SAM2-hiera-tiny-OpenVINO

SAM2-hiera-tiny for Intel NPUs, GPUs and CPUs: sam2_decoder and sam2_encoder from mlboydaisuke/SAM2-hiera-tiny-LiteRT as ONNX (tf2onnx 1.17.0, opset 17), OpenVINO IR with fp32 weights, and precompiled Intel NPU blobs. tf2onnx 1.17.0 drops the bias of 6 FULLY_CONNECTED layers of sam2_decoder; these files were converted with a one-line fix of that handler (convert/patch_tf2onnx.py in the GitHub repo), and their outputs match the .tflite. On the NPU of a Core Ultra 9 386H the blob runs in 7.97 ms (sam2_decoder) and 1972.66 ms (sam2_encoder) per run, numpy input to numpy output.

Scripts, raw records, and the same table for all 52 graphs of LiteRT-Models: https://github.com/john-rocky/windows-zoo

Files

file	bytes	sha256
`sam2_decoder.onnx`	24,932,648	`bef9a8b8f8adb6bb41755f6d9153742ed51d8f51b105fbbfb67ead7d59f449ea`
`openvino/sam2_decoder.xml`	755,917	`775d174179109655fbb26e1e7bc1880cfa02d2dc2f4c7fd31d0d42de8f1800df`
`openvino/sam2_decoder.bin`	24,421,952	`ce3bc5ab2832b822eaa6a9413802e2731120cb7184a01c74b1a8d6e8de617c59`
`npu/sam2_decoder.blob`	21,188,737	`2326a9d867fb98556a44bbb442d2f201477e9ee3453b3e1345f2e5a515675797`
`sam2_encoder.onnx`	134,201,095	`2377332c91031aa842d01d7bd53c1e3b772a2310d9c2925aed2b11e74fb59b4f`
`openvino/sam2_encoder.xml`	890,720	`91c371c1af7b9791f5bf214b4edcd7a7f7142f902481b44658630191801e6a26`
`openvino/sam2_encoder.bin`	134,055,208	`e33d8c1ebbd01b6b272ae008a32dd42b33b67153ae08501c4baeafed7a68163c`
`npu/sam2_encoder.blob`	146,653,313	`e107d56b61d093e1d03be3b3b5d69c5436deb7126d2258ea9ffb5986da196c3b`

The NPU blobs were compiled with OpenVINO 2026.4.1, Intel NPU driver 32.0.100.4841 (OpenVINO reports NPU_DRIVER_VERSION 1004841, NPU_COMPILER_VERSION 524292). If import_model refuses a blob on another driver or OpenVINO version, compile the IR instead (the commented line in the snippet).

How it was measured

MSI Stealth 16 AI+ (Core Ultra 9 386H, RTX 5080 Laptop GPU at a 95 W power limit, Windows 11 build 26200.9711, on AC power), 2026-11-07. On the laptop, each row is one process: load, one run, 5 warmup runs, then the median of 30 timed runs (2 warmup + 10 timed runs when the first run takes more than 3 s), numpy input to numpy output, every runtime at its default settings, one process at a time, each started after the CPU, the RTX and the Intel GPU had gone idle (the thresholds are in METHOD.md of the GitHub repo). Output check: each output against the CPU output in fp32 on the same input (ONNX Runtime's CPU EP); PASS = cosine ≥ 0.99, no NaN or Inf, and top-1 agreement ≥ 0.99 for class scores. The Snapdragon rows are Qualcomm AI Hub profile jobs of the same ONNX files (their own timer). No Snapdragon rows for sam2_decoder: Qualcomm AI Hub measured the file before the bias fix (its numbers are in the GitHub table).

sam2_decoder

processor	runtime	file	ms	load s	output check
Intel AI Boost NPU	OpenVINO 2026.4.1	`npu/` blob (`import_model`)	7.97	0.02	FAIL: output_0_output: cos 0.1192 < 0.99
Intel AI Boost NPU	ONNX Runtime 1.31.1 + OpenVINO EP 1.8.1	`.onnx`, compiled when the session loads	9.82	4.15	PASS (cos ≥ 0.9998)
Intel Graphics (iGPU)	ONNX Runtime 1.31.1 + OpenVINO EP 1.8.1	`.onnx`	11.29	1.19	PASS (cos ≥ 0.9999)
RTX 5080 Laptop (95 W limit)	ONNX Runtime 1.31.1, CUDA EP	fp16 copy of the `.onnx` (not in this repo)	4.58	1.23	PASS (cos ≥ 0.9999)
RTX 5080 Laptop (95 W limit)	ONNX Runtime 1.31.1, CUDA EP	`.onnx`	4.97	1.33	PASS (cos ≥ 0.9999)
Core Ultra 9 386H CPU	ONNX Runtime 1.31.1, CPU EP	`.onnx`	26.88	0.27	reference

