OpenAI Enables Models to Modify Circuits; AMD GPU PPA Improved by Up to 14%
21 hour ago / Read about 0 minute
Author:小编   

OpenAI has announced technical advancements in integrating artificial intelligence into the chip design process. Its in-house inference chip, Jalapeño, served as one of the test subjects: in early 2025, GPT-4o participated in circuit rewriting, significantly reducing the area of the compute module it redesigned, with the related design verified on actual silicon. In collaboration with AMD, OpenAI used models to rewrite SystemVerilog RTL, achieving up to a 14% PPA improvement on a highly optimized baseline, greatly compressing the optimization cycle, and some results have been incorporated into recently taped-out chip designs. OpenAI also conducted tests with two open-source RISC-V cores, with GPT-6.2 Astra achieving the highest average score on the related tasks. Additionally, OpenAI will collaborate with Synopsys to train specialized models to operate EDA tools, while noting that the models have a limitation: they may omit functions that do not contribute to throughput.