EdgeFirst/visdrone-yolov8-det

YOLOv8 Detection on VisDrone — EdgeFirst Model Zoo

YOLOv8 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/yolov8-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	3.2M	10.14%	18.60%	32.57%	18.11%	27.89%	45.53%	31.12%	v-1966
640	Small	11.2M	13.78%	22.28%	38.47%	21.95%	32.13%	45.41%	35.20%	v-196a
640	Medium	25.9M	15.89%	25.14%	42.14%	25.71%	36.96%	46.11%	37.97%	v-196e
1280	Nano	3.2M	20.02%	29.41%	48.26%	29.97%	40.91%	49.15%	44.38%	v-1967
1280	Small	11.2M	24.05%	32.77%	53.01%	33.97%	44.22%	51.08%	47.64%	v-196b
1280	Medium	25.9M	26.69%	35.88%	56.96%	37.44%	47.77%	57.21%	50.66%	v-1970
1344x768	Nano	3.2M	21.60%	30.29%	49.23%	30.87%	41.61%	53.41%	46.09%	v-1969
1344x768	Small	11.2M	25.91%	34.50%	55.59%	35.61%	46.20%	53.02%	49.93%	v-196d
1344x768	Medium	25.9M	28.46%	36.89%	58.87%	38.61%	48.80%	58.07%	51.96%	v-1972
tiled	Nano	3.2M	19.98%	28.48%	47.82%	28.65%	39.52%	48.53%	43.71%	v-1968
tiled	Small	11.2M	23.69%	32.42%	53.39%	33.52%	44.01%	47.30%	47.42%	v-196c
tiled	Medium	25.9M	26.32%	34.69%	56.34%	36.16%	46.29%	53.24%	49.84%	v-1971

---

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	—	10.14%	ref	18.60%	32.57%	9.51	335.9	v-1966
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	10.03%	-0.11	18.53%	32.50%	7.72	348.5	v-19bd
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	3.18%	-6.96	10.18%	21.10%	288.75	4.6	v-1b70
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	3.18%	-6.96	10.18%	21.10%	75.28	14.9	v-1a7c
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	3.18%	-6.96	10.18%	21.10%	75.32	14.9	v-1a7e
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	8.65%	-1.49	16.74%	29.91%	85.12	16.2	v-1a7f
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	3.48%	-6.66	10.74%	21.95%	91.92	14.8	v-1bf4
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	8.99%	-1.15	17.28%	31.03%	48.22	47.2	v-1702
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	8.99%	-1.15	17.28%	31.03%	49.56	60.3	v-19d7
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	8.70%	-1.44	16.91%	30.35%	34.80	50.6	v-1a01
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	2.61%	-7.53	9.46%	19.93%	41.66	29.7	v-19fd
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	3.10%	-7.04	10.29%	21.30%	42.88	28.6	v-19fb
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	8.70%	-1.44	16.91%	30.35%	53.73	80.1	v-1a03
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	2.61%	-7.53	9.46%	19.93%	60.76	75.7	v-19ff
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	3.10%	-7.04	10.29%	21.30%	58.82	81.8	v-19fc
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	8.70%	-1.44	16.91%	30.35%	53.33	78.7	v-175d
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	8.99%	-1.15	17.28%	31.03%	50.42	45.7	v-16f5
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	8.99%	-1.15	17.28%	31.03%	54.16	53.8	v-19ea
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	5.41%	-4.73	13.58%	25.94%	17.51	104.4	v-159d
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	5.40%	-4.74	13.58%	25.93%	18.52	191.4	v-1901
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	9.34%	-0.80	17.67%	31.78%	17.85	76.5	v-15d9
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	10.30%	+0.16	18.70%	32.87%	34.48	183.1	v-15b0
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	8.10%	-2.04	16.51%	30.20%	32.51	187.8	v-15b1
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	3.84%	-6.30	11.44%	22.90%	11.29	499.1	v-1536
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	9.20%	-0.94	17.58%	31.49%	10.56	295.9	v-1537
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	10.13%	-0.01	18.58%	32.55%	16.90	254.3	v-1539
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	9.01%	-1.13	17.19%	30.76%	10.82	298.2	v-1d2c
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	3.37%	-6.77	10.67%	21.68%	11.29	490.0	v-1d2d
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	9.87%	-0.27	18.42%	32.37%	17.91	249.0	v-1d2e
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	9.54%	-0.60	17.97%	32.13%	13.01	247.2	v-1d30
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	9.87%	-0.27	18.42%	32.43%	11.19	248.5	v-1d2f
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	13.78%	ref	22.28%	38.47%	11.62	244.6	v-196a
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	13.68%	-0.10	22.18%	38.37%	9.32	308.2	v-19c2
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	4.52%	-9.26	12.13%	25.46%	356.61	4.4	v-1b9b
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	4.52%	-9.26	12.14%	25.46%	131.25	8.1	v-1b0c
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	4.52%	-9.26	12.14%	25.46%	131.30	8.1	v-1b14
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	12.12%	-1.66	20.40%	35.98%	139.76	8.5	v-1b17
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	4.74%	-9.04	12.82%	26.59%	148.14	8.0	v-1c0f
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	12.19%	-1.59	20.72%	36.75%	68.58	29.2	v-171d
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	12.19%	-1.59	20.72%	36.75%	159.31	31.7	v-19ed
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	11.87%	-1.91	20.08%	35.94%	56.54	28.5	v-1a6f
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	4.60%	-9.18	12.32%	25.35%	66.08	17.0	v-1a6d
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	4.59%	-9.19	12.31%	25.32%	66.81	16.8	v-1a6a
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	11.87%	-1.91	20.08%	35.94%	148.73	31.6	v-1a70
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	4.60%	-9.18	12.31%	25.35%	134.78	30.5	v-1a6e
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	4.59%	-9.19	12.31%	25.32%	133.84	31.7	v-1a6b
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	11.87%	-1.91	20.08%	35.94%	148.93	31.6	v-1874
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	12.19%	-1.59	20.72%	36.75%	69.07	29.1	v-172d
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	12.19%	-1.59	20.72%	36.75%	160.56	31.4	v-19f4
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	7.50%	-6.28	16.05%	30.51%	21.20	83.3	v-15b7
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	7.50%	-6.28	16.05%	30.51%	19.54	144.1	v-1909
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	12.73%	-1.05	20.81%	36.75%	38.71	29.3	v-15f5
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	13.94%	+0.16	22.55%	38.85%	42.41	130.9	v-15bc
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	10.71%	-3.07	19.58%	34.84%	31.09	182.5	v-15be
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	5.21%	-8.57	13.08%	26.97%	11.60	432.0	v-1552
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	12.89%	-0.89	21.38%	37.52%	12.25	324.6	v-1553
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	13.79%	+0.01	22.22%	38.32%	21.40	200.6	v-1554
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	12.66%	-1.12	20.95%	36.87%	11.70	300.5	v-1d6a
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	4.94%	-8.84	12.36%	25.84%	12.93	431.6	v-1d6b
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	13.57%	-0.21	22.07%	38.18%	21.24	200.7	v-1d6c
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	12.83%	-0.95	21.37%	37.84%	14.43	187.4	v-1d6e
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	13.57%	-0.21	22.05%	38.22%	15.51	187.1	v-1d6d
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	15.89%	ref	25.14%	42.14%	20.56	165.7	v-196e
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	15.79%	-0.10	25.02%	41.99%	13.73	214.1	v-19c8
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	4.10%	-11.79	12.22%	24.93%	488.89	3.9	v-1bd5
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	4.10%	-11.79	12.22%	24.93%	235.60	4.4	v-1b33
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	4.10%	-11.79	12.22%	24.93%	235.67	4.4	v-1b37
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	11.36%	-4.53	20.77%	37.19%	244.22	4.5	v-1b40
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	12.14%	-3.75	21.76%	38.93%	131.54	12.0	v-1734
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	12.14%	-3.75	21.76%	38.93%	365.92	12.4	v-19f7
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	11.82%	-4.07	21.28%	38.27%	103.40	11.9	v-1a95
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	4.12%	-11.77	12.49%	25.62%	114.55	9.2	v-1a92
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	4.12%	-11.77	12.49%	25.62%	114.53	9.2	v-1a8e
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	11.82%	-4.07	21.29%	38.27%	340.52	12.5	v-1a97
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	4.12%	-11.77	12.49%	25.61%	329.44	12.2	v-1a94
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	4.12%	-11.77	12.49%	25.62%	327.42	12.5	v-1a90
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	11.82%	-4.07	21.29%	38.27%	341.33	12.5	v-187d
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	12.14%	-3.75	21.76%	38.93%	132.43	12.0	v-1746
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	12.14%	-3.75	21.76%	38.93%	364.79	12.4	v-1a06
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	6.29%	-9.60	15.42%	29.99%	32.24	44.0	v-15d6
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	6.29%	-9.60	15.42%	29.99%	28.96	54.2	v-1917
Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	13.23%	-2.66	22.51%	39.62%	66.69	16.2	v-160c
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	16.05%	+0.16	25.47%	42.76%	65.64	69.7	v-192b
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	11.19%	-4.70	20.82%	36.89%	49.64	102.6	v-192d
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	5.87%	-10.02	14.74%	29.57%	12.04	247.6	v-1581
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	13.90%	-1.99	23.61%	40.94%	15.95	259.6	v-1582
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	15.81%	-0.08	25.16%	42.14%	26.41	109.2	v-1583
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	12.60%	-3.29	22.04%	39.48%	15.06	259.7	v-1dc6
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	4.90%	-10.99	13.42%	27.77%	12.45	248.7	v-1dc7
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	15.37%	-0.52	24.85%	41.98%	29.44	108.7	v-1dc8
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	14.89%	-1.00	24.44%	41.70%	26.68	90.9	v-1dca
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	15.40%	-0.49	24.82%	41.96%	26.95	90.0	v-1dc9

