OpenAI Models Autonomously Redesign Chip Circuits

GPT-6 models achieved 14% performance gains and halved latency in semiconductor designs now being integrated into AMD production silicon.

OpenAI has revealed that its frontier AI models are now capable of autonomously redesigning semiconductor circuits, achieving performance gains that typically take human engineers years to reach. In a series of tests involving both internal hardware and AMD's flagship silicon, models including GPT-6.2 Astra and GPT-6.1 Sol demonstrated the ability to optimize Power, Performance, and Area (PPA) by up to 14%. These results are not merely theoretical; OpenAI confirmed the rewrites are production-viable, and AMD has already integrated several AI-generated findings into chip designs that recently moved to manufacturing.

The efficiency of the AI-driven design process significantly outpaced traditional timelines. Using a budget of approximately 112,000 tokens, the models completed roughly one year's worth of complex design optimizations in less than five months. The impact was particularly visible on OpenAI's in-house chip, codenamed Jalapeño. The AI reduced the Jalapeño vector compute unit's cell area by 13% and its multiplier area by 15%. Most notably, the models managed to slash the compute unit's latency in half, bringing it down from two clock cycles to just one.

Beyond Optimization to Autonomous Design

While the AI showed strong results on existing hardware, its performance varied depending on the task's constraints. The models achieved a 14% PPA improvement on an AMD flagship GPU IP and verified single-digit PPA gains on key modules for AMD's next-generation flagship CPU. Interestingly, the AI performed better when tasked with designing a RISC-V core from scratch rather than optimizing an existing architecture. However, OpenAI noted a trade-off: when asked to redesign instead of optimize, the models occasionally omitted features present in the original specifications.

The transition to AI-assisted hardware development is already creating a massive downstream workload for software maintainers. As these optimized designs move toward deployment, Linux engineers are reportedly managing thousands of bugs identified by the AI. As the industry moves forward, the focus shifts to balancing these rapid architectural leaps with the rigorous verification required to ensure feature parity in production environments.