YOLO26-Nano
6%Tiny variant tuned for sub-watt accelerators. Where the Coral TPU still earns its place.
Cross-chip benchmark matrix
Every supported chip, in matched-pair runs from the Fo’c’sle HIL lab. Sortable by any column — click a header. Cells where the chip can’t run this model show Not supported.
| Chip platform | Quant | Latency p50(ms) | Latency p95(ms) | Throughput | Acc. retention(%) | Power(W) | Memory(MB) | Tested by |
|---|---|---|---|---|---|---|---|---|
Q Qualcomm QCS8550 48 TOPS · SoC | INT8 | 4.71 | 6.01 | 212 FPS | 98.5 | 12.9 | 4 | Community |
A Ambarella CV72 32 TOPS · SoC | INT8 | 6.25 | 7.98 | 160 FPS | 97.7 | 5.2 | 4 | Community |
π Raspberry Pi 5 + Hailo HAT 26 TOPS · HAT | INT8 | 7.70 | 9.62 | 130 FPS | 97.9 | 6.9 | 4 | Publisher |
H Hailo-8 26 TOPS · M.2 | INT8 | 8.28 | 10.5 | 121 FPS | 97.5 | 2.3 | 4 | Publisher |
Q Qualcomm QCS6490 12 TOPS · SoC | INT8 | 12.5 | 15.2 | 80 FPS | 98.9 | 8.2 | 4 | Fo’c’sle HIL |
A Ambarella CV5 16 TOPS · SoC | INT8 | 14.3 | 18.2 | 70 FPS | 97.2 | 4.0 | 4 | Community |
R Rockchip RK3588 6 TOPS · SoC | INT8 | 19.8 | 25.9 | 51 FPS | 98.0 | 9.3 | 4 | Publisher |
G Google Coral Edge TPU 4 TOPS · USB | INT8 | 24.1 | 30.1 | 42 FPS | 97.8 | 1.5 | 4 | Community |
I Intel Movidius Myriad X 4 TOPS · SoC | INT8 | 26.3 | 32.3 | 38 FPS | 97.8 | 2.4 | 4 | Fo’c’sle HIL |
M MediaTek Genio 700 4 TOPS · SoC | In progress | |||||||
H Hailo-10H 40 TOPS · M.2 | Not supported | |||||||
Q Snapdragon 8 Gen 3 NPU 45 TOPS · SoC | Not supported | |||||||
N NVIDIA Jetson Orin Nano 40 TOPS · SoM | Not supported | |||||||
N NVIDIA Jetson AGX Orin 275 TOPS · Module | Not supported | |||||||
N NVIDIA Jetson Thor 2070 TOPS · Module | Not supported | |||||||
A AMD Versal AI Edge VE2302 22 TOPS · SoC | Not supported | |||||||
Apple Neural Engine (M4) 38 TOPS · SoC | Not supported | |||||||
HIL conditions
All numbers measured on Fo’c’sle HIL rigs in Tel Aviv (primary), Munich (secondary), and Pittsburgh (robotics). Single-stream, batch-1, real preprocessing, real downstream consumer. p50/p95 are over 10,000-frame steady-state windows after a 30-second warm-up. Power draw is package power, not wall power. Memory footprint is the resident model + activations footprint at peak — not on-disk.
Submitted publisher numbers are accepted only if they reproduce within ±8% of an HIL-lab matched run on the same chip in the same input mode. Otherwise they live separately under the Discussion tab.