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	10.14%	18.60%	32.57%	18.11%	27.89%	45.53%	8.64%	23.31%	31.12%	21.21%	44.61%	63.16%
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	10.03%	18.53%	32.50%	18.03%	27.86%	45.53%	8.61%	23.20%	31.00%	21.02%	44.63%	62.99%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	3.18%	10.18%	21.10%	8.83%	16.59%	36.45%	5.89%	15.47%	21.27%	11.47%	34.62%	48.75%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	3.18%	10.18%	21.10%	8.83%	16.59%	36.45%	5.89%	15.47%	21.27%	11.47%	34.62%	48.75%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	3.18%	10.18%	21.10%	8.83%	16.59%	36.45%	5.89%	15.47%	21.27%	11.47%	34.62%	48.75%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	8.65%	16.74%	29.91%	16.14%	25.31%	40.37%	7.90%	21.42%	28.62%	19.23%	41.25%	52.89%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	INT8	3.48%	10.74%	21.95%	9.35%	17.59%	40.41%	6.19%	16.14%	22.23%	12.10%	36.38%	53.92%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	8.99%	17.28%	31.03%	16.74%	26.40%	43.56%	8.08%	22.17%	29.67%	19.68%	43.49%	59.34%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	8.99%	17.28%	31.03%	16.74%	26.40%	43.56%	8.08%	22.17%	29.67%	19.68%	43.49%	59.34%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	8.70%	16.91%	30.35%	16.40%	25.77%	40.71%	8.07%	21.74%	29.00%	19.63%	42.09%	53.02%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	2.61%	9.46%	19.93%	8.03%	16.09%	35.92%	5.75%	14.75%	20.25%	10.83%	33.08%	46.78%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	3.10%	10.29%	21.30%	8.92%	17.10%	36.77%	6.10%	15.71%	21.57%	11.59%	35.55%	48.85%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	8.70%	16.91%	30.35%	16.40%	25.77%	40.71%	8.07%	21.74%	29.00%	19.63%	42.09%	53.02%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	2.61%	9.46%	19.93%	8.03%	16.09%	35.92%	5.75%	14.75%	20.25%	10.83%	33.08%	46.78%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	3.10%	10.29%	21.30%	8.91%	17.10%	36.77%	6.10%	15.71%	21.57%	11.59%	35.55%	48.85%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	8.70%	16.91%	30.35%	16.40%	25.77%	40.71%	8.07%	21.74%	29.00%	19.63%	42.09%	53.02%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	8.99%	17.28%	31.03%	16.74%	26.40%	43.56%	8.08%	22.17%	29.67%	19.68%	43.49%	59.34%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	8.99%	17.28%	31.03%	16.74%	26.40%	43.56%	8.08%	22.17%	29.67%	19.68%	43.49%	59.34%
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	5.41%	13.58%	25.94%	12.56%	21.83%	38.40%	7.02%	18.42%	25.07%	15.79%	37.46%	49.12%
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	5.40%	13.58%	25.93%	12.56%	21.83%	38.41%	7.02%	18.42%	25.07%	15.79%	37.46%	49.12%
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	9.34%	17.67%	31.78%	17.05%	26.88%	43.82%	8.31%	22.63%	30.22%	20.32%	43.58%	57.91%
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	10.30%	18.70%	32.87%	18.14%	28.02%	45.20%	8.63%	23.68%	31.47%	21.46%	45.17%	64.34%
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	8.10%	16.51%	30.20%	15.83%	25.53%	44.32%	7.97%	21.22%	28.59%	18.93%	41.88%	60.85%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	3.84%	11.44%	22.90%	10.27%	18.82%	42.05%	6.37%	16.77%	22.95%	12.71%	37.43%	55.99%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	9.20%	17.58%	31.49%	16.92%	26.78%	45.19%	8.16%	22.27%	29.72%	20.09%	43.10%	61.16%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	10.13%	18.58%	32.55%	18.09%	27.85%	44.72%	8.61%	23.38%	31.09%	21.32%	44.60%	60.01%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	9.01%	17.19%	30.76%	16.74%	26.04%	44.90%	7.99%	21.91%	29.35%	19.71%	42.41%	59.72%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	3.37%	10.67%	21.68%	9.50%	17.69%	39.36%	6.01%	16.12%	22.23%	12.56%	35.61%	52.03%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	9.87%	18.42%	32.37%	17.82%	27.84%	45.10%	8.61%	23.23%	30.84%	20.92%	44.55%	61.05%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	9.54%	17.97%	32.13%	17.21%	27.19%	44.09%	8.41%	22.62%	30.17%	20.57%	43.02%	58.68%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	9.87%	18.42%	32.43%	17.81%	27.84%	45.06%	8.61%	23.20%	30.81%	20.87%	44.48%	61.30%
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	13.78%	22.28%	38.47%	21.95%	32.13%	45.41%	9.64%	26.73%	35.20%	26.38%	47.46%	57.38%
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	13.68%	22.18%	38.37%	21.90%	32.10%	45.16%	9.62%	26.63%	35.10%	26.29%	47.44%	57.34%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	4.52%	12.13%	25.46%	10.38%	19.90%	38.71%	6.72%	17.71%	24.18%	14.37%	37.69%	50.91%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	4.52%	12.14%	25.46%	10.38%	19.90%	38.72%	6.72%	17.71%	24.18%	14.37%	37.69%	50.91%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	4.52%	12.14%	25.46%	10.38%	19.90%	38.72%	6.72%	17.71%	24.18%	14.37%	37.69%	50.91%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	12.12%	20.40%	35.98%	19.94%	29.58%	42.70%	9.03%	24.94%	33.08%	24.23%	45.29%	55.23%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	INT8	4.74%	12.82%	26.59%	11.18%	21.31%	41.50%	7.06%	18.70%	25.64%	15.24%	40.33%	56.04%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	12.19%	20.72%	36.75%	20.29%	30.60%	44.30%	9.31%	25.94%	34.40%	25.14%	47.45%	60.26%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	12.19%	20.72%	36.75%	20.29%	30.60%	44.30%	9.31%	25.94%	34.40%	25.14%	47.45%	60.26%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	11.87%	20.08%	35.94%	19.61%	29.69%	40.79%	9.23%	25.00%	33.28%	24.59%	45.60%	54.83%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	4.60%	12.32%	25.35%	10.93%	20.02%	37.81%	7.01%	18.11%	24.67%	14.92%	38.69%	51.58%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	4.59%	12.31%	25.32%	10.90%	20.01%	37.78%	6.99%	18.10%	24.66%	14.94%	38.64%	51.59%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	11.87%	20.08%	35.94%	19.61%	29.69%	40.79%	9.23%	25.00%	33.28%	24.59%	45.60%	54.83%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	4.60%	12.31%	25.35%	10.93%	20.02%	37.81%	7.01%	18.11%	24.67%	14.92%	38.69%	51.58%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	4.59%	12.31%	25.32%	10.90%	20.01%	37.78%	6.99%	18.10%	24.66%	14.94%	38.64%	51.59%
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	11.87%	20.08%	35.94%	19.61%	29.69%	40.79%	9.23%	25.00%	33.28%	24.59%	45.60%	54.83%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	12.19%	20.72%	36.75%	20.29%	30.60%	44.30%	9.31%	25.94%	34.40%	25.14%	47.45%	60.26%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	12.19%	20.72%	36.75%	20.29%	30.60%	44.30%	9.31%	25.94%	34.40%	25.14%	47.45%	60.26%
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	7.50%	16.05%	30.51%	15.10%	25.21%	41.77%	8.09%	21.36%	29.04%	20.29%	41.04%	53.43%
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	7.50%	16.05%	30.51%	15.10%	25.21%	41.77%	8.09%	21.36%	29.04%	20.29%	41.04%	53.43%
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	12.73%	20.81%	36.75%	20.43%	30.65%	42.78%	9.23%	25.42%	33.67%	24.83%	46.09%	56.71%
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	13.94%	22.55%	38.85%	22.52%	32.64%	46.24%	9.86%	27.04%	35.71%	26.79%	48.59%	58.92%
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	10.71%	19.58%	34.84%	19.28%	29.29%	44.70%	9.04%	24.66%	33.15%	24.11%	45.75%	56.72%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	5.21%	13.08%	26.97%	11.33%	20.96%	39.71%	7.07%	18.62%	25.35%	15.85%	38.89%	52.31%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	12.89%	21.38%	37.52%	21.05%	31.21%	44.31%	9.41%	25.85%	34.28%	25.45%	46.63%	56.10%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	13.79%	22.22%	38.32%	21.89%	32.01%	45.15%	9.63%	26.75%	35.22%	26.56%	47.30%	56.66%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	12.66%	20.95%	36.87%	20.73%	30.67%	43.06%	9.21%	25.52%	34.06%	25.41%	46.15%	54.78%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	4.94%	12.36%	25.84%	10.74%	20.53%	38.78%	6.72%	18.09%	24.80%	15.09%	38.59%	49.95%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	13.57%	22.07%	38.18%	21.64%	31.94%	45.07%	9.56%	26.51%	34.89%	26.15%	47.09%	56.76%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	12.83%	21.37%	37.84%	20.68%	31.36%	44.12%	9.39%	25.73%	33.94%	25.26%	45.62%	55.00%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	13.57%	22.05%	38.22%	21.57%	31.91%	45.02%	9.55%	26.49%	34.86%	26.14%	47.05%	56.81%
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	15.89%	25.14%	42.14%	25.71%	36.96%	46.11%	10.60%	29.12%	37.97%	29.14%	50.93%	59.24%
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	15.79%	25.02%	41.99%	25.60%	36.76%	46.14%	10.60%	28.99%	37.78%	28.92%	50.84%	59.23%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	4.10%	12.22%	24.93%	10.98%	21.13%	39.82%	6.97%	18.06%	24.81%	14.52%	39.79%	51.33%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	4.10%	12.22%	24.93%	10.98%	21.13%	39.82%	6.97%	18.06%	24.81%	14.52%	39.79%	51.33%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	4.10%	12.22%	24.93%	10.98%	21.13%	39.82%	6.97%	18.06%	24.81%	14.52%	39.79%	51.33%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	11.36%	20.77%	37.19%	20.37%	32.05%	44.38%	9.29%	25.41%	33.84%	24.54%	47.15%	55.96%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	12.14%	21.76%	38.93%	21.26%	33.75%	49.14%	9.66%	26.51%	35.08%	25.61%	49.30%	63.79%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	12.14%	21.76%	38.93%	21.26%	33.75%	49.14%	9.66%	26.51%	35.08%	25.61%	49.30%	63.79%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	11.82%	21.28%	38.27%	20.93%	32.92%	44.32%	9.59%	25.70%	34.25%	25.26%	47.27%	55.81%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	4.12%	12.49%	25.62%	11.19%	21.73%	39.55%	7.04%	18.23%	25.00%	14.91%	39.92%	52.48%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	4.12%	12.49%	25.62%	11.17%	21.74%	39.55%	7.04%	18.23%	25.00%	14.92%	39.89%	52.48%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	11.82%	21.29%	38.27%	20.93%	32.93%	44.32%	9.59%	25.70%	34.25%	25.26%	47.27%	55.81%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	4.12%	12.49%	25.61%	11.19%	21.73%	39.55%	7.04%	18.23%	25.00%	14.91%	39.92%	52.48%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	4.12%	12.49%	25.62%	11.17%	21.74%	39.55%	7.04%	18.23%	25.00%	14.92%	39.89%	52.48%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	11.82%	21.29%	38.27%	20.93%	32.92%	44.32%	9.59%	25.70%	34.25%	25.26%	47.27%	55.81%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	12.14%	21.76%	38.93%	21.26%	33.75%	49.14%	9.66%	26.51%	35.08%	25.61%	49.30%	63.79%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	12.14%	21.76%	38.93%	21.26%	33.75%	49.14%	9.66%	26.51%	35.08%	25.61%	49.30%	63.79%
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	6.29%	15.42%	29.99%	14.17%	25.92%	41.18%	7.94%	21.30%	29.05%	19.28%	42.77%	54.83%
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	6.29%	15.42%	29.99%	14.17%	25.92%	41.18%	7.94%	21.30%	29.05%	19.28%	42.77%	54.83%
Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	13.23%	22.51%	39.62%	22.11%	34.05%	44.52%	9.65%	26.69%	35.14%	26.20%	48.26%	55.31%
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	16.05%	25.47%	42.76%	25.94%	37.61%	49.36%	10.72%	29.55%	38.53%	29.33%	52.38%	64.90%
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	11.19%	20.82%	36.89%	20.88%	32.50%	46.95%	9.32%	25.81%	33.95%	24.64%	48.18%	58.01%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	5.87%	14.74%	29.57%	13.44%	24.47%	41.23%	7.85%	20.15%	27.12%	16.78%	42.47%	54.69%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	13.90%	23.61%	40.94%	23.47%	35.85%	45.72%	10.15%	27.65%	36.19%	26.91%	50.00%	58.50%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	15.81%	25.16%	42.14%	25.70%	36.99%	45.68%	10.58%	29.06%	37.89%	28.99%	51.03%	59.46%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	12.60%	22.04%	39.48%	21.51%	33.54%	44.81%	9.66%	26.53%	34.91%	25.77%	48.02%	60.55%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	4.90%	13.42%	27.77%	11.82%	23.04%	41.07%	7.22%	19.17%	26.06%	15.83%	41.02%	53.92%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	15.37%	24.85%	41.98%	25.24%	36.97%	45.61%	10.58%	28.85%	37.50%	28.42%	51.08%	59.30%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	14.89%	24.44%	41.70%	24.63%	36.39%	45.15%	10.36%	28.12%	36.76%	27.89%	49.29%	57.71%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	15.40%	24.82%	41.96%	25.26%	36.90%	45.69%	10.56%	28.83%	37.50%	28.43%	51.04%	59.40%

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	—	20.02%	ref	29.41%	48.26%	22.94	136.6	v-1967
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	19.77%	-0.25	29.18%	48.03%	15.76	192.2	v-19be
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	5.64%	-14.38	14.06%	28.19%	1157.15	1.1	v-1b77
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	5.64%	-14.38	14.06%	28.19%	290.24	3.6	v-1a81
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	5.64%	-14.38	14.06%	28.19%	289.64	3.6	v-1a84
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	18.35%	-1.67	27.22%	45.40%	300.10	4.1	v-1a88
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	5.79%	-14.23	14.47%	28.85%	331.43	3.6	v-1bfa
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	18.54%	-1.48	27.45%	45.80%	133.21	16.8	v-1704
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	18.54%	-1.48	27.45%	45.80%	278.69	19.1	v-19d9
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	18.56%	-1.46	27.51%	46.02%	106.92	16.6	v-1a0c
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	5.78%	-14.24	14.36%	28.77%	165.30	6.7	v-1a08
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	5.78%	-14.24	14.36%	28.77%	166.46	6.7	v-1a04
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	18.56%	-1.46	27.52%	46.02%	260.01	18.8	v-1a0d
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	5.78%	-14.24	14.36%	28.77%	241.56	17.7	v-1a0b
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	5.78%	-14.24	14.36%	28.77%	243.54	17.6	v-1a07
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	18.56%	-1.46	27.52%	46.02%	260.43	18.9	v-1865
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	18.54%	-1.48	27.45%	45.80%	133.05	16.5	v-16f8
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	18.54%	-1.48	27.45%	45.80%	285.73	18.8	v-19eb
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	9.72%	-10.30 ⚠	19.33%	34.71%	50.71	36.5	v-161a
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	9.72%	-10.30 ⚠	19.33%	34.71%	40.28	70.6	v-1902
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	19.21%	-0.81	28.63%	47.37%	45.30	31.6	v-15dc
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	19.83%	-0.19	29.26%	48.17%	75.01	77.5	v-15b2
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	18.11%	-1.91	27.21%	46.06%	66.78	103.5	v-15b3
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	6.40%	-13.62	16.33%	31.09%	18.79	198.6	v-153a
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	19.58%	-0.44	28.89%	47.82%	24.70	129.3	v-153b
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	19.96%	-0.06	29.32%	48.21%	54.46	65.4	v-153e
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	18.83%	-1.19	28.03%	46.68%	24.54	128.8	v-1d31
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	5.48%	-14.54	14.90%	29.25%	16.31	202.6	v-1d32
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	19.88%	-0.14	29.29%	48.19%	55.00	65.2	v-1d33
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	19.51%	-0.51	29.06%	47.98%	43.07	59.3	v-1d35
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	19.91%	-0.11	29.31%	48.24%	42.11	59.5	v-1d34
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	24.05%	ref	32.77%	53.01%	35.27	105.0	v-196b
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	23.75%	-0.30	32.51%	52.80%	23.11	148.4	v-19c4
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	6.04%	-18.01	13.70%	29.55%	1438.82	1.1	v-1ba4
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	6.04%	-18.01	13.70%	29.55%	523.13	2.0	v-1b1b
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	6.04%	-18.01	13.70%	29.55%	522.13	2.0	v-1b21
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	19.87%	-4.18	28.64%	48.38%	529.62	2.1	v-1b29
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	6.18%	-17.87	13.92%	30.06%	568.80	1.9	v-1c17
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	20.18%	-3.87	28.71%	48.49%	251.50	6.6	v-171f
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	20.18%	-3.87	28.71%	48.49%	677.25	6.9	v-19ee
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	20.17%	-3.88	28.64%	48.39%	194.44	6.6	v-1a76
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	6.11%	-17.94	13.50%	29.37%	251.91	4.2	v-1a74
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	6.11%	-17.94	13.50%	29.37%	256.22	4.2	v-1a72
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	20.17%	-3.88	28.64%	48.39%	618.73	6.9	v-1a77
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	6.11%	-17.94	13.50%	29.38%	597.36	6.8	v-1a75
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	6.11%	-17.94	13.50%	29.37%	600.24	6.8	v-1a73
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	20.17%	-3.88	28.64%	48.39%	619.15	6.9	v-1875
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	20.17%	-3.88	28.70%	48.49%	254.00	6.6	v-1730
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	20.17%	-3.88	28.70%	48.49%	661.10	6.9	v-19f5
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	10.90%	-13.15 ⚠	18.76%	35.33%	74.28	19.6	v-15b9
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	10.90%	-13.15 ⚠	18.76%	35.33%	61.68	25.3	v-190b
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	20.36%	-3.69	29.41%	49.77%	82.63	15.0	v-15fa
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	23.91%	-0.14	32.70%	52.99%	105.93	45.2	v-15c0
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	19.52%	-4.53	28.62%	48.19%	71.01	76.5	v-15c2
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	6.83%	-17.22	16.66%	32.58%	24.61	118.2	v-1556
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	22.43%	-1.62	31.53%	51.87%	33.15	136.0	v-1558
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	24.16%	+0.11	32.78%	53.00%	70.25	47.6	v-155a
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	20.75%	-3.30	29.47%	49.53%	32.20	134.7	v-1d6f
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	6.26%	-17.79	14.37%	30.62%	22.58	118.5	v-1d70
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	24.03%	-0.02	32.72%	52.94%	71.17	46.7	v-1d71
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	23.58%	-0.47	32.43%	52.73%	54.10	43.2	v-1d73
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	23.45%	-0.60	32.33%	52.77%	54.81	42.7	v-1d72
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	26.69%	ref	35.88%	56.96%	83.58	53.6	v-1970
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	26.38%	-0.31	35.60%	56.84%	55.49	86.7	v-19c9
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	5.20%	-21.49	10.24%	23.26%	2004.13	1.0	v-1bef
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	5.25%	-21.44	10.24%	23.33%	973.73	1.0	v-1b43
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	5.20%	-21.49	10.24%	23.25%	971.71	1.0	v-1b4f
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	14.37%	-12.32 ⚠	23.45%	42.67%	980.86	1.1	v-1b62
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	14.47%	-12.22 ⚠	23.88%	43.07%	448.49	2.9	v-1738
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	14.52%	-12.17 ⚠	23.87%	43.09%	1453.99	3.0	v-19fa
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	14.80%	-11.89 ⚠	23.64%	43.08%	382.81	2.9	v-1aa1
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	5.40%	-21.29	10.70%	24.01%	448.90	2.3	v-1a9b
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	5.45%	-21.24	10.73%	24.08%	445.83	2.3	v-1a98
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	14.83%	-11.86 ⚠	23.73%	43.14%	1367.38	3.0	v-1aa3
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	5.39%	-21.30	10.69%	24.04%	1349.23	3.0	v-1a9f
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	5.43%	-21.26	10.77%	24.16%	1345.29	3.0	v-1a9a
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	14.76%	-11.93 ⚠	23.70%	43.14%	1368.00	3.0	v-187f
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	14.57%	-12.12 ⚠	23.87%	43.19%	446.33	2.9	v-1777
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	14.56%	-12.13 ⚠	23.90%	43.15%	1433.30	3.0	v-1a09
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	9.06%	-17.63 ⚠	14.31%	29.11%	132.93	9.2	v-15e1
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	9.11%	-17.58 ⚠	14.34%	29.11%	121.74	10.2	v-1920
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	26.41%	-0.28	35.64%	56.92%	202.80	21.4	v-192f
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	11.59%	-15.10 ⚠	19.81%	36.20%	134.61	35.4	v-1933
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	8.22%	-18.47	18.16%	35.18%	36.27	66.0	v-15bd
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	25.33%	-1.36	34.52%	55.94%	53.66	62.3	v-15bf
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	26.76%	+0.07	35.89%	56.97%	113.09	24.5	v-15c3
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	19.05%	-7.64	29.16%	49.28%	54.01	63.4	v-1dad
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	5.78%	-20.91	13.85%	28.68%	37.57	65.7	v-1dae
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	26.72%	+0.03	35.86%	56.89%	113.69	24.3	v-1daf
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	26.14%	-0.55	35.52%	56.73%	92.09	23.6	v-1db1
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	26.69%	+0.00	35.80%	56.91%	91.18	23.7	v-1db0

