OpenAI says its models improved circuits in its Jalapeño chip and in AMD designs
OpenAI describes how its models have been used for chip design, and gives its own figures. They have not been independently checked.
- Jalapeño, OpenAI's first custom inference chip: OpenAI says its team estimates that model-proposed changes cut standard-cell area by about 13% in the warp engine and 10% in the matrix unit, and that the changes were silicon-proven during bring-up. One multiplier redesign used about 15% less cell area and went from two clock cycles to one.
- AMD: OpenAI says it saw up to 14% better power, performance and area than heavily optimized baselines on AMD's flagship GPU IP, by rewriting the SystemVerilog RTL, and that the model checked each rewrite with Synopsys tools. It says AMD has put some of the changes into a recent design sent for manufacturing, and that AMD verified single-digit percentage gains on key blocks of its next-generation flagship CPU.
- RISC-V tests: OpenAI tested models on two open-source 32-bit cores, PicoRV32 and Ibex, on raising instruction throughput and on designing a core from a written specification. At a budget of about 112,000 tokens per run, it says GPT-6.2 Astra scored highest on both cores, and about twice GPT-6 Astra on the existing PicoRV32 core. Designing from a specification scored higher than improving the existing design for all five models, which OpenAI says may partly be because the models can leave out features such as performance counters. The evaluation covered 9,881 scored runs.
OpenAI also says it is working with Synopsys to train GPT-Synopsys, a model that uses Synopsys electronic design automation tools.
Sources
Discussions
Get news like this as a push.
AI Tracker alerts you when something changes for the tools you use.