
tiled/imx95/visdrone-yolo26m-det-int8-smart-tiled.imx95.tflite |
Smart · INT8 |
Download |
| Medium |
tiled/imx95/visdrone-yolo26m-det-int8-tiled.imx95.tflite |
Logical · INT8 |
Download |
NXP Ara240 — Ara240 DNPU compiled model.
| Size |
File |
Variant |
Download |
| Nano |
tiled/ara240/visdrone-yolo26n-det-int16-tiled.dvm |
Smart · INT16 |
Download |
| Small |
tiled/ara240/visdrone-yolo26s-det-int16-tiled.dvm |
Smart · INT16 |
Download |
| Medium |
tiled/ara240/visdrone-yolo26m-det-int16-tiled.dvm |
Smart · INT16 |
Download |
RPi5 + Hailo-8L (13 TOPS) — Hailo-8L 13 TOPS accelerator.
| Size |
File |
Variant |
Download |
| Nano |
tiled/hailo/visdrone-yolo26n-det-int8-tiled.hailo8l.hef |
INT8 |
Download |
| Small |
tiled/hailo/visdrone-yolo26s-det-int8-tiled.hailo8l.hef |
INT8 |
Download |
| Medium |
tiled/hailo/visdrone-yolo26m-det-int8-tiled.hailo8l.hef |
INT8 |
Download |
NVIDIA Jetson (TensorRT) — Jetson FP16 and INT8 engines.
| Size |
File |
Variant |
Download |
| Nano |
tiled/jetson/visdrone-yolo26n-det-fp16-tiled.engine |
FP16 |
Download |
| Nano |
tiled/jetson/visdrone-yolo26n-det-int8-tiled.engine |
INT8 |
Download |
| Small |
tiled/jetson/visdrone-yolo26s-det-fp16-tiled.engine |
FP16 |
Download |
| Small |
tiled/jetson/visdrone-yolo26s-det-int8-tiled.engine |
INT8 |
Download |
| Medium |
tiled/jetson/visdrone-yolo26m-det-fp16-tiled.engine |
FP16 |
Download |
| Medium |
tiled/jetson/visdrone-yolo26m-det-int8-tiled.engine |
INT8 |
Download |
Qualcomm Dragonwing IQ-9075 (Hexagon v73) — Offline-compiled Hexagon context binary, run through the ONNX Runtime QNN execution provider. Loads only on the Dragonwing IQ-9075 (Hexagon v73).
| Size |
File |
Variant |
Download |
| Nano |
tiled/qnn/visdrone-yolo26n-det-w8a16-logical-in8-tiled.iq9075.qnn.onnx |
Logical · INT16 · Standard |
Download |
| Nano |
tiled/qnn/visdrone-yolo26n-det-w8a16-logical-tfe-seqmse-in8-tiled.iq9075.qnn.onnx |
Logical · INT16 · TFE + SeqMSE |
Download |
| Nano |
tiled/qnn/visdrone-yolo26n-det-w8a16-smart-in8-tiled.iq9075.qnn.onnx |
Smart · INT16 · Standard |
Download |
| Nano |
tiled/qnn/visdrone-yolo26n-det-w8a16-smart-tfe-seqmse-in8-tiled.iq9075.qnn.onnx |
Smart · INT16 · TFE + SeqMSE |
Download |
| Nano |
tiled/qnn/visdrone-yolo26n-det-w8a8-logical-tfe-seqmse-tiled.iq9075.qnn.onnx |
Logical · INT8 · TFE + SeqMSE |
Download |
| Nano |
tiled/qnn/visdrone-yolo26n-det-w8a8-logical-tiled.iq9075.qnn.onnx |
Logical · INT8 · Standard |
Download |
| Nano |
tiled/qnn/visdrone-yolo26n-det-w8a8-smart-tfe-seqmse-tiled.iq9075.qnn.onnx |
Smart · INT8 · TFE + SeqMSE |
Download |
| Nano |
tiled/qnn/visdrone-yolo26n-det-w8a8-smart-tiled.iq9075.qnn.onnx |
Smart · INT8 · Standard |
Download |
| Small |
tiled/qnn/visdrone-yolo26s-det-w8a16-logical-in8-tiled.iq9075.qnn.onnx |
Logical · INT16 · Standard |
Download |
| Small |
tiled/qnn/visdrone-yolo26s-det-w8a16-logical-tfe-seqmse-in8-tiled.iq9075.qnn.onnx |
Logical · INT16 · TFE + SeqMSE |
Download |
| Small |
tiled/qnn/visdrone-yolo26s-det-w8a16-smart-in8-tiled.iq9075.qnn.onnx |
Smart · INT16 · Standard |
Download |
| Small |
tiled/qnn/visdrone-yolo26s-det-w8a16-smart-tfe-seqmse-in8-tiled.iq9075.qnn.onnx |
Smart · INT16 · TFE + SeqMSE |
Download |
| Small |
tiled/qnn/visdrone-yolo26s-det-w8a8-logical-tfe-seqmse-tiled.iq9075.qnn.onnx |
Logical · INT8 · TFE + SeqMSE |
Download |
| Small |
tiled/qnn/visdrone-yolo26s-det-w8a8-logical-tiled.iq9075.qnn.onnx |
Logical · INT8 · Standard |
Download |
| Small |
tiled/qnn/visdrone-yolo26s-det-w8a8-smart-tfe-seqmse-tiled.iq9075.qnn.onnx |