Sizes	Platform	Runtime	Decode	Status
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	20.02%	29.41%	48.26%	29.97%	40.91%	49.15%	11.54%	33.66%	44.38%	35.81%	56.60%	64.29%
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	19.77%	29.18%	48.03%	29.80%	40.76%	49.15%	11.47%	33.41%	44.07%	35.39%	56.48%	64.34%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	5.64%	14.06%	28.19%	12.49%	23.31%	39.10%	7.84%	20.67%	28.71%	18.80%	43.17%	52.14%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	5.64%	14.06%	28.19%	12.49%	23.31%	39.10%	7.84%	20.67%	28.71%	18.80%	43.17%	52.14%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	5.64%	14.06%	28.19%	12.49%	23.31%	39.10%	7.84%	20.67%	28.71%	18.80%	43.17%	52.14%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	18.35%	27.22%	45.40%	27.56%	37.91%	43.79%	11.14%	32.01%	42.25%	33.88%	53.86%	60.18%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	INT8	5.79%	14.47%	28.85%	12.67%	24.20%	43.16%	7.71%	20.94%	29.24%	19.26%	44.04%	57.98%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	18.54%	27.45%	45.80%	27.81%	38.40%	45.42%	11.02%	31.91%	42.19%	33.78%	53.94%	61.37%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	18.54%	27.45%	45.80%	27.81%	38.40%	45.42%	11.02%	31.91%	42.19%	33.78%	53.94%	61.37%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	18.56%	27.51%	46.02%	27.90%	38.46%	47.10%	11.12%	31.97%	42.25%	34.04%	53.77%	60.56%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	5.78%	14.36%	28.77%	12.51%	23.82%	44.12%	7.90%	20.89%	29.09%	19.27%	43.67%	55.96%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	5.78%	14.36%	28.77%	12.51%	23.82%	44.12%	7.90%	20.89%	29.09%	19.27%	43.67%	55.96%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	18.56%	27.52%	46.02%	27.90%	38.47%	47.10%	11.12%	31.97%	42.25%	34.04%	53.77%	60.56%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	5.78%	14.36%	28.77%	12.51%	23.82%	44.12%	7.90%	20.89%	29.09%	19.27%	43.67%	55.96%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	5.78%	14.36%	28.77%	12.50%	23.82%	44.12%	7.90%	20.89%	29.09%	19.27%	43.67%	55.96%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	18.56%	27.52%	46.02%	27.90%	38.47%	47.10%	11.12%	31.97%	42.25%	34.04%	53.77%	60.56%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	18.54%	27.45%	45.80%	27.81%	38.40%	45.42%	11.02%	31.91%	42.19%	33.78%	53.94%	61.37%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	18.54%	27.45%	45.80%	27.81%	38.40%	45.42%	11.02%	31.91%	42.19%	33.78%	53.94%	61.37%
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	9.72%	19.33%	34.71%	19.53%	30.57%	42.75%	8.92%	24.72%	34.48%	25.94%	46.14%	54.73%
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	9.72%	19.33%	34.71%	19.53%	30.56%	43.53%	8.92%	24.72%	34.48%	25.94%	46.14%	54.73%
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	19.21%	28.63%	47.37%	29.01%	40.30%	46.43%	11.33%	32.67%	43.15%	34.18%	55.76%	61.39%
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	19.83%	29.26%	48.17%	29.85%	40.89%	49.08%	11.51%	33.51%	44.18%	35.38%	56.68%	64.09%
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	18.11%	27.21%	46.06%	27.44%	38.20%	46.76%	10.89%	32.03%	42.62%	33.72%	55.34%	63.63%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	6.40%	16.33%	31.09%	15.48%	27.47%	44.83%	8.41%	22.46%	31.27%	20.87%	46.75%	56.85%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	19.58%	28.89%	47.82%	29.23%	40.32%	48.23%	11.42%	33.24%	43.85%	35.09%	56.38%	62.74%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	19.96%	29.32%	48.21%	29.88%	40.74%	49.32%	11.45%	33.57%	44.31%	35.73%	56.54%	63.06%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	18.83%	28.03%	46.68%	28.68%	39.07%	46.60%	11.23%	32.59%	43.15%	34.47%	55.23%	61.98%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	5.48%	14.90%	29.25%	13.45%	25.77%	41.80%	7.97%	21.33%	29.57%	19.40%	44.44%	56.35%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	19.88%	29.29%	48.19%	29.75%	40.82%	49.06%	11.41%	33.55%	44.27%	35.56%	56.70%	63.46%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	19.51%	29.06%	47.98%	29.59%	40.72%	49.39%	11.36%	33.10%	43.72%	35.10%	55.82%	62.72%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	19.91%	29.31%	48.24%	29.75%	40.74%	48.94%	11.41%	33.56%	44.28%	35.58%	56.66%	63.25%
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	24.05%	32.77%	53.01%	33.97%	44.22%	51.08%	12.62%	36.31%	47.64%	39.91%	58.82%	66.68%
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	23.75%	32.51%	52.80%	33.59%	44.05%	50.85%	12.58%	36.07%	47.35%	39.56%	58.58%	66.14%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	6.04%	13.70%	29.55%	10.75%	24.40%	46.48%	7.73%	21.40%	30.48%	20.67%	45.08%	55.31%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	6.04%	13.70%	29.55%	10.75%	24.40%	46.49%	7.73%	21.40%	30.48%	20.67%	45.08%	55.31%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	6.04%	13.70%	29.55%	10.75%	24.40%	46.49%	7.73%	21.40%	30.48%	20.67%	45.08%	55.31%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	19.87%	28.64%	48.38%	29.18%	39.77%	51.38%	11.52%	33.65%	45.11%	37.13%	56.22%	63.03%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	INT8	6.18%	13.92%	30.06%	10.71%	24.82%	43.97%	7.73%	21.51%	30.53%	20.73%	45.30%	55.94%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	20.18%	28.71%	48.49%	29.27%	39.93%	46.68%	11.56%	33.92%	45.50%	37.67%	56.72%	63.13%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	20.18%	28.71%	48.49%	29.27%	39.93%	46.68%	11.56%	33.92%	45.50%	37.67%	56.72%	63.13%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	20.17%	28.64%	48.39%	29.18%	39.94%	45.40%	11.35%	33.71%	45.34%	37.46%	56.56%	62.24%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	6.11%	13.50%	29.37%	10.40%	24.31%	39.18%	7.51%	21.21%	30.28%	20.71%	44.95%	51.74%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	6.11%	13.50%	29.37%	10.40%	24.31%	39.18%	7.51%	21.21%	30.28%	20.71%	44.95%	51.74%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	20.17%	28.64%	48.39%	29.18%	39.93%	45.40%	11.35%	33.71%	45.34%	37.46%	56.56%	62.24%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	6.11%	13.50%	29.38%	10.40%	24.31%	39.18%	7.51%	21.21%	30.28%	20.71%	44.95%	51.74%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	6.11%	13.50%	29.37%	10.40%	24.31%	39.18%	7.51%	21.21%	30.28%	20.71%	44.95%	51.74%
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	20.17%	28.64%	48.39%	29.19%	39.94%	45.40%	11.35%	33.71%	45.34%	37.46%	56.56%	62.24%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	20.17%	28.70%	48.49%	29.27%	39.93%	46.68%	11.56%	33.92%	45.50%	37.67%	56.72%	63.13%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	20.17%	28.70%	48.49%	29.27%	39.93%	46.68%	11.56%	33.92%	45.50%	37.67%	56.72%	63.13%
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	10.90%	18.76%	35.33%	18.09%	30.55%	46.21%	8.87%	26.12%	37.62%	29.62%	49.34%	57.49%
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	10.90%	18.76%	35.33%	18.09%	30.54%	46.21%	8.87%	26.12%	37.62%	29.62%	49.34%	57.49%
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	20.36%	29.41%	49.77%	29.60%	40.58%	48.34%	11.69%	33.89%	44.95%	37.09%	55.93%	64.24%
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	23.91%	32.70%	52.99%	33.86%	44.17%	51.11%	12.63%	36.18%	47.58%	39.90%	58.67%	66.87%
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	19.52%	28.62%	48.19%	29.00%	40.51%	45.70%	11.53%	32.95%	43.71%	35.80%	55.03%	62.42%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	6.83%	16.66%	32.58%	15.20%	27.90%	45.81%	8.86%	23.36%	32.25%	22.38%	47.27%	58.84%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	22.43%	31.53%	51.87%	32.46%	43.01%	51.13%	12.21%	35.42%	46.70%	38.87%	57.98%	67.87%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	24.16%	32.78%	53.00%	33.90%	44.19%	51.22%	12.61%	36.47%	47.76%	40.06%	58.88%	66.82%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	20.75%	29.47%	49.53%	30.00%	40.76%	49.24%	11.69%	34.10%	45.71%	37.73%	57.20%	65.62%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	6.26%	14.37%	30.62%	11.52%	25.52%	44.66%	7.86%	21.59%	30.67%	20.66%	45.74%	56.25%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	24.03%	32.72%	52.94%	33.73%	44.15%	51.06%	12.63%	36.33%	47.66%	39.87%	58.85%	66.69%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	23.58%	32.43%	52.73%	33.20%	44.03%	51.56%	12.50%	35.95%	47.09%	39.29%	58.01%	66.64%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	23.45%	32.33%	52.77%	33.05%	43.96%	51.30%	12.53%	35.80%	47.00%	39.14%	58.05%	65.19%
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	26.69%	35.88%	56.96%	37.44%	47.77%	57.21%	13.53%	38.79%	50.66%	42.74%	62.02%	68.55%
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	26.38%	35.60%	56.84%	37.14%	47.53%	57.39%	13.46%	38.52%	50.31%	42.38%	61.71%	68.83%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	5.20%	10.24%	23.26%	7.61%	18.35%	39.44%	6.82%	19.67%	29.37%	20.60%	42.92%	51.46%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	5.25%	10.24%	23.33%	7.52%	18.31%	39.44%	6.82%	19.67%	29.37%	20.60%	42.92%	51.46%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	5.20%	10.24%	23.25%	7.66%	18.36%	39.44%	6.82%	19.67%	29.37%	20.60%	42.92%	51.46%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	14.37%	23.45%	42.67%	22.97%	36.20%	48.75%	10.26%	30.03%	41.47%	33.35%	53.29%	62.14%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	14.47%	23.88%	43.07%	23.62%	36.90%	54.33%	10.51%	30.65%	42.08%	33.83%	54.15%	65.22%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	14.52%	23.87%	43.09%	23.60%	36.78%	54.26%	10.45%	30.61%	42.05%	33.84%	54.07%	65.19%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	14.80%	23.64%	43.08%	23.58%	35.92%	53.37%	10.44%	30.39%	41.91%	33.88%	53.55%	63.67%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	5.40%	10.70%	24.01%	8.20%	18.91%	43.27%	6.91%	20.03%	29.83%	20.89%	43.74%	56.14%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	5.45%	10.73%	24.08%	8.07%	18.91%	43.34%	6.99%	20.05%	29.84%	20.90%	43.74%	56.03%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	14.83%	23.73%	43.14%	23.73%	35.97%	53.45%	10.44%	30.44%	41.93%	33.86%	53.64%	63.75%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	5.39%	10.69%	24.04%	8.17%	18.87%	43.21%	6.90%	20.01%	29.82%	20.89%	43.74%	56.04%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	5.43%	10.77%	24.16%	8.05%	18.92%	43.49%	6.92%	19.99%	29.81%	20.80%	43.75%	56.04%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	14.76%	23.70%	43.14%	23.64%	36.00%	53.34%	10.43%	30.41%	41.93%	33.85%	53.65%	63.65%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	14.57%	23.87%	43.19%	23.56%	36.72%	54.17%	10.45%	30.64%	42.12%	33.93%	54.14%	65.11%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	14.56%	23.90%	43.15%	23.79%	36.84%	54.24%	10.48%	30.61%	42.03%	33.78%	54.13%	65.20%
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	9.06%	14.31%	29.11%	12.47%	25.59%	39.91%	8.03%	23.79%	35.33%	29.05%	46.09%	50.08%
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	9.11%	14.34%	29.11%	12.52%	25.56%	39.91%	8.03%	23.79%	35.33%	29.05%	46.09%	50.08%
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	26.41%	35.64%	56.92%	37.06%	47.49%	57.55%	13.45%	38.53%	50.32%	42.42%	61.63%	68.06%
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	11.59%	19.81%	36.20%	19.46%	32.26%	38.39%	8.74%	25.74%	35.68%	26.96%	49.72%	49.01%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	8.22%	18.16%	35.18%	16.64%	29.42%	50.31%	9.51%	25.17%	34.61%	24.51%	49.76%	63.25%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	25.33%	34.52%	55.94%	35.70%	46.16%	56.59%	13.13%	37.85%	49.61%	41.80%	60.81%	67.19%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	26.76%	35.89%	56.97%	37.41%	47.54%	56.44%	13.48%	38.83%	50.69%	42.92%	61.79%	67.34%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	19.05%	29.16%	49.28%	29.71%	42.12%	52.16%	11.74%	34.12%	45.67%	37.34%	57.78%	64.00%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	5.78%	13.85%	28.68%	11.37%	23.95%	47.94%	7.93%	22.26%	31.77%	22.04%	46.32%	62.08%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	26.72%	35.86%	56.89%	37.39%	47.60%	56.48%	13.54%	38.77%	50.62%	42.75%	61.86%	67.28%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	26.14%	35.52%	56.73%	36.78%	47.51%	56.58%	13.41%	38.45%	50.14%	42.20%	61.23%	67.30%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	26.69%	35.80%	56.91%	37.37%	47.54%	56.72%	13.52%	38.74%	50.59%	42.73%	61.88%	67.64%

