OpenAI models achieve 14% PPA gains on AMD GPU IP and outperform predecessors in RISC-V design benchmarks

·2026.11.18 09:00

Key point

GPT-6.2 Astra achieved the highest scores in RISC-V core optimization and design tasks, while model-driven changes reduced standard-cell area by 13% in OpenAI's Jalapeño chip.

Details

  • OpenAI's frontier models, including GPT-4o and GPT-6.2 Astra, have demonstrated significant improvements in chip design, achieving up to 14% PPA (power, performance, area) gains on AMD's flagship GPU IP and reducing standard-cell area by 13% in the warp engine of OpenAI's custom Jalapeño inference chip.
  • In RISC-V benchmarks, GPT-6.2 Astra achieved the highest average scores on both PicoRV32 and Ibex cores, outperforming GPT-6.1 Sol and GPT-5.6 Sol, with a score approximately twice that of GPT-6 Astra at a budget of 112,000 tokens per run.
  • Designing a core from written specifications yielded higher scores than optimizing existing RTL for all five tested models, with GPT-6.2 Astra averaging 0.75 on Ibex and 0.7 on PicoRV32 near the 112,000-token budget.
  • OpenAI is partnering with Synopsys to integrate frontier intelligence with electronic design automation (EDA) tools, aiming to train a new model, GPT-Synopsys, to use these tools like expert chip designers.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.