Task image classification · model-size class mobile-class CNN · batch 1 · 16 platforms
✓ Like-for-like: same model-size class and batch.
| Platform | Result | Model · precision · stack | Grade | Power | Price | Fit perf/power/budget | Source |
|---|---|---|---|---|---|---|---|
| Qualcomm Snapdragon X2 Elite (CRD) | 3953 FPS inference_time b=1 | mobilenet_v2 w8a16 · Qualcomm AI Hub onnx (QAIRT 2.50.0.260828221209) | D | unknown no power published | unknown | – / – / – | Qualcomm AI Hub Models perf.yaml (BSD-3) |
| Hailo-10H (M.2/PCIe, x86 host) | 2798 FPS throughput b=1 | mobilenet_v2_1.0 INT8 (Hailo quantized) · Hailo DFC v5.4.0 / HailoRT (model zoo) | D | 2.50 W reported in result source — spec | unknown | – / – / – | Hailo Model Zoo public models (Hailo-10H) |
| Hailo-8 (M.2/PCIe, x86 host) | 2597 FPS throughput b=1 | mobilenet_v2_1.0 INT8 (Hailo quantized) · Hailo DFC v2.20.0 / HailoRT (model zoo) | D | 2.50 W reported in result source — spec | unknown | – / – / – | Hailo Model Zoo public models (Hailo-8) |
| Hailo-8L (M.2/PCIe, x86 host) | 1738 FPS throughput b=1 | mobilenet_v2_1.0 INT8 (Hailo quantized) · Hailo DFC v2.20.0 / HailoRT (model zoo) | D | 1.50 W reported in result source — spec | unknown | – / – / – | Hailo Model Zoo public models (Hailo-8L) |
| Qualcomm Snapdragon X Elite (CRD) | 2203 FPS inference_time b=1 | mobilenet_v2 w8a8 · Qualcomm AI Hub onnx (QAIRT 2.50.0.260828221209) | D | unknown no power published | unknown | – / – / – | Qualcomm AI Hub Models perf.yaml (BSD-3) |
| Qualcomm Dragonwing IQ-9075 EVK | 1965 FPS inference_time b=1 | mobilenet_v2 w8a8 · Qualcomm AI Hub qnn_dlc (QAIRT 2.50.0.260828221209) | D | unknown no power published | unknown | – / – / – | Qualcomm AI Hub Models perf.yaml (BSD-3) |
| Qualcomm Dragonwing RB3 Gen 2 (QCS6490) | 940 FPS inference_time b=1 | mobilenet_v2 w8a8 · Qualcomm AI Hub onnx (QAIRT 2.50.0.260828221209) | D | unknown no power published | unknown | – / – / – | Qualcomm AI Hub Models perf.yaml (BSD-3) |
| Microchip PolarFire SoC Video Kit (VectorBlox CoreVectorBlox) | 672 FPS inference_time b=1 | mobilenet-v1-0.25-128 INT8 (VNNX) · VectorBlox SDK (VNNX), no compression | D | unknown no power published | unknown | – / – / – | Microchip VectorBlox SDK tutorial metrics |
| NXP i.MX 9 with eIQ Neutron-S NPU (i.MX 95/952/943; doc does not specify) | 619 FPS inference_time b=1 | mobilenetv2-12 (neutron-converted ONNX) INT8 (quantized) · ONNX Runtime 1.24.3 onnxruntime_perf_test; delegate: Neutron execution provider | D | unknown no power published | unknown | – / – / – | NXP UG10166 (vendor document) |
| NXP i.MX 95 | 588 FPS inference_time b=1 | mobilenet_v1_1.0_224_quant INT8 (quantized) · TensorFlow Lite 2.19.0 label_image.py (Python); delegate: Neutron Delegate | D | unknown no power published | unknown | – / – / – | NXP UG10166 (vendor document) |
| Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0 | 417 FPS inference_time b=1 | MobileNet v1 (224x224) INT8 · TensorFlow Lite C++ benchmark (Edge TPU runtime) | D | 2.00 W platform spec envelope (not measured with this result) | unknown | – / – / – | Google Coral Edge TPU benchmarks page (Table 1) |
| Google Coral Dev Board (Edge TPU) | 417 FPS inference_time b=1 | MobileNet v1 (224x224) INT8 · TensorFlow Lite C++ benchmark (Edge TPU runtime) | D | 2.00 W platform spec envelope (not measured with this result) | unknown | – / – / – | Google Coral Edge TPU benchmarks page (Table 1) |
| NXP i.MX 8M Plus | 390 FPS inference_time b=1 | mobilenet_v1_1.0_224_quant INT8 (quantized) · TensorFlow Lite 2.19.0 label_image (C++); delegate: VX Delegate (USE_GPU_INFERENCE=0) | D | unknown no power published | unknown | – / – / – | NXP UG10166 (vendor document) |
| NXP i.MX 93 | 256 FPS inference_time b=1 | mobilenet_v1_1.0_224_quant (vela) INT8 (quantized) · TensorFlow Lite 2.19.0 label_image (C++); delegate: Ethos-U Delegate (Vela-compiled model) | D | unknown no power published | unknown | – / – / – | NXP UG10166 (vendor document) |
| Intel Xeon Gold 6154 desktop (CPU only, TFLite) | 19.61 FPS inference_time b=1 | MobileNet v2 (224x224) FP32/INT8 TFLite on CPU · TensorFlow Lite C++ benchmark (Edge TPU runtime) | D | unknown no power published | unknown | – / – / – | Google Coral Edge TPU benchmarks page (Table 1) |
| Coral Dev Board CPU (quad Cortex-A53 @1.5 GHz, CPU only) | 8.20 FPS inference_time b=1 | MobileNet v2 (224x224) TFLite on CPU · TensorFlow Lite C++ benchmark (Edge TPU runtime) | D | unknown no power published | unknown | – / – / – | Google Coral Edge TPU benchmarks page (Table 1) |
Confidence: A audited MLPerf closed, B documented third-party, C community, D vendor-published, E estimate. Unknown is never assumed. Other like-for-like groups: mobile-class CNN b1 (16); mid-size CNN b1 (13); nano (YOLO-n) b1 (4); mobile-class CNN b8 (3); mid-size CNN b8 (3); large (YOLO-l) b1 (1); medium (YOLO-m) b1 (1); small (YOLO-s) b1 (1); xlarge (YOLO-x) b1 (1); mid-size CNN bunknown (41); mobile-class CNN bunknown (31); unknown b1 (5); unknown bunknown (6)