💻 AI Designed the Redwood Accelerator from Scratch in 2 Weeks
Architect Labs published a paper on arXiv about Redwood — an accelerator for low-power inference in robots and drones. Based on a specification by two people, the ALP system autonomously generated RTL code, verification, formal proofs, firmware, and kernels: 95% coverage, with edits reaching hardware in 48 hours.
🌍 If the numbers are confirmed, chip development will be compressed from years to two weeks: according to Wilson Research/Siemens 2024, only 14% of ASIC projects work on the first silicon. Custom silicon will become available to startups, not just giants.
👤 Measured so far only on FPGA: Redwood Nano on an AMD Versal VPK180 board delivers 13 tokens/s on Qwen3-0.6B. The claimed 3.4x per watt against Jetson Orin Nano is a projection to Samsung 8 nm, not a finished chip. This is a demonstration of the method, not the overthrow of NVIDIA.
Source 1: https://arxiv.org/abs/2608.26418 Source 2: https://architectlabs.com/redwood-paper