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	—	21.60%	ref	30.29%	49.23%	17.39	187.3	v-1969
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	21.37%	-0.23	30.08%	49.14%	12.78	253.7	v-19c0
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	6.88%	-14.72	16.17%	30.87%	745.26	1.8	v-1b91
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	6.88%	-14.72	16.17%	30.87%	193.39	5.5	v-1afd
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	6.88%	-14.72	16.17%	30.87%	193.36	5.6	v-1b02
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	17.74%	-3.86	26.38%	45.20%	202.37	6.2	v-1b07
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	6.93%	-14.67	16.48%	31.36%	232.03	5.5	v-1c09
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	18.04%	-3.56	26.65%	45.78%	101.31	23.6	v-171b
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	18.04%	-3.56	26.65%	45.78%	185.06	28.7	v-19ec
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	17.82%	-3.78	26.55%	45.71%	75.54	24.4	v-1a68
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	6.85%	-14.75	16.29%	31.04%	110.33	10.3	v-1a66
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	6.87%	-14.73	16.34%	31.11%	112.76	10.0	v-1a63
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	17.82%	-3.78	26.55%	45.71%	171.61	29.0	v-1a69
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	6.85%	-14.75	16.29%	31.04%	159.50	27.2	v-1a67
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	6.87%	-14.73	16.34%	31.11%	151.53	29.1	v-1a64
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	17.82%	-3.78	26.55%	45.71%	171.67	29.0	v-1872
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	18.04%	-3.56	26.65%	45.78%	107.55	22.5	v-172b
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	18.04%	-3.56	26.65%	45.78%	169.37	28.0	v-19f3
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	11.15%	-10.45 ⚠	21.11%	37.26%	33.89	54.7	v-15d2
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	11.14%	-10.46 ⚠	21.10%	37.25%	29.70	124.6	v-1907
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	18.83%	-2.77	27.27%	46.75%	37.06	37.5	v-15f1
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	21.48%	-0.12	30.12%	49.18%	68.12	105.7	v-190d
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	18.52%	-3.08	27.29%	46.02%	62.32	116.9	v-190f
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	8.96%	-12.64	18.10%	34.31%	12.54	301.1	v-154f
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	20.47%	-1.13	29.04%	48.07%	19.64	156.7	v-1550
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	21.53%	-0.07	30.14%	49.20%	33.70	117.7	v-1551
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	18.94%	-2.66	27.15%	46.26%	19.55	156.2	v-1d5b
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	7.44%	-14.16	16.67%	31.61%	14.35	302.6	v-1d5c
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	21.38%	-0.22	30.01%	48.97%	34.86	117.7	v-1d5d
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	21.16%	-0.44	29.90%	49.05%	27.33	104.2	v-1d5f
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	21.36%	-0.24	29.98%	48.98%	29.81	103.2	v-1d5e
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	25.91%	ref	34.50%	55.59%	23.64	142.9	v-196d
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	25.75%	-0.16	34.32%	55.43%	15.42	204.0	v-19c7
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	8.63%	-17.28	18.30%	35.44%	934.43	1.7	v-1bc8
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	8.64%	-17.27	18.30%	35.44%	341.99	3.0	v-1b24
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	8.64%	-17.27	18.30%	35.44%	341.85	3.0	v-1b28
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	22.40%	-3.51	31.03%	51.73%	352.00	3.2	v-1b30
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	22.80%	-3.11	31.49%	52.43%	164.25	10.4	v-172e
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	22.80%	-3.11	31.49%	52.43%	431.33	11.0	v-19f1
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	22.51%	-3.40	31.11%	52.22%	131.72	10.4	v-1a8b
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	8.74%	-17.17	18.12%	35.59%	166.52	6.5	v-1a89
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	8.74%	-17.17	18.12%	35.59%	169.34	6.3	v-1a86
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	22.51%	-3.40	31.11%	52.22%	394.00	11.2	v-1a8d
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	8.74%	-17.17	18.12%	35.58%	375.04	10.8	v-1a8a
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	8.74%	-17.17	18.12%	35.58%	369.94	11.2	v-1a87
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	22.51%	-3.40	31.11%	52.22%	395.28	11.1	v-187a
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	22.80%	-3.11	31.49%	52.43%	171.19	10.5	v-1740
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	22.80%	-3.11	31.49%	52.43%	438.11	11.1	v-19f8
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	14.20%	-11.71 ⚠	23.75%	42.26%	53.36	29.2	v-15c5
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	14.22%	-11.69 ⚠	23.75%	42.26%	46.32	37.7	v-1913
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	23.67%	-2.24	31.64%	53.51%	64.18	18.8	v-1607
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	25.70%	-0.21	34.29%	55.43%	79.88	65.0	v-15c8
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	21.22%	-4.69	30.46%	50.81%	60.29	104.4	v-15ca
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	9.77%	-16.14	20.37%	37.95%	13.91	216.2	v-157d
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	24.22%	-1.69	33.03%	54.22%	22.08	179.9	v-157f
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	25.87%	-0.04	34.48%	55.56%	42.03	81.9	v-1580
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	23.33%	-2.58	31.73%	52.79%	22.22	156.4	v-1d99
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	8.89%	-17.02	17.54%	34.61%	16.92	217.4	v-1d9a
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	25.82%	-0.09	34.45%	55.51%	41.71	80.9	v-1d9b
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	25.42%	-0.49	34.13%	55.43%	34.93	73.1	v-1d9d
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	25.53%	-0.38	34.23%	55.41%	35.32	72.8	v-1d9c
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	28.46%	ref	36.89%	58.87%	58.28	80.7	v-1972
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	28.29%	-0.17	36.75%	58.68%	39.12	125.9	v-19cb
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	5.67%	-22.79	11.83%	25.11%	1288.98	1.5	v-1c25
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	5.66%	-22.80	11.82%	25.09%	618.11	1.7	v-1b94
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	5.66%	-22.80	11.82%	25.07%	618.03	1.7	v-1ba0
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	15.58%	-12.88 ⚠	24.55%	44.10%	624.35	1.7	v-1ba5
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	15.63%	-12.83 ⚠	25.41%	44.75%	313.02	4.6	v-175a
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	15.59%	-12.87 ⚠	25.37%	44.73%	947.97	4.7	v-1a0a
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	15.33%	-13.13 ⚠	25.07%	44.33%	254.92	4.6	v-1ad9
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	5.77%	-22.69	11.92%	25.26%	296.18	3.6	v-1acf
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	5.78%	-22.68	11.94%	25.29%	300.10	3.5	v-1acb
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	15.31%	-13.15 ⚠	25.07%	44.32%	876.82	4.8	v-1add
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	5.74%	-22.72	11.92%	25.25%	856.46	4.7	v-1ad7
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	5.76%	-22.70	11.92%	25.30%	859.19	4.8	v-1acd
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	15.31%	-13.15 ⚠	25.07%	44.32%	877.38	4.8	v-1889
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	15.59%	-12.87 ⚠	25.38%	44.74%	316.21	4.6	v-1764
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	15.59%	-12.87 ⚠	25.38%	44.74%	940.25	4.7	v-1a14
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	9.14%	-19.32 ⚠	16.51%	31.20%	100.86	13.5	v-15fc
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	8.92%	-19.54 ⚠	16.40%	31.08%	92.70	15.1	v-1925
Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	21.12%	-7.34	30.99%	52.35%	118.24	9.3	v-1625
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	28.31%	-0.15	36.72%	58.74%	142.59	32.3	v-1939
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	14.90%	-13.56 ⚠	24.99%	44.56%	95.98	52.4	v-193b
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	9.89%	-18.57	20.41%	38.37%	23.55	106.3	v-162b
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	23.07%	-5.39	32.38%	54.25%	32.27	113.4	v-1630
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	28.35%	-0.11	36.80%	58.70%	71.92	39.7	v-1632
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	17.64%	-10.82 ⚠	26.41%	46.49%	33.35	116.1	v-1dda
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	6.94%	-21.52	13.00%	27.42%	26.99	108.3	v-1ddb
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	28.38%	-0.08	36.82%	58.80%	71.78	39.7	v-1ddc
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	28.04%	-0.42	36.60%	58.50%	60.89	38.0	v-1dde
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	28.05%	-0.41	36.59%	58.60%	62.15	37.5	v-1ddd

