Skip to content
audiotale.net
Sign in
NewsNews

OpenAI models cut AMD GPU design PPA by up to 14%, Jalapeño area 13%

OpenAI says its models improved heavily optimised AMD GPU circuits, with some changes already in a design sent to manufacturing, and is building GPT-Synopsys with Synopsys.

audiotale.net team

· 6 min read

Written by AI from the sources linked below. How we use AI

On this page
  1. Key points
  2. What happened
  3. Background
  4. The numbers
  5. What it means
  6. What we don't know yet
  7. What to watch next
  8. Sources

OpenAI says its models have found up to 14% improvements in power, performance and area (PPA) on AMD's flagship GPU designs, measured against circuits AMD had already tuned heavily. The company published the results on 18 November 2026, together with details of how its models helped design its own Jalapeño inference chip, a new benchmark that scores models on designing processors, and a partnership with Synopsys to build a model that runs chip-design software. The claim matters because chip design is one of the slowest, most expensive engineering jobs in computing, and the figures here come from work that AMD has partly sent to manufacturing. They are also self-reported by OpenAI, a point we return to below.

We could not open OpenAI's own post from our systems (the page returned an access error), so the figures below come from the reports of The Next Web, Unite.AI, AlphaSignal and The Decoder, which agree with each other on the headline numbers. Where they differ or where only one outlet gives a figure, we say so or leave it out.

Key points

  • OpenAI says its models improved power, performance and area by up to 14% on AMD's flagship GPU intellectual property, against heavily optimised baselines.
  • OpenAI equates the gain to about a year of progress achieved in under five months, and says AMD has put some of the changes in a design sent for manufacturing.
  • On Jalapeño, OpenAI's first custom inference chip, model-proposed changes cut area by about 13% in the warp engine and about 10% in the matrix unit, and were reproduced in the finished silicon.
  • OpenAI's new RISC-V test ran 9,881 scored runs across five model generations; GPT-6.2 Astra scored highest on both cores.
  • OpenAI and Synopsys are training GPT-Synopsys to operate Synopsys design tools; no release date or price has been announced.

What happened

According to the reports, OpenAI's model rewrote SystemVerilog RTL, the code that describes a chip's circuits, for blocks of AMD's flagship GPU design. Each proposed rewrite was checked with Synopsys software to confirm it still behaved like the original, and only candidates that passed were assessed for power, performance and area. OpenAI says the work compressed roughly a year of conventional progress into less than five months.

AMD has included some of the model-written changes in a recent design that was sent for manufacturing, and added others to production design files for future versions, according to the reports. OpenAI also says AMD verified single-digit percentage PPA gains across key blocks of an upcoming flagship CPU using model-written designs.

The 14% figure is "up to", and it is a block-level measure rather than a statement about whole-GPU speed. AlphaSignal notes that OpenAI states no equivalent figure for a complete GPU. Neither OpenAI nor the outlets we read name the specific GPU part or the blocks behind the maximum, and we found no statement from AMD of its own confirming the numbers.

Background

PPA are the three coupled quantities chip designers trade against each other. A smaller circuit can use less power but miss its timing target; a faster one can need more energy or silicon. Teams therefore compare designs under fixed functional, timing and manufacturing limits, which is why a verified gain on an already optimised block is meaningful.

AI in chip design is not new. Tom's Hardware recounts that Synopsys launched DSO.ai, a reinforcement-learning optimiser, in March 2020, and added a generative-AI assistant in November 2023 through a collaboration involving the Azure OpenAI service. Synopsys, Cadence and Siemens EDA have long dominated the tools market, and all three are now pitching autonomous design agents.

OpenAI's own chip work is the other thread. It began using GPT-4o to help design Jalapeño, built with Broadcom, in early 2025, and Tom's Hardware reports the chip was unveiled in June 2026 and went from first RTL to tape-out in about nine months. OpenAI's AMD relationship is also commercial: AlphaSignal reports that OpenAI committed to 6 gigawatts of AMD GPU capacity and holds warrants for up to 160 million AMD shares, equivalent to nearly a 10% stake under the announced terms.

The numbers

On Jalapeño, Unite.AI reports that OpenAI estimates model-proposed changes reduced standard-cell area by about 13% in the warp engine, which manages parallel execution, and by about 10% in the matrix unit, an arithmetic block. The models explored alternative logic, different ways to split computation across pipeline stages, hardware sharing and less intermediate storage. OpenAI says these optimisations were proven in silicon when fabricated chips were brought up.

Reported gains from model-written chip changes%
Reported gains from model-written chip changes (%)
%
AMD GPU IP (up to)14
Jalapeño warp engine area cut13
Jalapeño matrix unit area cut10

Source: OpenAI results as reported by The Next Web and Unite.AI

Two detailed examples are given in Unite.AI's account. A redesigned multiplier in the matrix unit used about 15% less cell area in synthesis and cut latency from two clock cycles to one. A separate adder redesign cut cell area by about 15% while keeping two-cycle latency. Note that the three headline figures measure different things: the AMD number is a combined PPA gain, the Jalapeño numbers are area reductions in specific blocks.

For the benchmark, OpenAI tested GPT-6.2 Astra, GPT-6 Astra, GPT-6.1 Sol, GPT-5.6 Sol and GPT-5.5 on two open-source 32-bit RISC-V cores: PicoRV32, designed for FPGAs, and Ibex, the embedded core in the OpenTitan security chip. There were two tasks: improving an existing core's RTL for throughput, and building a core from written requirements. Across 9,881 scored runs, a design scored zero if it failed a simulation check or used more than 16 times the original area. At about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores in optimisation, and in specification-to-RTL it scored above 0.7 on PicoRV32 and about 0.75 on Ibex. All five models scored higher on the from-scratch task than on optimisation.

What it means

For chip teams, the result suggests a practical role for models: generate many candidate rewrites of blocks, and let established verification and synthesis tools filter them. The hard guarantee stays with the tools, not the model, and engineers remain responsible for correctness, timing and signoff.

For the tools market, the Synopsys deal is the notable move. Tom's Hardware reports that Synopsys and OpenAI signed a multi-year agreement on 30 September to develop GPT-Synopsys, a model that runs Synopsys's tools, interprets their output and iterates toward verified results for an engineer to review. It would run on OpenAI-hosted infrastructure, and the companies plan to share revenue. Competitors Cadence and Siemens are pursuing agents of their own.

For smaller chip companies, the possible upside is access to optimisation work that normally needs large specialist teams, though pricing is unknown.

What we don't know yet

  • The exact AMD parts and blocks behind the 14% figure, and the baseline each was measured against.
  • Whether AMD will confirm the numbers independently. The figures are OpenAI's account of the collaboration, not an independent test of finished hardware.
  • How much the OpenAI-AMD financial relationship, including the warrants, should colour the interpretation.
  • Whether the from-scratch benchmark scores reflect real design skill. OpenAI itself says models could omit features present in the original cores, such as performance counters and extra instructions, which may explain part of the higher scores. A tighter area limit, required features and post-synthesis timing checks would make the test closer to production.
  • When GPT-Synopsys ships and what it costs.

What to watch next

Watch for AMD or Synopsys statements naming the products affected, independent measurements once the manufactured parts are tested, and whether OpenAI publishes the RISC-V benchmark in a form others can rerun. Also watch Cadence and Siemens EDA for comparable announcements. Our earlier coverage of the newest model in the test is in GPT-6.2 Astra and Sol: API prices, context window and what independent tests show.

Sources

Written by

audiotale.net team

The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.