Show HN: AutoGPU – AI designs a real 7nm GPU, from Verilog to GDSII (github.com)

A groundbreaking project called AutoGPU has been introduced, demonstrating a transformative leap in AI-driven silicon chip design. This initiative empowers AI agents to autonomously design a 7nm GPU, transitioning from Verilog code to a finalized GDSII layout—an intricate process that traditionally relied heavily on human engineers. With agents managing every facet of the design, including die floorplan creation, RTL editing, and hardening checks, the project showcases a fully automated workflow that leverages 7,000 lines of Markdown to guide the agents’ decision-making processes. The significance of AutoGPU lies in its potential to revolutionize how semiconductor designs are produced, addressing the growing demand for computational power while reducing the time and costs associated with traditional methods. The successfully generated chip includes a 32×32 systolic array for fp8 matrix multiplication, a critical component for advanced AI applications, and meets stringent accuracy requirements. This initiative not only highlights the capabilities of AI in managing complex engineering tasks typically reserved for human expertise, but it also sets a precedent for future chip design workflows, potentially reshaping the landscape of semiconductor manufacturing as the AI/ML community seeks more efficient and scalable solutions.