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	21.60%	30.29%	49.23%	30.87%	41.61%	53.41%	12.00%	34.80%	46.09%	37.20%	59.07%	71.32%
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	21.37%	30.08%	49.14%	30.60%	41.41%	53.46%	11.92%	34.54%	45.78%	36.88%	58.80%	71.40%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	6.88%	16.17%	30.87%	15.12%	26.29%	46.88%	8.62%	22.88%	31.65%	21.69%	46.28%	63.41%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	6.88%	16.17%	30.87%	15.12%	26.29%	46.88%	8.62%	22.88%	31.65%	21.69%	46.28%	63.41%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	6.88%	16.17%	30.87%	15.12%	26.29%	46.88%	8.62%	22.88%	31.65%	21.69%	46.28%	63.41%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	17.74%	26.38%	45.20%	26.57%	36.73%	51.11%	10.89%	31.18%	41.81%	33.21%	53.86%	67.85%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	INT8	6.93%	16.48%	31.36%	15.40%	27.03%	47.97%	8.75%	23.28%	32.24%	22.06%	47.37%	66.29%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	18.04%	26.65%	45.78%	26.83%	37.39%	51.65%	11.05%	31.55%	42.24%	33.43%	55.08%	68.12%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	18.04%	26.65%	45.78%	26.83%	37.39%	51.65%	11.05%	31.55%	42.24%	33.43%	55.08%	68.12%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	17.82%	26.55%	45.71%	26.95%	37.26%	50.99%	11.19%	31.47%	42.19%	33.43%	54.79%	70.35%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	6.85%	16.29%	31.04%	15.35%	26.95%	47.57%	8.71%	23.17%	31.95%	21.67%	47.24%	66.70%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	6.87%	16.34%	31.11%	15.38%	27.00%	47.87%	8.73%	23.20%	31.99%	21.68%	47.31%	67.09%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	17.82%	26.55%	45.71%	26.95%	37.26%	50.99%	11.19%	31.47%	42.19%	33.43%	54.79%	70.35%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	6.85%	16.29%	31.04%	15.35%	26.95%	47.57%	8.71%	23.17%	31.95%	21.67%	47.24%	66.70%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	6.87%	16.34%	31.11%	15.38%	27.00%	47.87%	8.73%	23.20%	31.99%	21.68%	47.31%	67.09%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	17.82%	26.55%	45.71%	26.95%	37.26%	50.99%	11.19%	31.47%	42.19%	33.43%	54.79%	70.35%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	18.04%	26.65%	45.78%	26.83%	37.39%	51.65%	11.05%	31.55%	42.24%	33.43%	55.08%	68.12%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	18.04%	26.65%	45.78%	26.83%	37.39%	51.65%	11.05%	31.55%	42.24%	33.43%	55.08%	68.12%
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	11.15%	21.11%	37.26%	21.67%	33.24%	47.35%	9.88%	27.02%	37.36%	28.50%	50.15%	64.09%
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	11.14%	21.10%	37.25%	21.67%	33.24%	47.36%	9.88%	27.02%	37.36%	28.50%	50.15%	64.09%
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	18.83%	27.27%	46.75%	27.74%	37.99%	51.63%	11.24%	32.00%	42.84%	34.26%	55.45%	65.31%
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	21.48%	30.12%	49.18%	30.75%	41.38%	53.45%	11.97%	34.51%	45.80%	37.02%	58.79%	70.89%
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	18.52%	27.29%	46.02%	27.73%	38.39%	50.69%	11.15%	32.20%	43.02%	34.20%	55.81%	66.94%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	8.96%	18.10%	34.31%	16.85%	28.88%	46.25%	9.04%	24.69%	33.95%	24.11%	48.98%	62.07%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	20.47%	29.04%	48.07%	29.71%	40.41%	49.31%	11.73%	33.39%	44.42%	35.76%	57.26%	64.94%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	21.53%	30.14%	49.20%	30.73%	41.38%	52.89%	12.07%	34.63%	45.89%	37.11%	58.73%	67.86%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	18.94%	27.15%	46.26%	27.52%	37.72%	49.90%	11.27%	31.76%	42.55%	34.15%	54.71%	65.13%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	7.44%	16.67%	31.61%	15.96%	27.16%	47.87%	8.80%	23.37%	32.34%	22.49%	47.39%	60.91%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	21.38%	30.01%	48.97%	30.77%	41.35%	52.51%	12.04%	34.46%	45.60%	36.82%	58.55%	67.38%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	21.16%	29.90%	49.05%	30.41%	41.29%	52.72%	11.98%	34.29%	45.43%	36.65%	58.02%	67.42%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	21.36%	29.98%	48.98%	30.71%	41.24%	52.53%	12.03%	34.42%	45.58%	36.81%	58.56%	67.39%
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	25.91%	34.50%	55.59%	35.61%	46.20%	53.02%	13.27%	38.23%	49.93%	42.02%	62.01%	72.51%
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	25.75%	34.32%	55.43%	35.42%	46.03%	53.13%	13.22%	38.03%	49.70%	41.77%	61.83%	72.45%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	8.63%	18.30%	35.44%	17.00%	30.54%	48.47%	9.33%	25.39%	34.93%	25.29%	50.12%	66.01%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	8.64%	18.30%	35.44%	17.00%	30.54%	48.47%	9.33%	25.39%	34.93%	25.29%	50.12%	66.01%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	8.64%	18.30%	35.44%	17.00%	30.54%	48.47%	9.33%	25.39%	34.93%	25.29%	50.12%	66.01%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	22.40%	31.03%	51.73%	32.05%	42.70%	49.27%	12.33%	35.50%	46.78%	38.81%	59.03%	68.28%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	22.80%	31.49%	52.43%	32.36%	43.27%	51.21%	12.49%	35.92%	47.52%	39.37%	59.83%	70.80%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	22.80%	31.49%	52.43%	32.36%	43.27%	51.21%	12.49%	35.92%	47.52%	39.37%	59.83%	70.80%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	22.51%	31.11%	52.22%	31.90%	42.74%	50.67%	12.15%	35.89%	47.25%	39.14%	59.52%	70.49%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	8.74%	18.12%	35.59%	16.55%	30.37%	46.77%	8.89%	25.34%	35.26%	25.79%	50.38%	65.99%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	8.74%	18.12%	35.59%	16.55%	30.37%	46.77%	8.89%	25.34%	35.26%	25.79%	50.38%	65.99%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	22.51%	31.11%	52.22%	31.90%	42.74%	50.67%	12.15%	35.89%	47.25%	39.14%	59.52%	70.49%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	8.74%	18.12%	35.58%	16.55%	30.37%	46.77%	8.89%	25.34%	35.26%	25.79%	50.38%	65.99%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	8.74%	18.12%	35.58%	16.55%	30.36%	46.77%	8.89%	25.34%	35.26%	25.79%	50.38%	65.99%
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	22.51%	31.11%	52.22%	31.90%	42.74%	50.67%	12.15%	35.89%	47.25%	39.14%	59.52%	70.49%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	22.80%	31.49%	52.43%	32.35%	43.27%	51.21%	12.49%	35.92%	47.52%	39.37%	59.83%	70.80%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	22.80%	31.49%	52.43%	32.35%	43.27%	51.21%	12.49%	35.92%	47.52%	39.37%	59.83%	70.80%
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	14.20%	23.75%	42.26%	23.86%	36.67%	47.65%	10.26%	29.45%	40.77%	32.60%	53.11%	67.61%
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	14.22%	23.75%	42.26%	23.86%	36.67%	47.65%	10.26%	29.45%	40.77%	32.60%	53.11%	67.61%
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	23.67%	31.64%	53.51%	32.28%	43.13%	50.40%	12.46%	36.49%	48.34%	40.66%	60.20%	69.93%
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	25.70%	34.29%	55.43%	35.36%	46.00%	53.92%	13.26%	38.01%	49.68%	41.68%	61.88%	72.23%
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	21.22%	30.46%	50.81%	31.05%	43.05%	53.10%	12.34%	34.89%	46.13%	37.71%	58.73%	70.67%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	9.77%	20.37%	37.95%	19.69%	33.37%	48.65%	10.01%	27.08%	36.92%	27.07%	52.31%	67.86%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	24.22%	33.03%	54.22%	33.99%	45.09%	51.88%	12.85%	37.04%	48.60%	40.33%	61.08%	72.01%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	25.87%	34.48%	55.56%	35.61%	46.21%	51.74%	13.24%	38.26%	49.98%	41.98%	62.25%	72.51%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	23.33%	31.73%	52.79%	32.51%	43.28%	50.84%	12.42%	35.89%	47.50%	39.86%	59.35%	69.39%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	8.89%	17.54%	34.61%	15.85%	29.68%	46.23%	9.16%	25.04%	35.07%	25.74%	50.17%	64.91%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	25.82%	34.45%	55.51%	35.66%	46.17%	52.11%	13.39%	38.21%	49.85%	41.90%	62.07%	72.32%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	25.42%	34.13%	55.43%	34.94%	46.05%	52.03%	13.13%	37.87%	49.41%	41.49%	61.48%	72.45%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	25.53%	34.23%	55.41%	35.17%	46.09%	52.36%	13.32%	37.93%	49.49%	41.59%	61.46%	72.14%
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	28.46%	36.89%	58.87%	38.61%	48.80%	58.07%	13.65%	39.97%	51.96%	44.62%	63.52%	73.16%
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	28.29%	36.75%	58.68%	38.66%	48.72%	58.04%	13.61%	39.80%	51.77%	44.37%	63.40%	73.10%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	5.67%	11.83%	25.11%	9.95%	21.95%	42.54%	7.06%	21.13%	32.31%	23.60%	45.68%	53.62%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	5.66%	11.82%	25.09%	9.93%	21.99%	42.54%	7.06%	21.13%	32.31%	23.60%	45.68%	53.62%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	5.66%	11.82%	25.07%	9.95%	21.98%	42.54%	7.06%	21.13%	32.31%	23.60%	45.68%	53.62%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	15.58%	24.55%	44.10%	24.06%	36.84%	53.38%	10.50%	31.32%	43.40%	35.57%	54.82%	64.47%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	15.63%	25.41%	44.75%	25.16%	37.85%	52.71%	10.81%	31.73%	43.97%	36.07%	55.27%	68.69%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	15.59%	25.37%	44.73%	25.11%	37.85%	52.71%	10.81%	31.73%	43.97%	36.07%	55.27%	68.69%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	15.33%	25.07%	44.33%	25.10%	37.67%	52.74%	10.85%	31.55%	43.80%	35.92%	55.36%	67.10%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	5.77%	11.92%	25.26%	10.01%	21.72%	45.21%	7.34%	21.86%	33.08%	24.40%	46.58%	56.75%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	5.78%	11.94%	25.29%	10.05%	21.74%	45.23%	7.36%	21.88%	33.12%	24.40%	46.69%	56.76%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	15.31%	25.07%	44.32%	25.10%	37.67%	52.74%	10.85%	31.55%	43.80%	35.92%	55.36%	67.10%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	5.74%	11.92%	25.25%	10.01%	21.70%	45.21%	7.34%	21.86%	33.08%	24.40%	46.58%	56.75%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	5.76%	11.92%	25.30%	10.03%	21.76%	45.23%	7.36%	21.88%	33.12%	24.40%	46.69%	56.76%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	15.31%	25.07%	44.32%	25.10%	37.67%	52.74%	10.85%	31.55%	43.80%	35.92%	55.36%	67.10%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	15.59%	25.38%	44.74%	25.11%	37.85%	52.71%	10.81%	31.73%	43.97%	36.07%	55.27%	68.69%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	15.59%	25.38%	44.74%	25.11%	37.85%	52.71%	10.81%	31.73%	43.97%	36.07%	55.27%	68.69%
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	9.14%	16.51%	31.20%	16.04%	28.67%	46.70%	8.70%	25.33%	37.98%	30.87%	48.58%	56.88%
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	8.92%	16.40%	31.08%	15.96%	28.72%	46.68%	8.59%	25.28%	37.98%	30.87%	48.68%	57.34%
Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	21.12%	30.99%	52.35%	31.69%	44.58%	55.21%	12.60%	35.27%	46.61%	38.12%	59.32%	72.68%
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	28.31%	36.72%	58.74%	38.51%	48.62%	58.50%	13.66%	39.84%	51.73%	44.38%	63.32%	73.88%
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	14.90%	24.99%	44.56%	23.74%	39.08%	54.17%	10.66%	30.86%	41.50%	32.84%	54.40%	67.98%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	9.89%	20.41%	38.37%	19.50%	33.56%	52.44%	10.10%	27.49%	37.51%	28.06%	52.59%	68.63%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	23.07%	32.38%	54.25%	32.79%	44.75%	56.38%	12.35%	36.41%	48.44%	40.65%	60.39%	74.86%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	28.35%	36.80%	58.70%	38.55%	48.74%	57.96%	13.66%	39.94%	51.81%	44.32%	63.53%	73.66%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	17.64%	26.41%	46.49%	26.12%	39.07%	52.55%	10.94%	32.90%	45.51%	37.94%	57.15%	68.79%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	6.94%	13.00%	27.42%	10.78%	23.54%	46.48%	7.53%	22.62%	34.09%	26.01%	47.38%	57.83%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	28.38%	36.82%	58.80%	38.51%	48.69%	58.40%	13.66%	39.89%	51.86%	44.45%	63.47%	74.50%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	28.04%	36.60%	58.50%	38.34%	48.74%	57.79%	13.57%	39.70%	51.48%	43.98%	63.00%	73.22%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	28.05%	36.59%	58.60%	38.10%	48.75%	58.31%	13.57%	39.63%	51.50%	44.00%	63.08%	73.90%

