█

AI/TLDR

OpenAI · 2026-11-18 · major

OpenAI models redesign AMD GPU circuits — up to 14% better PPA

In short

OpenAI says its models rewrote SystemVerilog for AMD's flagship GPU IP and found up to 14% better power, performance and area in under five months. AMD put some changes in a design sent for manufacturing.

OpenAI · 2026-11-18 · 3 sources checked

Illustration for Implicator's report on OpenAI model-made circuit changes for AMD's GPU design
Implicator.ai

OpenAI's models rewrote chip circuit code for AMD and for OpenAI's own Jalapeño chip, and some of the changes are going into real silicon.

Key specs

Ppa gain on amd gpu ipup to 14%

Quick facts

MakerOpenAI, with AMD
What changedModel-written SystemVerilog RTL
AMD GPU IP resultUp to 14% PPA vs heavily optimized baselines
Time takenUnder five months
Best model on RISC-V testGPT-6.2 Astra (~112K tokens per run)
CheckingSynopsys tools verify behavior and PPA

What is it?

OpenAI published results on Nov 18, 2026 showing its models improving chip designs. On AMD's flagship GPU IP, model-proposed circuit changes gave up to 14% better power, performance and area (PPA) against heavily optimized baselines, which OpenAI calls about a year of progress done in under five months. AMD included some of the changes in a recent design sent for manufacturing and added others to its production design files.

How does it work?

The models rewrite SystemVerilog RTL, the code that describes how a chip's circuits behave. Synopsys software then checks that each rewrite keeps the block's behavior the same and measures its PPA trade-offs. On an open RISC-V benchmark (PicoRV32 and Ibex cores, 9,881 runs, designs capped at 16x the original area), GPT-6.2 Astra had the highest average score on both cores at about 112,000 tokens per run.

Why does it matter?

Chip blocks that human teams have tuned for years still had room to improve, and a chip maker is shipping the changes, not just reading a paper. OpenAI also used the method on its own Jalapeño inference chip. The figures are self-reported: OpenAI gives no breakdown of the 14% and no measurements from finished AMD silicon.

Who is it for?

chip designers, EDA teams and hardware engineers

Frequently asked questions

Is the 14% gain on AMD's GPU verified on real chips?
No. OpenAI's 14% figure for AMD's flagship GPU IP is an upper bound measured against heavily optimized design baselines, not finished silicon. OpenAI does not split it into power, performance and area, does not name the AMD product, and the coverage notes the post includes no statement from AMD. AMD did put some of the changes into a design sent for manufacturing.
Which OpenAI model did best at chip design?
GPT-6.2 Astra had the highest average score in OpenAI's RISC-V benchmark at a comparable budget of about 112,000 tokens per run. The test used two open cores, PicoRV32 and Ibex, across 9,881 runs. Designs had to pass simulation checks and stay within 16 times the original core's area, and failed designs scored zero.
Did OpenAI use the same method on its own Jalapeño chip?
Yes. OpenAI applied model-proposed circuit changes to Jalapeño, its first custom inference chip, and says the optimizations were validated during silicon bring-up. OpenAI reports area reductions in Jalapeño's compute units, but outlets quote different percentages for which units, so the exact figures are best read in OpenAI's own post rather than in secondhand summaries.
Can other companies use OpenAI models for chip design?
OpenAI is working with Synopsys on GPT-Synopsys, a specialized model that operates Synopsys chip-design tools on OpenAI-hosted infrastructure. The plan was announced under a September 30 agreement, and early customer engagements are underway. Customer design data is encrypted, kept out of model training and covered by configurable retention and access controls.
Does OpenAI have a business stake in AMD?
Yes. Under an October 2025 deal for six gigawatts of AMD GPUs, OpenAI received the right to buy up to 160 million AMD shares at a low fixed price, roughly a 10% stake, with vesting tied to deployment milestones and AMD share-price targets. Readers should weigh the self-reported chip-design results with that partnership in mind.

Sources · 3 outlets

Tags

  • openai
  • amd
  • chip-design
  • gpu
  • systemverilog
  • rtl
  • eda
  • synopsys
  • jalapeno
  • risc-v
  • gpt-6-2-astra
  • ai-for-hardware

← All releases · Learn AI