Depth-Anything-Edge

41%

Distilled depth model that holds 0.32 AbsRel on KITTI while running 60 FPS on a Pi 5 + Hailo HAT. The reference release.

DepthFocsle-ResearchINT8FP16MIXEDdistilledmonocularkitti

198K downloads 15K deploymentsUpdated Apr 16, 2028

Headline:16.8ms · Raspberry Pi 5 + Hailo HAT · INT8

arXiv

Overview Benchmarks9 Sim Results Deploy9 Files Discussion23

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
N NVIDIA Jetson Thor 2070 TOPS · Module	MIXED	5.85	7.64	171 FPS	99.3	91.7	68	Fo’c’sle HIL
N NVIDIA Jetson AGX Orin 275 TOPS · Module	MIXED	9.52	12.5	105 FPS	99.1	55.1	68	Fo’c’sle HIL
H Hailo-10H 40 TOPS · M.2	INT8	23.8	29.7	42 FPS	97.6	3.3	49	Fo’c’sle HIL
N NVIDIA Jetson Orin Nano 40 TOPS · SoM	INT8	25.1	30.1	40 FPS	97.3	12.2	49	Fo’c’sle HIL
Q Qualcomm QCS8550 48 TOPS · SoC	INT8	26.4	31.7	38 FPS	97.4	11.7	49	Fo’c’sle HIL
Apple Neural Engine (M4) 38 TOPS · SoC	MIXED	29.2	38.7	34 FPS	98.7	6.9	68	Fo’c’sle HIL
A Ambarella CV72 32 TOPS · SoC	INT8	30.2	38.6	33 FPS	97.3	5.7	49	Fo’c’sle HIL
π Raspberry Pi 5 + Hailo HAT 26 TOPS · HAT	INT8	36.6	47.3	27 FPS	97.9	7.6	49	Fo’c’sle HIL
H Hailo-8 26 TOPS · M.2	INT8	40.9	52.1	24 FPS	98.6	2.3	49	Fo’c’sle HIL
Q Qualcomm QCS6490 12 TOPS · SoC	Not supported
Q Snapdragon 8 Gen 3 NPU 45 TOPS · SoC	Not supported
A Ambarella CV5 16 TOPS · SoC	Not supported
M MediaTek Genio 700 4 TOPS · SoC	Not supported
G Google Coral Edge TPU 4 TOPS · USB	Not supported
I Intel Movidius Myriad X 4 TOPS · SoC	Not supported
A AMD Versal AI Edge VE2302 22 TOPS · SoC	Not supported
R Rockchip RK3588 6 TOPS · SoC	Not supported

Leader per columnKITTI 2015 stereo eval · 1216×352 · monocular

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.