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	—	19.98%	ref	28.48%	47.82%	33.48	58.5	v-1968
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	19.85%	-0.13	28.39%	47.80%	28.20	71.3	v-19bf
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	11.86%	-8.12	20.60%	41.34%	1476.47	0.9	v-1b86
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	11.86%	-8.12	20.60%	41.34%	372.97	2.9	v-1a8c
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	11.86%	-8.12	20.60%	41.34%	373.59	2.9	v-1a91
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	17.66%	-2.32	25.26%	45.29%	416.68	3.2	v-1a96
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	11.74%	-8.24	20.71%	41.39%	410.82	2.9	v-1c03
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	17.88%	-2.10	25.73%	45.79%	169.41	11.4	v-1706
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	17.88%	-2.10	25.74%	45.79%	212.48	7.8	v-19e9
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	18.01%	-1.97	25.96%	45.88%	164.06	11.4	v-1a17
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	11.86%	-8.12	20.98%	41.86%	219.14	5.5	v-1a13
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	11.82%	-8.16	20.96%	41.83%	223.04	5.3	v-1a11
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	18.01%	-1.97	25.96%	45.88%	180.27	10.5	v-1a1a
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	11.86%	-8.12	20.98%	41.86%	277.48	12.9	v-1a16
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	11.82%	-8.16	20.96%	41.83%	243.18	15.6	v-1a12
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	18.01%	-1.97	25.96%	45.88%	185.29	9.9	v-1866
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	17.88%	-2.10	25.74%	45.79%	178.68	11.1	v-16fb
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	17.88%	-2.10	25.73%	45.79%	215.91	7.6	v-19f2
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	14.58%	-5.40	22.89%	43.52%	83.87	21.2	v-15b4
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	14.56%	-5.42	22.86%	43.51%	68.20	54.3	v-1903
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	18.94%	-1.04	27.20%	47.30%	92.19	15.0	v-15e0
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	19.88%	-0.10	28.43%	47.80%	128.30	46.6	v-190a
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	17.33%	-2.65	25.83%	45.51%	118.93	56.0	v-190c
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	13.78%	-6.20	22.82%	44.42%	39.40	125.5	v-154a
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	18.89%	-1.09	27.41%	47.19%	48.34	66.6	v-154c
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	19.86%	-0.12	28.38%	47.75%	92.11	51.9	v-154e
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	18.16%	-1.82	26.46%	46.67%	47.69	68.7	v-1d36
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	12.15%	-7.83	21.48%	42.58%	39.42	127.1	v-1d37
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	19.67%	-0.31	28.26%	47.66%	87.46	52.2	v-1d38
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	19.51%	-0.47	28.14%	47.63%	69.20	48.9	v-1d3a
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	19.67%	-0.31	28.23%	47.67%	64.22	48.5	v-1d39
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	23.69%	ref	32.42%	53.39%	46.82	50.0	v-196c
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	23.62%	-0.07	32.41%	53.38%	34.22	65.4	v-19c5
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	13.13%	-10.56	22.37%	44.64%	1832.52	0.9	v-1bb2
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	13.13%	-10.56	22.37%	44.64%	666.98	1.6	v-1b01
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	13.13%	-10.56	22.37%	44.64%	667.20	1.6	v-1b0f
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	19.77%	-3.92	28.39%	49.12%	712.89	1.6	v-1b1e
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	13.19%	-10.50	22.26%	44.77%	708.98	1.6	v-1c27
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	20.11%	-3.58	28.43%	49.55%	300.17	5.6	v-1723
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	20.11%	-3.58	28.43%	49.55%	576.66	5.7	v-19f0
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	20.31%	-3.38	28.60%	49.74%	289.23	5.5	v-1a82
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	13.51%	-10.18	22.35%	44.83%	341.56	3.3	v-1a7d
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	13.49%	-10.20	22.32%	44.83%	348.29	3.2	v-1a79
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	20.31%	-3.38	28.59%	49.73%	652.79	6.2	v-1a83
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	13.51%	-10.18	22.34%	44.83%	675.09	5.5	v-1a80
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	13.49%	-10.20	22.32%	44.83%	647.08	6.2	v-1a7b
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	20.31%	-3.38	28.59%	49.74%	647.07	6.2	v-1879
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	20.11%	-3.58	28.43%	49.55%	304.93	5.6	v-1735
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	20.11%	-3.58	28.43%	49.55%	574.63	5.8	v-19f6
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	16.20%	-7.49	24.43%	46.18%	108.06	15.9	v-15bb
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	16.16%	-7.53	24.39%	46.15%	88.06	27.8	v-190e
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	22.09%	-1.60	29.85%	51.54%	200.10	5.8	v-15fd
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	23.61%	-0.08	32.36%	53.32%	170.50	28.7	v-15c4
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	18.42%	-5.27	27.79%	49.14%	122.68	42.7	v-15c6
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	15.74%	-7.95	25.22%	48.74%	34.85	91.4	v-155f
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	22.79%	-0.90	31.23%	52.33%	52.11	76.4	v-1561
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	23.70%	+0.01	32.40%	53.38%	101.44	39.2	v-1562
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	20.48%	-3.21	28.74%	49.55%	57.99	67.1	v-1d74
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	13.76%	-9.93	22.70%	44.91%	38.68	92.0	v-1d75
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	23.72%	+0.03	32.36%	53.31%	90.85	39.3	v-1d76
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	23.36%	-0.33	32.15%	53.14%	76.56	36.5	v-1d78
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	23.55%	-0.14	32.26%	53.16%	72.21	36.4	v-1d77
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	26.32%	ref	34.69%	56.34%	88.64	34.6	v-1971
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	26.28%	-0.04	34.70%	56.35%	59.92	46.9	v-19ca
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	9.94%	-16.38	16.68%	38.01%	2516.62	0.8	v-1c0b
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	9.86%	-16.46	16.62%	37.95%	1209.94	0.9	v-1b6e
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	9.86%	-16.46	16.62%	37.95%	1209.94	0.9	v-1b7f
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	14.03%	-12.29 ⚠	22.30%	43.78%	1247.25	0.9	v-1b89
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	14.13%	-12.19 ⚠	22.42%	44.24%	570.95	2.3	v-1747
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	14.13%	-12.19 ⚠	22.43%	44.26%	1665.52	2.4	v-1a02
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	14.78%	-11.54 ⚠	22.93%	44.83%	527.95	2.3	v-1ac0
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	10.11%	-16.21	16.80%	38.63%	586.34	1.8	v-1ab6
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	10.24%	-16.08	16.86%	38.74%	593.57	1.8	v-1aa8
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	14.73%	-11.59 ⚠	22.90%	44.82%	1653.81	2.4	v-1ac3
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	10.12%	-16.20	16.80%	38.63%	1676.50	2.3	v-1abb
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	10.18%	-16.14	16.80%	38.65%	1645.51	2.4	v-1ab2
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	14.73%	-11.59 ⚠	22.90%	44.82%	1655.29	2.4	v-1884
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	14.12%	-12.20 ⚠	22.43%	44.29%	573.63	2.3	v-1752
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	14.12%	-12.20 ⚠	22.43%	44.29%	1658.73	2.4	v-1a0f
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	12.07%	-14.25 ⚠	18.47%	39.66%	176.08	8.5	v-15ee
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	12.10%	-14.22 ⚠	18.51%	39.77%	150.94	10.6	v-1922
Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	21.75%	-4.57	30.39%	53.92%	344.62	3.1	v-1619
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	26.24%	-0.08	34.68%	56.36%	342.11	14.2	v-1934
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	15.36%	-10.96 ⚠	24.80%	46.82%	209.52	21.6	v-1937
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	18.05%	-8.27	27.63%	52.27%	56.82	47.8	v-15a7
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	25.04%	-1.28	33.12%	55.18%	89.13	49.6	v-15a8
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	26.35%	+0.03	34.76%	56.28%	150.17	21.5	v-15a9
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	17.94%	-8.38	26.46%	48.48%	91.54	49.8	v-1dcb
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	12.40%	-13.92	19.87%	42.46%	57.90	48.9	v-1dcc
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	26.42%	+0.10	34.74%	56.35%	144.82	21.1	v-1dcd
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	26.04%	-0.28	34.54%	56.18%	131.28	17.8	v-1dcf
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	26.14%	-0.18	34.55%	56.24%	134.07	17.5	v-1dce

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	19.98%	28.48%	47.82%	28.65%	39.52%	48.53%	11.45%	33.82%	43.71%	35.04%	56.23%	58.42%
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	19.85%	28.39%	47.80%	28.55%	39.44%	48.56%	11.41%	33.74%	43.61%	34.94%	56.11%	58.77%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	11.86%	20.60%	41.34%	18.31%	31.89%	45.28%	9.45%	26.26%	33.20%	23.47%	47.16%	53.59%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	11.86%	20.60%	41.34%	18.31%	31.89%	45.28%	9.45%	26.26%	33.20%	23.47%	47.15%	53.59%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	11.86%	20.60%	41.34%	18.32%	31.90%	45.28%	9.45%	26.26%	33.20%	23.47%	47.16%	53.59%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	17.66%	25.26%	45.29%	24.78%	34.83%	48.31%	10.66%	31.11%	40.07%	31.76%	51.71%	56.88%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	INT8	11.74%	20.71%	41.39%	18.48%	32.50%	42.61%	9.39%	26.22%	33.37%	23.34%	47.81%	52.81%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	17.88%	25.73%	45.79%	25.26%	35.68%	44.66%	10.77%	31.41%	40.67%	32.15%	52.67%	55.47%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	17.88%	25.74%	45.79%	25.26%	35.68%	44.66%	10.77%	31.41%	40.67%	32.15%	52.67%	55.47%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	18.01%	25.96%	45.88%	25.67%	35.78%	46.33%	10.82%	31.50%	40.72%	32.30%	52.56%	57.30%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	11.86%	20.98%	41.86%	18.59%	32.83%	42.89%	9.57%	26.40%	33.63%	23.33%	48.21%	55.41%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	11.82%	20.96%	41.83%	18.50%	32.81%	42.57%	9.56%	26.37%	33.60%	23.30%	48.18%	55.15%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	18.01%	25.96%	45.88%	25.67%	35.78%	46.32%	10.82%	31.50%	40.72%	32.30%	52.56%	57.30%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	11.86%	20.98%	41.86%	18.59%	32.84%	42.90%	9.57%	26.40%	33.63%	23.33%	48.21%	55.41%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	11.82%	20.96%	41.83%	18.50%	32.82%	42.57%	9.56%	26.37%	33.60%	23.30%	48.18%	55.15%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	18.01%	25.96%	45.88%	25.67%	35.78%	46.33%	10.82%	31.50%	40.72%	32.30%	52.56%	57.30%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	17.88%	25.74%	45.79%	25.26%	35.68%	44.66%	10.77%	31.41%	40.67%	32.15%	52.67%	55.47%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	17.88%	25.73%	45.79%	25.26%	35.68%	44.66%	10.77%	31.41%	40.67%	32.15%	52.67%	55.47%
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	14.58%	22.89%	43.52%	21.35%	33.49%	42.50%	10.02%	28.58%	36.70%	27.56%	49.80%	53.27%
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	14.56%	22.86%	43.51%	21.33%	33.46%	42.49%	10.01%	28.54%	36.66%	27.54%	49.76%	53.27%
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	18.94%	27.20%	47.30%	26.79%	37.74%	47.01%	11.04%	32.33%	41.77%	33.13%	54.14%	57.59%
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	19.88%	28.43%	47.80%	28.57%	39.46%	48.60%	11.40%	33.79%	43.67%	34.95%	56.19%	59.36%
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	17.33%	25.83%	45.51%	25.35%	36.58%	42.90%	10.87%	31.37%	40.50%	31.92%	52.37%	51.74%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	13.78%	22.82%	44.42%	20.98%	34.73%	42.97%	9.96%	28.32%	36.05%	25.90%	51.06%	54.88%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	18.89%	27.41%	47.19%	27.10%	38.20%	46.03%	11.19%	32.70%	42.37%	33.59%	55.15%	57.09%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	19.86%	28.38%	47.75%	28.50%	39.41%	47.91%	11.41%	33.73%	43.60%	34.97%	56.08%	57.69%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	18.16%	26.46%	46.67%	26.08%	37.04%	44.36%	10.98%	31.74%	41.13%	32.42%	53.67%	53.90%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	12.15%	21.48%	42.58%	19.07%	33.41%	43.26%	9.77%	26.88%	34.06%	23.91%	48.47%	53.52%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	19.67%	28.26%	47.66%	28.35%	39.43%	48.13%	11.41%	33.60%	43.35%	34.59%	56.05%	58.20%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	19.51%	28.14%	47.63%	28.23%	39.34%	47.53%	11.35%	33.41%	43.11%	34.45%	55.54%	56.77%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	19.67%	28.23%	47.67%	28.33%	39.34%	48.14%	11.40%	33.58%	43.33%	34.60%	56.01%	58.21%
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	23.69%	32.42%	53.39%	33.52%	44.01%	47.30%	12.81%	36.92%	47.42%	39.40%	58.96%	61.56%
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	23.62%	32.41%	53.38%	33.44%	44.07%	47.27%	12.82%	36.89%	47.38%	39.31%	59.01%	61.49%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	13.13%	22.37%	44.64%	20.36%	35.32%	38.99%	10.08%	28.19%	35.74%	26.04%	49.94%	49.80%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	13.13%	22.37%	44.64%	20.36%	35.32%	38.99%	10.08%	28.19%	35.74%	26.04%	49.94%	49.80%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	13.13%	22.37%	44.64%	20.36%	35.32%	38.99%	10.08%	28.19%	35.74%	26.04%	49.94%	49.80%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	19.77%	28.39%	49.12%	28.50%	39.62%	41.28%	11.64%	33.94%	43.71%	35.83%	55.16%	52.35%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Phytec) — latency	LiteRT	Logical	INT8	13.19%	22.26%	44.77%	20.16%	35.35%	40.41%	9.91%	28.23%	36.00%	26.68%	49.92%	52.94%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	20.11%	28.43%	49.55%	28.18%	39.95%	40.29%	11.65%	34.13%	44.12%	36.35%	55.75%	53.96%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	20.11%	28.43%	49.55%	28.18%	39.95%	40.29%	11.65%	34.13%	44.12%	36.35%	55.75%	53.96%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	20.31%	28.60%	49.74%	28.63%	40.30%	40.24%	11.67%	34.26%	44.37%	36.65%	56.26%	53.60%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	13.51%	22.35%	44.83%	20.11%	35.27%	38.06%	9.92%	28.37%	36.06%	26.79%	50.26%	50.57%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	13.49%	22.32%	44.83%	20.10%	35.29%	38.14%	9.91%	28.35%	36.03%	26.75%	50.25%	50.66%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	20.31%	28.59%	49.73%	28.63%	40.30%	40.24%	11.67%	34.26%	44.37%	36.65%	56.26%	53.60%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	13.51%	22.34%	44.83%	20.11%	35.27%	38.06%	9.92%	28.37%	36.06%	26.79%	50.26%	50.57%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	13.49%	22.32%	44.83%	20.10%	35.28%	38.14%	9.91%	28.35%	36.03%	26.75%	50.25%	50.66%
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	20.31%	28.59%	49.74%	28.63%	40.30%	40.24%	11.67%	34.26%	44.37%	36.65%	56.26%	53.60%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	20.11%	28.43%	49.55%	28.18%	39.95%	40.29%	11.65%	34.13%	44.12%	36.35%	55.75%	53.96%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	20.11%	28.43%	49.55%	28.18%	39.95%	40.29%	11.65%	34.13%	44.12%	36.35%	55.75%	53.96%
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	16.20%	24.43%	46.18%	23.22%	36.40%	40.00%	10.44%	30.47%	39.45%	31.37%	51.11%	49.33%
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	16.16%	24.39%	46.15%	23.15%	36.35%	40.17%	10.44%	30.46%	39.42%	31.34%	51.12%	49.35%
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	22.09%	29.85%	51.54%	29.89%	40.59%	43.98%	11.92%	35.48%	45.60%	38.17%	56.56%	57.07%
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	23.61%	32.36%	53.32%	33.37%	43.98%	47.10%	12.71%	36.73%	47.32%	39.29%	58.90%	60.86%
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	18.42%	27.79%	49.14%	27.18%	40.79%	44.12%	11.44%	32.98%	42.12%	33.06%	55.73%	60.24%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	15.74%	25.22%	48.74%	22.89%	38.50%	43.69%	10.85%	30.13%	37.97%	28.02%	52.71%	58.46%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	22.79%	31.23%	52.33%	32.10%	42.51%	45.11%	12.28%	36.14%	46.48%	38.67%	57.82%	60.27%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	23.70%	32.40%	53.38%	33.57%	43.95%	46.73%	12.67%	36.84%	47.38%	39.43%	58.96%	60.95%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	20.48%	28.74%	49.55%	28.72%	39.97%	40.71%	11.86%	34.59%	44.70%	36.97%	56.39%	54.81%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	13.76%	22.70%	44.91%	20.65%	35.51%	39.10%	10.28%	28.92%	36.77%	27.28%	51.24%	53.16%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	23.72%	32.36%	53.31%	33.45%	43.89%	46.42%	12.70%	36.76%	47.25%	39.28%	58.93%	60.03%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	23.36%	32.15%	53.14%	32.92%	43.93%	46.41%	12.60%	36.60%	46.96%	38.96%	58.55%	60.65%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	23.55%	32.26%	53.16%	33.18%	43.84%	46.28%	12.64%	36.64%	47.03%	39.12%	58.45%	59.93%
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	26.32%	34.69%	56.34%	36.16%	46.29%	53.24%	13.35%	38.76%	49.84%	42.44%	61.07%	64.69%
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	26.28%	34.70%	56.35%	36.14%	46.34%	53.23%	13.33%	38.71%	49.75%	42.32%	61.09%	64.80%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	9.94%	16.68%	38.01%	12.49%	28.66%	34.77%	7.89%	25.11%	33.15%	24.70%	46.57%	44.41%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	9.86%	16.62%	37.95%	12.38%	28.65%	34.77%	7.89%	25.11%	33.15%	24.70%	46.57%	44.41%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	9.86%	16.62%	37.95%	12.38%	28.64%	34.77%	7.89%	25.11%	33.15%	24.70%	46.57%	44.41%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	14.03%	22.30%	43.78%	19.80%	33.46%	40.89%	9.70%	29.76%	39.01%	31.42%	51.07%	49.26%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	14.13%	22.42%	44.24%	19.73%	33.98%	42.11%	9.78%	30.17%	39.62%	31.53%	52.22%	55.64%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	14.13%	22.43%	44.26%	19.73%	33.98%	42.11%	9.78%	30.17%	39.62%	31.53%	52.22%	55.64%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	14.78%	22.93%	44.83%	20.19%	34.59%	44.37%	9.82%	30.39%	39.83%	31.86%	52.41%	57.96%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	10.11%	16.80%	38.63%	12.63%	29.43%	35.47%	7.85%	25.45%	33.59%	24.98%	47.46%	50.33%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	10.24%	16.86%	38.74%	12.63%	29.47%	35.40%	7.86%	25.50%	33.64%	25.03%	47.56%	50.02%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	14.73%	22.90%	44.82%	20.14%	34.60%	44.37%	9.82%	30.39%	39.83%	31.86%	52.41%	57.96%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	10.12%	16.80%	38.63%	12.63%	29.43%	35.47%	7.85%	25.45%	33.59%	24.98%	47.46%	50.33%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	10.18%	16.80%	38.65%	12.60%	29.46%	35.40%	7.86%	25.50%	33.64%	25.03%	47.56%	50.02%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	14.73%	22.90%	44.82%	20.14%	34.59%	44.37%	9.82%	30.39%	39.83%	31.86%	52.41%	57.96%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	14.12%	22.43%	44.29%	19.72%	33.98%	42.10%	9.78%	30.17%	39.62%	31.53%	52.22%	55.64%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	14.12%	22.43%	44.29%	19.72%	33.98%	42.10%	9.78%	30.17%	39.62%	31.53%	52.22%	55.64%
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	12.07%	18.47%	39.66%	15.25%	30.17%	32.58%	8.50%	27.40%	36.53%	28.83%	48.95%	43.98%
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	12.10%	18.51%	39.77%	15.33%	30.22%	32.79%	8.54%	27.45%	36.57%	28.85%	49.04%	44.16%
Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	21.75%	30.39%	53.92%	30.06%	43.04%	50.51%	12.21%	35.18%	44.64%	36.08%	57.65%	63.01%
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	26.24%	34.68%	56.36%	36.03%	46.34%	53.21%	13.39%	38.70%	49.72%	42.27%	60.96%	64.66%
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	15.36%	24.80%	46.82%	23.14%	38.81%	47.20%	10.75%	30.85%	39.37%	30.16%	53.49%	58.61%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	18.05%	27.63%	52.27%	25.71%	41.38%	50.53%	11.63%	32.21%	40.55%	30.77%	55.51%	62.51%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	25.04%	33.12%	55.18%	34.05%	44.69%	48.39%	12.93%	37.85%	48.73%	41.27%	60.37%	62.65%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	26.35%	34.76%	56.28%	36.23%	46.26%	53.31%	13.39%	38.76%	49.83%	42.44%	61.02%	64.71%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	17.94%	26.46%	48.48%	25.38%	38.20%	47.22%	11.29%	32.77%	42.54%	34.31%	55.21%	58.06%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	12.40%	19.87%	42.46%	16.55%	32.38%	38.48%	9.30%	27.36%	35.55%	26.65%	49.81%	51.04%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	26.42%	34.74%	56.35%	36.19%	46.18%	53.35%	13.44%	38.74%	49.79%	42.46%	60.90%	64.76%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	26.04%	34.54%	56.18%	35.98%	46.20%	53.71%	13.34%	38.51%	49.44%	42.01%	60.53%	64.64%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	26.14%	34.55%	56.24%	35.99%	46.19%	53.55%	13.39%	38.56%	49.47%	42.07%	60.56%	64.75%

