Home / Technology / AI Designs Chips in Weeks, Not Years
AI Designs Chips in Weeks, Not Years
18 Sep
Summary
- AI system designed a chip from spec in under two weeks.
- Changes are verified and on FPGA in less than 48 hours.
- This could allow hardware to adapt to fast-changing AI needs.
Researchers at Architect Labs have unveiled an AI system capable of completing the entire chip design flow, from specification to FPGA implementation, in a matter of weeks. This platform successfully designed a low-power inference chip, dubbed Redwood, with minimal human intervention beyond the initial high-level specification. The entire process took under two weeks, and design modifications could be verified and deployed on an FPGA in under 48 hours.
This AI-driven approach contrasts sharply with conventional methods where design stages are typically frozen, leading to delays or deferred updates. The company's innovation aims to enable hardware teams to adapt quickly to evolving AI workloads, rather than making long-term commitments based on uncertain future needs. Architect Labs believes this reimagines the entire hardware design flow, not just accelerates existing processes.
Measurements on an FPGA showed promise, though projected silicon performance indicates a significant advantage over existing solutions like Nvidia's Jetson Orin Nano. Specifically, projections suggest Redwood Nano could achieve 1.75 times the throughput of the Jetson while consuming roughly half the power, marking a 3.4X improvement in performance per watt. The company is working towards a physical tapeout of Redwood.
