AI Accelerator Designed, Verified, and Deployed from Scratch in 2 Weeks by AI (arxiv.org)

In a groundbreaking achievement, researchers have developed Redwood, an AI accelerator designed, verified, and deployed entirely by an AI system in just two weeks. This end-to-end solution addresses the increasing disconnect between evolving AI workloads and the lengthy hardware design cycles traditionally required. By integrating hardware and software optimization into a single loop, Redwood autonomously produced its performance model, RTL design, firmware, and other components from high-level human specifications without further human intervention. Notably, it achieved a remarkable coverage rate of 95% using state-of-the-art tools and formal verification methods. The significance of Redwood lies in its potential to transform how AI accelerators are developed, marking the first instance of a production-ready AI accelerator built entirely by an AI for modern AI models. Notably, Redwood boasts a performance-per-watt increase of 3.4 times compared to the Jetson Orin Nano, all while successfully running multi-billion-parameter models like Llama and Qwen. This innovation not only showcases a new frontier in AI hardware design but also sets the stage for recursive self-improvement in AI systems, where deployed models can contribute to designing even more efficient successors.