Smart · INT8 · TFE + SeqMSE |
Download |
| Small |
tiled/qnn/visdrone-yolo26s-det-w8a8-smart-tiled.iq9075.qnn.onnx |
Smart · INT8 · Standard |
Download |
| Medium |
tiled/qnn/visdrone-yolo26m-det-w8a16-logical-in8-tiled.iq9075.qnn.onnx |
Logical · INT16 · Standard |
Download |
| Medium |
tiled/qnn/visdrone-yolo26m-det-w8a16-logical-tfe-seqmse-in8-tiled.iq9075.qnn.onnx |
Logical · INT16 · TFE + SeqMSE |
Download |
| Medium |
tiled/qnn/visdrone-yolo26m-det-w8a16-smart-in8-tiled.iq9075.qnn.onnx |
Smart · INT16 · Standard |
Download |
| Medium |
tiled/qnn/visdrone-yolo26m-det-w8a16-smart-tfe-seqmse-in8-tiled.iq9075.qnn.onnx |
Smart · INT16 · TFE + SeqMSE |
Download |
| Medium |
tiled/qnn/visdrone-yolo26m-det-w8a8-logical-tfe-seqmse-tiled.iq9075.qnn.onnx |
Logical · INT8 · TFE + SeqMSE |
Download |
| Medium |
tiled/qnn/visdrone-yolo26m-det-w8a8-logical-tiled.iq9075.qnn.onnx |
Logical · INT8 · Standard |
Download |
| Medium |
tiled/qnn/visdrone-yolo26m-det-w8a8-smart-tfe-seqmse-tiled.iq9075.qnn.onnx |
Smart · INT8 · TFE + SeqMSE |
Download |
| Medium |
tiled/qnn/visdrone-yolo26m-det-w8a8-smart-tiled.iq9075.qnn.onnx |
Smart · INT8 · Standard |
Download |
Traceability
Every measurement in the tables above is reachable through the EdgeFirst Studio validation framework. The v-XXXX Source link on each row resolves to a public Studio URL of the form:
https://edgefirst.studio/public/validation/v-XXXX/details?mode=charts
From there, the full provenance chain is one click deeper: training session ID, dataset version, calibration artifact, converter chain, validation parameters, and the host hardware description. The model file you download from this repository embeds the same chain in its edgefirst.json metadata.
See also
Limitations
- Aerial bias — VisDrone is drone footage from Chinese cities: the scenes, object scales, viewpoints and the 10 classes do not transfer to ground-level cameras without retraining.
- Quantization loss — integer quantization costs accuracy against FP32, most visibly on small objects; the Δ APs vs FP32 column shows it per platform.
- Memory limits — the 1280 and 1344x768 variants exceed some accelerators' on-chip memory or attention limits; those cells read
Not supported with the reason.
License
Model weights in this repository are derived from Ultralytics YOLO and remain © Ultralytics Inc., licensed AGPL-3.0 — use requires AGPL-3.0 compliance or an Ultralytics Enterprise License. The VisDrone dataset is © the AISKYEYE team, Tianjin University, and is subject to its own terms.
The validation results, this model card, and its metadata are Au-Zone Technologies' own contribution, licensed CC BY-NC 4.0 (Attribution — NonCommercial) — see the repository LICENSE for the full text and citation requirements.
Citation
@software{edgefirst_visdrone_yolo26_det,
title = { {YOLO26 Detection on VisDrone — EdgeFirst Model Zoo} },
author = {Au-Zone Technologies},
url = {https://huggingface.co/EdgeFirst/visdrone-yolo26-det},
year = {2026},
license = {CC-BY-NC-4.0},
}
EdgeFirst Studio · GitHub · Docs · Au-Zone Technologies
Model weights © Ultralytics Inc. (AGPL-3.0) · Validation results & card © 2026 Au-Zone Technologies (CC BY-NC 4.0)
NXP®, i.MX, eIQ®, Neutron, and Ara240 are trademarks or products of NXP Semiconductors. Hailo is a trademark of Hailo Technologies Ltd. Jetson is a trademark of NVIDIA Corporation. Qualcomm and Dragonwing are trademarks of Qualcomm Incorporated. All other trademarks are the property of their respective owners.