---

Validation pipeline

1. 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.
2. EdgeFirst Studio receives the metrics, predictions, charts and trace, and publishes them as the v-XXXX validation session every row above cites.
3. 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-yolov8{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-yolov8n-det-fp16-640.onnx	FP16	Download
Nano	640/onnx/visdrone-yolov8n-det-fp32-640.onnx	FP32	Download
Small	640/onnx/visdrone-yolov8s-det-fp16-640.onnx	FP16	Download
Small	640/onnx/visdrone-yolov8s-det-fp32-640.onnx	FP32	Download
Medium	640/onnx/visdrone-yolov8m-det-fp16-640.onnx	FP16	Download
Medium	640/onnx/visdrone-yolov8m-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-yolov8n-det-int8-640.tflite	Logical · INT8	Download
Nano	640/tflite/visdrone-yolov8n-det-int8-smart-640.tflite	Smart · INT8	Download
Small	640/tflite/visdrone-yolov8s-det-int8-640.tflite	Logical · INT8	Download
Small	640/tflite/visdrone-yolov8s-det-int8-smart-640.tflite	Smart · INT8	Download
Medium	640/tflite/visdrone-yolov8m-det-int8-640.tflite	Logical · INT8	Download
Medium	640/tflite/visdrone-yolov8m-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-yolov8n-det-int8-640.imx95.tflite	Logical · INT8	Download
Nano	640/imx95/visdrone-yolov8n-det-int8-smart-640.imx95.tflite	Smart · INT8	Download
Small	640/imx95/visdrone-yolov8s-det-int8-640.imx95.tflite	Logical · INT8	Download
Small	640/imx95/visdrone-yolov8s-det-int8-smart-640.imx95.tflite	Smart · INT8	Download
Medium	640/imx95/visdrone-yolov8m-det-int8-640.imx95.tflite	Logical · INT8	Download
Medium	640/imx95/visdrone-yolov8m-det-int8-smart-640.imx95.tflite	Smart · INT8	Download

NXP Ara240 — Ara240 DNPU compiled model.

Size	File	Variant	Download
Nano	640/ara240/visdrone-yolov8n-det-int16-640.dvm	Smart · INT16	Download
Small	640/ara240/visdrone-yolov8s-det-int16-640.dvm	Smart · INT16	Download
Medium	640/ara240/visdrone-yolov8m-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-yolov8n-det-int8-640.hailo8l.hef	INT8	Download
Small	640/hailo/visdrone-yolov8s-det-int8-640.hailo8l.hef	INT8	Download
Medium	640/hailo/visdrone-yolov8m-det-int8-640.hailo8l.hef	INT8	Download

NVIDIA Jetson (TensorRT) — Jetson FP16 and INT8 engines.

Size	File	Variant	Download
Nano	640/jetson/visdrone-yolov8n-det-fp16-640.engine	FP16	Download
Nano	640/jetson/visdrone-yolov8n-det-int8-640.engine	INT8	Download
Small	640/jetson/visdrone-yolov8s-det-fp16-640.engine	FP16	Download
Small	640/jetson/visdrone-yolov8s-det-int8-640.engine	INT8	Download
Medium	640/jetson/visdrone-yolov8m-det-fp16-640.engine	FP16	Download
Medium	640/jetson/visdrone-yolov8m-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-yolov8n-det-w8a16-logical-in8-640.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Nano	640/qnn/visdrone-yolov8n-det-w8a16-logical-tfe-seqmse-in8-640.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Nano	640/qnn/visdrone-yolov8n-det-w8a16-smart-in8-640.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Nano	640/qnn/visdrone-yolov8n-det-w8a16-smart-tfe-seqmse-in8-640.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Nano	640/qnn/visdrone-yolov8n-det-w8a8-logical-640.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Nano	640/qnn/visdrone-yolov8n-det-w8a8-logical-tfe-seqmse-640.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Nano	640/qnn/visdrone-yolov8n-det-w8a8-smart-640.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Nano	640/qnn/visdrone-yolov8n-det-w8a8-smart-tfe-seqmse-640.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Small	640/qnn/visdrone-yolov8s-det-w8a16-logical-in8-640.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Small	640/qnn/visdrone-yolov8s-det-w8a16-logical-tfe-seqmse-in8-640.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Small	640/qnn/visdrone-yolov8s-det-w8a16-smart-in8-640.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Small	640/qnn/visdrone-yolov8s-det-w8a16-smart-tfe-seqmse-in8-640.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Small	640/qnn/visdrone-yolov8s-det-w8a8-logical-640.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Small	640/qnn/visdrone-yolov8s-det-w8a8-logical-tfe-seqmse-640.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Small	640/qnn/visdrone-yolov8s-det-w8a8-smart-640.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Small	640/qnn/visdrone-yolov8s-det-w8a8-smart-tfe-seqmse-640.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Medium	640/qnn/visdrone-yolov8m-det-w8a16-logical-in8-640.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Medium	640/qnn/visdrone-yolov8m-det-w8a16-logical-tfe-seqmse-in8-640.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Medium	640/qnn/visdrone-yolov8m-det-w8a16-smart-in8-640.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Medium	640/qnn/visdrone-yolov8m-det-w8a16-smart-tfe-seqmse-in8-640.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Medium	640/qnn/visdrone-yolov8m-det-w8a8-logical-640.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Medium	640/qnn/visdrone-yolov8m-det-w8a8-logical-tfe-seqmse-640.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Medium	640/qnn/visdrone-yolov8m-det-w8a8-smart-640.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Medium	640/qnn/visdrone-yolov8m-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-yolov8n-det-fp16-1280.onnx	FP16	Download
Nano	1280/onnx/visdrone-yolov8n-det-fp32-1280.onnx	FP32	Download
Small	1280/onnx/visdrone-yolov8s-det-fp16-1280.onnx	FP16	Download
Small	1280/onnx/visdrone-yolov8s-det-fp32-1280.onnx	FP32	Download
Medium	1280/onnx/visdrone-yolov8m-det-fp16-1280.onnx	FP16	Download
Medium	1280/onnx/visdrone-yolov8m-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-yolov8n-det-int8-1280.tflite	Logical · INT8	Download
Nano	1280/tflite/visdrone-yolov8n-det-int8-smart-1280.tflite	Smart · INT8	Download
Small	1280/tflite/visdrone-yolov8s-det-int8-1280.tflite	Logical · INT8	Download
Small	1280/tflite/visdrone-yolov8s-det-int8-smart-1280.tflite	Smart · INT8	Download
Medium	1280/tflite/visdrone-yolov8m-det-int8-1280.tflite	Logical · INT8	Download
Medium	1280/tflite/visdrone-yolov8m-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-yolov8n-det-int8-1280.imx95.tflite	Logical · INT8	Download
Nano	1280/imx95/visdrone-yolov8n-det-int8-smart-1280.imx95.tflite	Smart · INT8	Download
Small	1280/imx95/visdrone-yolov8s-det-int8-1280.imx95.tflite	Logical · INT8	Download
Small	1280/imx95/visdrone-yolov8s-det-int8-smart-1280.imx95.tflite	Smart · INT8	Download
Medium	1280/imx95/visdrone-yolov8m-det-int8-1280.imx95.tflite	Logical · INT8	Download
Medium	1280/imx95/visdrone-yolov8m-det-int8-smart-1280.imx95.tflite	Smart · INT8	Download

NXP Ara240 — Ara240 DNPU compiled model.

Size	File	Variant	Download
Nano	1280/ara240/visdrone-yolov8n-det-int16-1280.dvm	Smart · INT16	Download
Small	1280/ara240/visdrone-yolov8s-det-int16-1280.dvm	Smart · INT16	Download
Medium	1280/ara240/visdrone-yolov8m-det-int16-1280.dvm	Smart · INT16	Download

RPi5 + Hailo-8L (13 TOPS) — Hailo-8L 13 TOPS accelerator.

Size	File	Variant	Download
Nano	1280/hailo/visdrone-yolov8n-det-int8-1280.hailo8l.hef	INT8	Download
Small	1280/hailo/visdrone-yolov8s-det-int8-1280.hailo8l.hef	INT8	Download

NVIDIA Jetson (TensorRT) — Jetson FP16 and INT8 engines.

