OpenAI reported that its artificial intelligence models have achieved up to 14% improvements in power, performance, and area (PPA) for AMD's flagship GPU designs. The company stated that the AI models identified optimizations by rewriting SystemVerilog register-transfer level (RTL) code, effectively condensing approximately one year of conventional engineering work into less than five months. Some of these AI-generated optimizations have already been incorporated into chip designs sent for manufacturing.
The collaboration extended to AMD's upcoming next-generation flagship CPU, where OpenAI reported single-digit PPA gains across key blocks. While the 14% figure represents improvements in specific GPU circuitry rather than overall chip performance, the company noted that the models were able to reason on RTL at the level of an expert microarchitect. OpenAI expects to achieve further PPA improvements in both GPU and CPU designs moving forward.
In separate testing, OpenAI evaluated its high-end large language models by tasking them with optimizing the PicoRV32 and Ibex open-source RISC-V chip cores. The company reported that its GPT-6.2 Astra model achieved the highest optimization scores and the lowest overall cost during these tests. The research indicated that the models provided superior results when redesigning chips according to initial specifications rather than attempting to optimize existing code.
This semiconductor work follows OpenAI's previous efforts to design its own custom AI inference accelerator, Jalapeño. The company reported that AI-assisted optimizations reduced the standard-cell area of Jalapeño's vector compute unit by approximately 13% and its matrix compute unit by 10%. One specific multiplier redesign reduced synthesized cell area by 15% while cutting operation latency from two clock cycles to one. OpenAI confirmed these optimizations were validated during the silicon bring-up phase.














