HomeAIAI Chips & ComputeOpenAI Says Models Improved AMD Flagship Chip Design 14%

OpenAI Says Models Improved AMD Flagship Chip Design 14%

OpenAI says its models have produced “up to 14% improvement relative to heavily optimized baselines on the flagship GPU IP” in a chip design collaboration with AMD, in a blog post published on 18 November 2026. The company says the gains came from rewriting SystemVerilog RTL, “represents roughly a year of progress achieved in less than 5 months,” and that “AMD has already included some of these changes in a recent design it sent for manufacturing.” Those figures and that timetable are OpenAI’s. AMD, in an X post the same day, quoted OpenAI president Greg Brockman and said it was “thankful for our partnership with @OpenAI as we use AI to advance chip design, together.” Lisa Su, AMD’s chief executive, wrote: “Proud of these fantastic results – @OpenAI + @AMD engineers pushing the frontiers of chip design together!”

In brief

  • OpenAI reports up to 14% PPA improvement on AMD flagship GPU intellectual property, against what it calls heavily optimised baselines, and single-digit percentage PPA gains on blocks of an upcoming next-generation flagship CPU, which it says AMD verified.
  • The “roughly a year of progress” comparison is OpenAI’s wording. AMD’s X posts welcome the work and quote Brockman; they leave the 14% figure unrepeated.
  • OpenAI also says its models earlier reduced standard-cell area on its own Jalapeño inference chip by about 13% in the vector unit and 10% in the matrix unit, figures it describes as a team estimate, later “silicon proven.”

OpenAI’s chip design figures are company measurements, welcomed on X

The blog describes a partnership “on PPA improvements” — power, performance and area, the usual three-way trade-off in chip design. OpenAI says the model “identified further improvements by rewriting the code that describes the chip’s circuits (SystemVerilog RTL),” used Synopsys tools to check that each rewrite kept the block’s behaviour, and evaluated the power, performance and area trade-offs. It adds that AMD “added others to its production design files for use in future versions,” and that “AMD also verified single-digit percentage gains in PPA across key blocks of its upcoming next-generation flagship CPU using designs written by our model.”

That last clause is still OpenAI’s account of what AMD verified. It is not a separate AMD technical paper. The X posts, which TSN verified through the X API, are the company’s public welcome: AMD’s account posted at 19:11:58 GMT on 18 November, and Su’s at 19:48:32 GMT. The AMD post leads with a Brockman quotation: “Working with AMD, we’re able to get a year’s worth of full generational improvements in five months, just using our models.” That is the same “year of progress” framing as the blog, spoken by an OpenAI executive and relayed by AMD, not an AMD laboratory measurement published on its own.

OpenAI first applied the same method to Jalapeño, the custom inference chip it discussed in August. The new post says the team estimates that model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit, and that those optimisations “were silicon proven during Jalapeño’s silicon bring-up.” It also describes earlier stepwise redesigns, including a multiplier that, in OpenAI’s telling, used about 15% less cell area in synthesis and halved latency from two clock cycles to one. Those numbers belong to the Jalapeño programme, not to the AMD GPU result.

A 14% PPA claim is not a finished product announcement

The industrial question is how far a model-written RTL change that looks good in synthesis survives the rest of the tape-out process. OpenAI says some of the AMD GPU changes have already gone into a design sent for manufacturing. That is a stronger claim than a paper study, and it is still a claim from the model vendor, backed by a public congratulation from AMD’s chief executive rather than by a detailed AMD datasheet.

It also sits in a familiar lane of vendor-reported design-tool gains. Cadence has said its Maestro AI models lifted chip-design quality by up to 30%, a different company, a different metric and a different baseline. The OpenAI–AMD chip design work is a collaboration on a production GPU and CPU, not a generic EDA benchmark, but the “up to” qualifier and the “heavily optimized baselines” are doing a lot of work. Without AMD’s own tables, readers cannot see which blocks moved, at which process node, or what was given up in the other two corners of PPA.

The commercial backdrop is already visible on TSN. TensorWave and OpenAI have agreed a 457MW deal for AMD MI455X systems, a capacity plan rather than a design result. AMD’s latest quarter showed data-centre sales nearly doubling, to a record overall figure of $13.4 billion. A 14% PPA improvement, if it holds in silicon, would matter to both of those stories. Until AMD publishes its own measurements, the right attribution remains the one on the blog: OpenAI’s models, OpenAI’s percentages, and an AMD partnership that the company has publicly embraced without repeating the number.

Related stories

Sources

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

More in this category

Latest stories

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.