An AI system can now take a chip from a written spec to working hardware on its own, and this paper argues that is what finally makes chip design fast, not better code generation.

Redwood is a small, low-power inference chip. 2 human architects wrote the spec, and the AI system generated everything below it: the chip design, the tests, the firmware, and the kernels.

That means a spec change was back on real hardware in under 48 hours.

Redwood ran Qwen3-0.6B on an AMD FPGA board at 12.1 tokens/s, against 28 on an NVIDIA Jetson Orin Nano. Built as a real chip on Samsung 8 nm, the team projects 49 tokens/s at 1.335 W, or 3.4x more tokens per watt than the Jetson.

More importantly, the speed did not come from skipping tests. Every block reached 95% coverage, and the first design sent to the FPGA had 0 bugs.

Hardware could start moving at the pace of model releases, instead of models waiting on hardware.

– arxiv. org/abs/2608.26418

Title: "Redwood: A Frontier AI Accelerator Designed, Verified, and Deployed from Scratch in 2 Weeks by AI"