Skip to main content

OpenAI says models improved AMD GPU IP by up to 14%

20 NOVEMBER 2026·2 MIN READ·4 SOURCES

OpenAI says it worked with AMD to improve power, performance and area (PPA) on AMD’s flagship GPU IP, reaching up to 14% improvement against heavily optimized baselines.

OpenAI says models improved AMD GPU IP by up to 14%

Key takeaways · 4

  • 01

    OpenAI says it partnered with AMD on PPA improvements to AMD’s flagship GPU IP, reaching up to 14% against heavily optimized baselines.

  • 02

    OpenAI says its model rewrote SystemVerilog RTL, with Synopsys tools checking behavior preservation and PPA trade-offs.

  • 03

    OpenAI says AMD included some model-generated changes in a recent design sent for manufacturing and added other changes to future design files.

  • 04

    OpenAI says model-written designs delivered single-digit PPA gains in key blocks of AMD’s upcoming flagship CPU.

GPU improvements against optimized designs

OpenAI says it partnered with AMD on power, performance and area improvements to AMD’s flagship GPU IP.[1] It reports gains of up to 14% against heavily optimized baselines, rather than against an unoptimized starting point.[1] OpenAI says AMD engineers had already spent years optimizing the designs underlying that comparison.[2] It also says the effort produced roughly a year of progress in less than five months.[1] The reported result is a maximum, not a claim that every GPU block or design received the same improvement.[1]

How the model worked

OpenAI says its model found further improvements by rewriting SystemVerilog RTL, the code that describes chip circuits.[1] It used Synopsys tools to check that each rewrite preserved circuit behavior and to assess the associated PPA trade-offs.[1] OpenAI says it observed the model reasoning about RTL at the level of an expert microarchitect.[1] These details describe a workflow that combines model-generated changes with checks of both behavior and design trade-offs; the reported results do not establish that every proposed rewrite was adopted.[1]

AMD CPU and OpenAI chip results

OpenAI says AMD verified single-digit percentage PPA gains in key blocks of its upcoming next-generation flagship CPU using model-written designs.[1] It says AMD included some model-generated changes in a recent design sent for manufacturing, while adding other changes to production design files for future versions.[1] Separately, OpenAI says frontier models helped develop its first custom inference chip, Jalapeño.[1] It estimates model-proposed changes reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit, and says those optimizations were proven in silicon during bring-up.[1]

Evaluation and wider design work

OpenAI tested models on two open-source 32-bit RISC-V cores and says its evaluation included 9,881 scored runs across tasks, cores and token budgets.[1] At comparable budgets of about 112,000 tokens per run, OpenAI says GPT-6.2 Astra had the highest average score on both cores.[1] It also found that all five tested models scored higher when designing from written requirements than when improving an existing design; OpenAI notes that this approach can leave out features in the original cores.[1] Separately, OpenAI and Synopsys announced a multi-year partnership to develop GPT-Synopsys for semiconductor design workflows.[4]

For chip-design teams, the results make model-assisted RTL changes worth evaluating against existing verification and PPA workflows. OpenAI reports up to 14% improvement against heavily optimized GPU baselines and single-digit PPA gains in key blocks of an upcoming CPU, so teams should distinguish measured results from expectations of future improvements.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 20 November 2026

    OpenAI says models improved AMD GPU IP by up to 14%

Sources

AI fluency, one session a day, built for your work.