OpenAI has reported that its advanced AI models successfully redesigned elements of AMD GPU and CPU architecture, resulting in significant performance, power, and area (PPA) improvements. These AI-driven optimizations are described as "production-viable," with AMD confirming the integration of some of these findings into chips currently in manufacturing. This collaboration highlights a new frontier in hardware design, where artificial intelligence assists in achieving gains that would typically require extensive human engineering time.
According to OpenAI's findings, the AI models worked on several key components, including an in-house "Jalapeño" chip, an AMD flagship GPU IP, and modules for AMD's next-generation flagship CPU. The specific improvements varied across these components. For the Jalapeño vector compute unit, AI-generated changes led to a roughly 13% reduction in cell area. The standard cell area for the Jalapeño compute unit saw a 10% decrease, with the multiplier area reduced by approximately 15% and latency cut from two clock cycles to one.
The most substantial gains were observed in the AMD flagship GPU, which experienced a 14% PPA improvement. For AMD's next-gen CPU, the company verified a single-digit PPA improvement resulting from the AI's work.
OpenAI noted that its models, including GPT-6.2 Astra, GPT-6 Astra, GPT-6.1 Sol, and GPT-5.6 Sol, performed better when tasked with designing a RISC-V core from scratch rather than optimizing an existing one. However, the company also acknowledged that models redesigning components from the ground up might omit features present in the original design, presenting a potential challenge.
The integration of generative AI into hardware development presents a complex landscape. While AI can accelerate design processes significantly, as seen in this collaboration with AMD, it also introduces new considerations. OpenAI's work accelerated about a year's worth of design optimization into less than five months. The broader implications of AI in the PC hardware space also include contributing factors to the current memory crisis and challenges in managing AI-identified vulnerabilities in software, such as the thousands found in the Linux kernel. Nonetheless, the potential for AI to drive future performance gains in silicon remains a compelling prospect.