Subscribe
0%
Illustration for: OpenAI says its models rewrote parts of AMD's flagship GPU for up to 14% PPA gains, and some went to the fab. What's proven

OpenAI says its models rewrote parts of AMD's flagship GPU for up to 14% PPA gains, and some went to the fab. What's proven

OpenAI said on November 18 that its models rewrote circuit code in AMD's flagship GPU design for up to 14% power-performance-area gains, and that some of those changes are in a design already sent for manufacturing. It's the first claim of OpenAI-written RTL shipping in an AMD part. But AMD hasn't named the chip, and every number so far comes from OpenAI.

Listen to this article0:00 / 5:07

OpenAI said on Wednesday, November 18, that its models found up to a 14% improvement in power, performance and area (PPA) in selected circuits of AMD's flagship GPU design, measured against baselines AMD had already heavily optimized. According to Unite.AI and Implicator.ai, AMD put some of the model-written changes into a design that has already been sent for manufacturing. Other changes went into production design files for future versions. OpenAI described the result as "roughly a year of progress achieved in less than five months."

AI coding tools have mostly been about software until now. Here, the models are writing SystemVerilog RTL, the hardware-description code that defines a chip's circuits, for a top-tier GPU vendor's flagship part. AI Weekly calls it the first case it has tracked of OpenAI-authored RTL shipping in an AMD design. If the gains hold up in silicon, design-cycle time becomes something AI labs can sell to chipmakers. That's the pitch behind OpenAI's planned GPT-Synopsys product.

What OpenAI claims

Claim Figure Status
AMD flagship GPU, selected circuits Up to 14% PPA vs optimized baseline OpenAI-reported; no chip named
Time compression on GPU work ~1 year of progress in <5 months OpenAI-reported
AMD next-gen flagship CPU "Single-digit" % PPA gains across key blocks, verified by AMD Reported via OpenAI; no AMD first-party statement
OpenAI's Jalapeño inference chip ~13% standard-cell area cut in vector unit, ~10% in matrix unit OpenAI says initial silicon tests support the estimates
Jalapeño design to tapeout 9 months with AI help OpenAI (August results)

How it worked: the models rewrote RTL blocks, and Synopsys tools checked that each rewrite behaved identically to the original and estimated the PPA trade-offs, per Unite.AI and Implicator. Unite.AI adds that on Jalapeño, a multiplier redesign in the matrix unit cut about 15% of cell area and reduced latency from two cycles to one.

Note: Unite.AI describes the 13% Jalapeño cut as being in the "warp engine," while AI Weekly quotes OpenAI's post as saying "vector unit." We've used AI Weekly's quoted wording.

The RISC-V test behind it

OpenAI also published a benchmark built on two small open RISC-V cores, PicoRV32 and Ibex. Ibex is the embedded processor in the OpenTitan security chip. Across 9,881 scored runs of five models, GPT-6.2 Astra had the highest average score on both cores at a budget of about 112,000 tokens per run, according to Unite.AI and Implicator. Any design that failed simulation or grew past 16x the original core's area scored zero. Throughput was estimated from IPC on hidden workloads multiplied by estimated clock speed.

OpenAI's own caveat matters here. All five models scored better when designing from a written spec than when improving existing code, and spec-built designs could leave out features the originals had, such as performance counters. OpenAI said tighter area limits and timing checks after physical synthesis are needed to make the test realistic.

What isn't proven

  • No named product. Neither OpenAI nor AMD has said which GPU or CPU is involved. Implicator notes that the first planned deployment under the October 2025 AMD–OpenAI GPU deal was the Instinct MI450, but the November 18 disclosure doesn't link the two.
  • No breakdown. "Up to 14% PPA" doesn't say how much is power, how much is performance and how much is area. It also applies to selected blocks, not the whole chip.
  • No silicon numbers for AMD. Synopsys equivalence and PPA estimates aren't the same as measured performance on finished chips. AI Weekly notes AMD hasn't issued its own statement, so the 14% figure is OpenAI-reported at both ends.
  • Short tests can mislead. Implicator points to an October 29 University of Tokyo preprint. In it, a modified CPU looked faster over a 10-iteration test but used 9.98% more cycles over 8,000 CoreMark iterations on an FPGA. That study didn't test OpenAI's claims, but it shows why longer validation matters.

Why it matters and what to watch

For chip teams, the practical signal is that frontier models plus formal equivalence checking are now being trusted on production RTL at a major vendor, at least for specific blocks. The commercial play is GPT-Synopsys, a planned model under OpenAI and Synopsys's September 30 deal that would drive Synopsys EDA tools on OpenAI infrastructure. The companies have said customer design data won't be used for training. It hasn't launched.

Watch for three things: AMD naming the part (and ideally quantifying the gain itself), any measured silicon results, and whether Cadence or Siemens EDA answer with their own model partnerships. Until then, read the 14% as a best-case, block-level, vendor-reported number.

OpenAI's own post was not readable for this article. OpenAI's claims above are as reported by Unite.AI, Implicator.ai and AI Weekly.

Sources

  1. Unite.AI — OpenAI Reports Model-Driven Gains in AMD Flagship Chip Designs unite.ai
  2. Implicator.ai — AMD Sends OpenAI Circuit Changes to Manufacturing implicator.ai
  3. AI Weekly — OpenAI Says Models Found 14% PPA Gains in AMD GPU Designs aiweekly.co

#OpenAI#AMD#Synopsys#chip design#RTL#SystemVerilog#Jalapeño#GPT-6.2 Astra#EDA#semiconductors