OpenAI released a technical report on the 18th local time, disclosing progress in AI-assisted chip design. Its in-house inference chip Jalapeño was already assisting GPT-4o in early 2025: the model rewrote circuits such as floating-point adders and multipliers. After synthesis, the new multiplier reduced area by about 15% and cut latency to a single clock cycle; overall, vector compute unit area dropped by about 13% and matrix compute unit area by about 10%, with silicon validation completed.

Accelerating AMD's Flagship, Compressing a Year's Worth of Work The model rewrote SystemVerilog RTL, the circuit description language, and squeezed up to 14% more PPA out of a highly optimized baseline on AMD's flagship GPU IP—compressing about a year of optimization into less than 5 months. Some results have been incorporated into chip designs recently sent for manufacturing, and key modules of the next-generation flagship CPU also demonstrated single-digit percentage gains in validation.
Exam: Building a RISC-V Core from Scratch OpenAI set the task using two open-source RISC-V cores, requiring improved throughput and passing simulation validation. Failing checks or exceeding the original design's area by 16x resulted in a score of zero. There were 9,881 scored runs in total. As a result, GPT-6.2 Astra had the highest average score; when designing a core from scratch, however, the five tested models actually scored higher. OpenAI is candid about the limitations: the model may omit features that are useless for throughput. Next, it will partner with Synopsys to train GPT-Synopsys, enabling the model to command EDA tools like a professional engineer.
Join Now