Ex-NVIDIA GPU architects build Physical AI chip claiming 5x performance, 80% cost cut

快鲤鱼 · wechat · 2026-10-18

Yuke Technology founder Xu Feixiang argues slow robots stem from chip architecture: GPUs are built for LLM throughput, not the millisecond-scale perceive-reason-act-feedback loop of Physical AI. His contrarian take: architecture accounts for 70% of a parallel chip's usability vs 30% for software, so the startup redesigned the processor rather than patching software.

The team includes veterans of multiple NVIDIA GPU generations with full 6nm-class tapeout-to-production experience; all their chips over the past decade lit up on first silicon.
First product Y10 uses a custom, highly programmable instruction set (not fixed ASIC) to balance generality and performance. Simulation vs customers' current x86 + NVIDIA Orin stacks shows 5x performance at 80% lower cost, with sim-to-silicon error within 3%. An FPGA platform ships late this year for testing with leading robot companies.
Go-to-market: start with smart industrial arms (today replacing only 2% of human operation), then humanoids and a small world-model server; target closed-loop latency under 1ms. He believes the industry vastly underestimates robot compute needs — today's level is "maybe a 1985 386 chip" — and NVIDIA's 30-year architecture legacy is the startup's opening.

Original post →

Ex-NVIDIA GPU architects build Physical AI chip claiming 5x performance, 80% cost cut

More from Embodied