Size	File	Variant	Download
Nano	1280/jetson/visdrone-yolov8n-det-fp16-1280.engine	FP16	Download
Nano	1280/jetson/visdrone-yolov8n-det-int8-1280.engine	INT8	Download
Small	1280/jetson/visdrone-yolov8s-det-fp16-1280.engine	FP16	Download
Small	1280/jetson/visdrone-yolov8s-det-int8-1280.engine	INT8	Download
Medium	1280/jetson/visdrone-yolov8m-det-fp16-1280.engine	FP16	Download
Medium	1280/jetson/visdrone-yolov8m-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-yolov8n-det-w8a16-logical-in8-1280.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Nano	1280/qnn/visdrone-yolov8n-det-w8a16-logical-tfe-seqmse-in8-1280.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Nano	1280/qnn/visdrone-yolov8n-det-w8a16-smart-in8-1280.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Nano	1280/qnn/visdrone-yolov8n-det-w8a16-smart-tfe-seqmse-in8-1280.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Nano	1280/qnn/visdrone-yolov8n-det-w8a8-logical-1280.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Nano	1280/qnn/visdrone-yolov8n-det-w8a8-logical-tfe-seqmse-1280.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Nano	1280/qnn/visdrone-yolov8n-det-w8a8-smart-1280.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Nano	1280/qnn/visdrone-yolov8n-det-w8a8-smart-tfe-seqmse-1280.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Small	1280/qnn/visdrone-yolov8s-det-w8a16-logical-in8-1280.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Small	1280/qnn/visdrone-yolov8s-det-w8a16-logical-tfe-seqmse-in8-1280.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Small	1280/qnn/visdrone-yolov8s-det-w8a16-smart-in8-1280.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Small	1280/qnn/visdrone-yolov8s-det-w8a16-smart-tfe-seqmse-in8-1280.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Small	1280/qnn/visdrone-yolov8s-det-w8a8-logical-1280.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Small	1280/qnn/visdrone-yolov8s-det-w8a8-logical-tfe-seqmse-1280.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Small	1280/qnn/visdrone-yolov8s-det-w8a8-smart-1280.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Small	1280/qnn/visdrone-yolov8s-det-w8a8-smart-tfe-seqmse-1280.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Medium	1280/qnn/visdrone-yolov8m-det-w8a16-logical-in8-1280.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Medium	1280/qnn/visdrone-yolov8m-det-w8a16-logical-tfe-seqmse-in8-1280.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Medium	1280/qnn/visdrone-yolov8m-det-w8a16-smart-in8-1280.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Medium	1280/qnn/visdrone-yolov8m-det-w8a16-smart-tfe-seqmse-in8-1280.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Medium	1280/qnn/visdrone-yolov8m-det-w8a8-logical-1280.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Medium	1280/qnn/visdrone-yolov8m-det-w8a8-logical-tfe-seqmse-1280.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Medium	1280/qnn/visdrone-yolov8m-det-w8a8-smart-1280.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Medium	1280/qnn/visdrone-yolov8m-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-yolov8n-det-fp16-1344x768.onnx	FP16	Download
Nano	1344x768/onnx/visdrone-yolov8n-det-fp32-1344x768.onnx	FP32	Download
Small	1344x768/onnx/visdrone-yolov8s-det-fp16-1344x768.onnx	FP16	Download
Small	1344x768/onnx/visdrone-yolov8s-det-fp32-1344x768.onnx	FP32	Download
Medium	1344x768/onnx/visdrone-yolov8m-det-fp16-1344x768.onnx	FP16	Download
Medium	1344x768/onnx/visdrone-yolov8m-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-yolov8n-det-int8-1344x768.tflite	Logical · INT8	Download
Nano	1344x768/tflite/visdrone-yolov8n-det-int8-smart-1344x768.tflite	Smart · INT8	Download
Small	1344x768/tflite/visdrone-yolov8s-det-int8-1344x768.tflite	Logical · INT8	Download
Small	1344x768/tflite/visdrone-yolov8s-det-int8-smart-1344x768.tflite	Smart · INT8	Download
Medium	1344x768/tflite/visdrone-yolov8m-det-int8-1344x768.tflite	Logical · INT8	Download
Medium	1344x768/tflite/visdrone-yolov8m-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-yolov8n-det-int8-1344x768.imx95.tflite	Logical · INT8	Download
Nano	1344x768/imx95/visdrone-yolov8n-det-int8-smart-1344x768.imx95.tflite	Smart · INT8	Download
Small	1344x768/imx95/visdrone-yolov8s-det-int8-1344x768.imx95.tflite	Logical · INT8	Download
Small	1344x768/imx95/visdrone-yolov8s-det-int8-smart-1344x768.imx95.tflite	Smart · INT8	Download
Medium	1344x768/imx95/visdrone-yolov8m-det-int8-1344x768.imx95.tflite	Logical · INT8	Download
Medium	1344x768/imx95/visdrone-yolov8m-det-int8-smart-1344x768.imx95.tflite	Smart · INT8	Download

NXP Ara240 — Ara240 DNPU compiled model.

Size	File	Variant	Download
Nano	1344x768/ara240/visdrone-yolov8n-det-int16-1344x768.dvm	Smart · INT16	Download
Small	1344x768/ara240/visdrone-yolov8s-det-int16-1344x768.dvm	Smart · INT16	Download
Medium	1344x768/ara240/visdrone-yolov8m-det-int16-1344x768.dvm	Smart · INT16	Download

RPi5 + Hailo-8L (13 TOPS) — Hailo-8L 13 TOPS accelerator.

Size	File	Variant	Download
Nano	1344x768/hailo/visdrone-yolov8n-det-int8-1344x768.hailo8l.hef	INT8	Download
Small	1344x768/hailo/visdrone-yolov8s-det-int8-1344x768.hailo8l.hef	INT8	Download
Medium	1344x768/hailo/visdrone-yolov8m-det-int8-1344x768.hailo8l.hef	INT8	Download

NVIDIA Jetson (TensorRT) — Jetson FP16 and INT8 engines.

Size	File	Variant	Download
Nano	1344x768/jetson/visdrone-yolov8n-det-fp16-1344x768.engine	FP16	Download
Nano	1344x768/jetson/visdrone-yolov8n-det-int8-1344x768.engine	INT8	Download
Small	1344x768/jetson/visdrone-yolov8s-det-fp16-1344x768.engine	FP16	Download
Small	1344x768/jetson/visdrone-yolov8s-det-int8-1344x768.engine	INT8	Download
Medium	1344x768/jetson/visdrone-yolov8m-det-fp16-1344x768.engine	FP16	Download
Medium	1344x768/jetson/visdrone-yolov8m-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-yolov8n-det-w8a16-logical-in8-1344x768.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Nano	1344x768/qnn/visdrone-yolov8n-det-w8a16-logical-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Nano	1344x768/qnn/visdrone-yolov8n-det-w8a16-smart-in8-1344x768.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Nano	1344x768/qnn/visdrone-yolov8n-det-w8a16-smart-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Nano	1344x768/qnn/visdrone-yolov8n-det-w8a8-logical-1344x768.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Nano	1344x768/qnn/visdrone-yolov8n-det-w8a8-logical-tfe-seqmse-1344x768.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Nano	1344x768/qnn/visdrone-yolov8n-det-w8a8-smart-1344x768.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Nano	1344x768/qnn/visdrone-yolov8n-det-w8a8-smart-tfe-seqmse-1344x768.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Small	1344x768/qnn/visdrone-yolov8s-det-w8a16-logical-in8-1344x768.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Small	1344x768/qnn/visdrone-yolov8s-det-w8a16-logical-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Small	1344x768/qnn/visdrone-yolov8s-det-w8a16-smart-in8-1344x768.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Small	1344x768/qnn/visdrone-yolov8s-det-w8a16-smart-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Small	1344x768/qnn/visdrone-yolov8s-det-w8a8-logical-1344x768.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Small	1344x768/qnn/visdrone-yolov8s-det-w8a8-logical-tfe-seqmse-1344x768.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Small	1344x768/qnn/visdrone-yolov8s-det-w8a8-smart-1344x768.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Small	1344x768/qnn/visdrone-yolov8s-det-w8a8-smart-tfe-seqmse-1344x768.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Medium	1344x768/qnn/visdrone-yolov8m-det-w8a16-logical-in8-1344x768.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Medium	1344x768/qnn/visdrone-yolov8m-det-w8a16-logical-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Medium	1344x768/qnn/visdrone-yolov8m-det-w8a16-smart-in8-1344x768.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Medium	1344x768/qnn/visdrone-yolov8m-det-w8a16-smart-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Medium	1344x768/qnn/visdrone-yolov8m-det-w8a8-logical-1344x768.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Medium	1344x768/qnn/visdrone-yolov8m-det-w8a8-logical-tfe-seqmse-1344x768.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Medium	1344x768/qnn/visdrone-yolov8m-det-w8a8-smart-1344x768.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Medium	1344x768/qnn/visdrone-yolov8m-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-yolov8n-det-fp16-tiled.onnx	FP16	Download
Nano	tiled/onnx/visdrone-yolov8n-det-fp32-tiled.onnx	FP32	Download
Small	tiled/onnx/visdrone-yolov8s-det-fp16-tiled.onnx	FP16	Download
Small	tiled/onnx/visdrone-yolov8s-det-fp32-tiled.onnx	FP32	Download
Medium	tiled/onnx/visdrone-yolov8m-det-fp16-tiled.onnx	FP16	Download
Medium	tiled/onnx/visdrone-yolov8m-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-yolov8n-det-int8-smart-tiled.tflite	Smart · INT8	Download
Nano	tiled/tflite/visdrone-yolov8n-det-int8-tiled.tflite	Logical · INT8	Download
Small	tiled/tflite/visdrone-yolov8s-det-int8-smart-tiled.tflite	Smart · INT8	Download
Small	tiled/tflite/visdrone-yolov8s-det-int8-tiled.tflite	Logical · INT8	Download
Medium	tiled/tflite/visdrone-yolov8m-det-int8-smart-tiled.tflite	Smart · INT8	Download
Medium	tiled/tflite/visdrone-yolov8m-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-yolov8n-det-int8-smart-tiled.imx95.tflite	Smart · INT8	Download
Nano	tiled/imx95/visdrone-yolov8n-det-int8-tiled.imx95.tflite	Logical · INT8	Download
Small	tiled/imx95/visdrone-yolov8s-det-int8-smart-tiled.imx95.tflite	Smart · INT8	Download
Small	tiled/imx95/visdrone-yolov8s-det-int8-tiled.imx95.tflite	Logical · INT8	Download
Medium	tiled/imx95/visdrone-yolov8m-det-int8-smart-tiled.imx95.tflite	Smart · INT8	Download
Medium	tiled/imx95/visdrone-yolov8m-det-int8-tiled.imx95.tflite	Logical · INT8	Download

NXP Ara240 — Ara240 DNPU compiled model.

Size	File	Variant	Download
Nano	tiled/ara240/visdrone-yolov8n-det-int16-tiled.dvm	Smart · INT16	Download
Small	tiled/ara240/visdrone-yolov8s-det-int16-tiled.dvm	Smart · INT16	Download
Medium	tiled/ara240/visdrone-yolov8m-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-yolov8n-det-int8-tiled.hailo8l.hef	INT8	Download
Small	tiled/hailo/visdrone-yolov8s-det-int8-tiled.hailo8l.hef	INT8	Download
Medium	tiled/hailo/visdrone-yolov8m-det-int8-tiled.hailo8l.hef	INT8	Download

NVIDIA Jetson (TensorRT) — Jetson FP16 and INT8 engines.

Size	File	Variant	Download
Nano	tiled/jetson/visdrone-yolov8n-det-fp16-tiled.engine	FP16	Download
Nano	tiled/jetson/visdrone-yolov8n-det-int8-tiled.engine	INT8	Download
Small	tiled/jetson/visdrone-yolov8s-det-fp16-tiled.engine	FP16	Download
Small	tiled/jetson/visdrone-yolov8s-det-int8-tiled.engine	INT8	Download
Medium	tiled/jetson/visdrone-yolov8m-det-fp16-tiled.engine	FP16	Download
Medium	tiled/jetson/visdrone-yolov8m-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-yolov8n-det-w8a16-logical-in8-tiled.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Nano	tiled/qnn/visdrone-yolov8n-det-w8a16-logical-tfe-seqmse-in8-tiled.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Nano	tiled/qnn/visdrone-yolov8n-det-w8a16-smart-in8-tiled.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Nano	tiled/qnn/visdrone-yolov8n-det-w8a16-smart-tfe-seqmse-in8-tiled.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Nano	tiled/qnn/visdrone-yolov8n-det-w8a8-logical-tfe-seqmse-tiled.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Nano	tiled/qnn/visdrone-yolov8n-det-w8a8-logical-tiled.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Nano	tiled/qnn/visdrone-yolov8n-det-w8a8-smart-tfe-seqmse-tiled.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Nano	tiled/qnn/visdrone-yolov8n-det-w8a8-smart-tiled.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Small	tiled/qnn/visdrone-yolov8s-det-w8a16-logical-in8-tiled.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Small	tiled/qnn/visdrone-yolov8s-det-w8a16-logical-tfe-seqmse-in8-tiled.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Small	tiled/qnn/visdrone-yolov8s-det-w8a16-smart-in8-tiled.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Small	tiled/qnn/visdrone-yolov8s-det-w8a16-smart-tfe-seqmse-in8-tiled.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Small	tiled/qnn/visdrone-yolov8s-det-w8a8-logical-tfe-seqmse-tiled.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Small	tiled/qnn/visdrone-yolov8s-det-w8a8-logical-tiled.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Small	tiled/qnn/visdrone-yolov8s-det-w8a8-smart-tfe-seqmse-tiled.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Small	tiled/qnn/visdrone-yolov8s-det-w8a8-smart-tiled.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Medium	tiled/qnn/visdrone-yolov8m-det-w8a16-logical-in8-tiled.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Medium	tiled/qnn/visdrone-yolov8m-det-w8a16-logical-tfe-seqmse-in8-tiled.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Medium	tiled/qnn/visdrone-yolov8m-det-w8a16-smart-in8-tiled.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Medium	tiled/qnn/visdrone-yolov8m-det-w8a16-smart-tfe-seqmse-in8-tiled.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Medium	tiled/qnn/visdrone-yolov8m-det-w8a8-logical-tfe-seqmse-tiled.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Medium	tiled/qnn/visdrone-yolov8m-det-w8a8-logical-tiled.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Medium	tiled/qnn/visdrone-yolov8m-det-w8a8-smart-tfe-seqmse-tiled.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Medium	tiled/qnn/visdrone-yolov8m-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

Model	Dataset	Link
YOLOv8 Detection	COCO	EdgeFirst/yolov8-det
YOLO11 Detection	VisDrone	EdgeFirst/visdrone-yolo11-det
YOLO26 Detection	VisDrone	EdgeFirst/visdrone-yolo26-det

---

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_yolov8_det,
  title = { {YOLOv8 Detection on VisDrone — EdgeFirst Model Zoo} },
  author = {Au-Zone Technologies},
  url = {https://huggingface.co/EdgeFirst/visdrone-yolov8-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.}