14g Orange Pi Zero Runs Stereo Depth Inference Under 60ms via Hand-written NPU Kernel

Developer yacineMTB ran stereo depth inference on an Orange Pi Zero 3W weighing just 14 grams with a 30mm×65mm PCB (about the size of three US quarters), achieving latency under 60ms. The key was hand-writing custom kernels for the board's NPU, removing any dependence on high-end GPUs like the RTX 4090. The next day he pushed average depth-estimation frame rate to 27.8 FPS, demonstrating the onboard vision inference potential of ultra-cheap, ultra-light edge devices.

Confirmed

• The hardware is an Orange Pi Zero 3W, 14 grams, 30mm×65mm PCB, roughly three US quarters in size (per @yacineMTB).
• Hand-written NPU kernels enabled sub-60ms stereo depth inference without high-end GPUs.
• On Aug 29 the author raised average depth-estimation frame rate to 27.8 FPS, calling the result hard to believe.
• The author noted his control loops typically run at 7–14Hz, so 27.8 FPS leaves ample headroom.
• @WilliamF relayed that the board costs around $30 (secondhand information).

Why it matters

• Hand-written NPU kernels show that onboard NPUs on small low-power devices have substantial untapped compute, allowing vision tasks like depth inference to run without cloud or high-end GPUs.
• With roughly double the headroom over 7–14Hz control needs, 14-gram-class boards could handle weight- and power-sensitive tasks such as onboard real-time depth perception.

2026-08-28 ~ 2026-08-30 · 7 related posts

• [source] Sub-60ms stereo depth inference on a 14g board — yacineMTB · 2026-08-28
• Sub-60ms stereo depth on Orange Pi Zero 3w via custom NPU kernels — _William_F_ · 2026-08-28
• [source] Optimization Record: Running Depth Estimation at 28 FPS on Orange Pi Zero — yacineMTB · 2026-08-29
• [source] Orange Pi Zero Achieves 27.8 FPS Depth Processing — yacineMTB · 2026-08-29

3 near-duplicate retellings: yacineMTB · ReedSealFoss · _jmwebb