The ONNX file against the source .tflite (LiteRT interpreter on the CPU, the same input): PASS, cosine ≥ 0.9999 for every output.

sam2_encoder

processor	runtime	file	ms	load s	output check
Intel AI Boost NPU	OpenVINO 2026.4.1	`npu/` blob (`import_model`)	1972.66	0.14	PASS (cos ≥ 0.9999)
Intel AI Boost NPU	ONNX Runtime 1.31.1 + OpenVINO EP 1.8.1	`.onnx`, compiled when the session loads	106.59	225.12	PASS (cos ≥ 0.9999)
Intel Graphics (iGPU)	ONNX Runtime 1.31.1 + OpenVINO EP 1.8.1	`.onnx`	156.36	1.55	PASS (cos ≥ 0.9999)
RTX 5080 Laptop (95 W limit)	ONNX Runtime 1.31.1, CUDA EP	fp16 copy of the `.onnx` (not in this repo)	28.01	1.27	PASS (cos ≥ 0.9999)
RTX 5080 Laptop (95 W limit)	ONNX Runtime 1.31.1, CUDA EP	`.onnx`	46.41	1.26	PASS (cos ≥ 0.9999)
Core Ultra 9 386H CPU	ONNX Runtime 1.31.1, CPU EP	`.onnx`	2200.98	0.33	reference
Snapdragon X2 Elite NPU	Qualcomm AI Hub, ONNX Runtime QNN EP	`.onnx`, compiled when it loads	191.09 (1 layers on the CPU)	139.87	not compared (AI Hub profile job, 2026-10-30)

The ONNX file against the source .tflite (LiteRT interpreter on the CPU, the same input): PASS, cosine ≥ 0.9999 for every output.

Run it on the Intel NPU

Download this repo, then run the snippets from its folder:

pip install huggingface_hub
hf download mlboydaisuke/SAM2-hiera-tiny-OpenVINO --local-dir SAM2-hiera-tiny-OpenVINO
cd SAM2-hiera-tiny-OpenVINO

OpenVINO (pip install openvino==2026.4.1):

import numpy as np
import openvino as ov

core = ov.Core()
if "NPU" not in core.available_devices:
    raise SystemExit(f"no Intel NPU visible to OpenVINO: {core.available_devices}")
with open("npu/sam2_decoder.blob", "rb") as f:
    compiled = core.import_model(f.read(), "NPU")  # precompiled: no compile at load
# or compile the IR on this machine: compiled = core.compile_model("openvino/sam2_decoder.xml", "NPU")
# replace the zeros with your preprocessed input (float32, the shapes printed by compiled.inputs)
feeds = {i.get_any_name(): np.zeros(list(i.shape), np.float32) for i in compiled.inputs}
outputs = compiled(feeds)
print({o.get_any_name(): outputs[o].shape for o in compiled.outputs})

ONNX Runtime with the OpenVINO EP (pip install onnxruntime==1.31.1 onnxruntime-ep-openvino==1.8.1):

import numpy as np
import onnxruntime as ort
import onnxruntime_ep_openvino as ov_ep

ort.register_execution_provider_library(ov_ep.get_ep_name(), ov_ep.get_library_path())
npu = [d for d in ort.get_ep_devices()
       if d.ep_name == ov_ep.get_ep_name() and d.ep_metadata.get("ov_device") == "NPU"]
if not npu:
    raise SystemExit("no Intel NPU visible to the OpenVINO EP")
so = ort.SessionOptions()
so.add_provider_for_devices(npu, {})
sess = ort.InferenceSession("sam2_decoder.onnx", sess_options=so)  # compiles for the NPU here
# replace the zeros with your preprocessed input (float32)
feeds = {i.name: np.zeros(i.shape, np.float32) for i in sess.get_inputs()}
outputs = sess.run(None, feeds)
print([o.shape for o in outputs])