Skip to content
← Back to Computer science

OpenAI says its models cut its own chip's circuit area by 10–13% and found up to 14% gains for AMD

Some changes are in silicon or on their way to it. Every figure is OpenAI's own, and AMD has not published its own account.

13%less cell area in the vector unit of OpenAI's Jalapeño chip

On 18 November 2026, OpenAI described how its models helped design Jalapeño, its first custom chip for running AI models. OpenAI engineers have used its models on the project since early 2025, starting with GPT-4o. The team estimates the models’ proposed changes cut the area taken by the basic logic cells by about 13% in the vector unit and 10% in the matrix unit. One multiplier redesign used about 15% less area and finished in one clock cycle instead of two. OpenAI says they worked in the manufactured chip.

OpenAI also worked with AMD on the circuit description of AMD’s flagship GPU, which was already heavily optimised. The model rewrote parts of that code and gained up to 14% on a combined measure of power, speed and area. OpenAI says AMD sent a design containing some of the changes for manufacturing, and that AMD verified gains in the single-digit percent range on key blocks of its next flagship CPU. OpenAI calls the GPU work a year of progress in under five months.

To compare its own models, OpenAI ran them on two small open-source RISC-V processors, PicoRV32 and Ibex, across 9,881 scored runs. At about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both. On PicoRV32 its score was about twice that of GPT-6 Astra. Designing a core from a written specification scored higher than improving the existing design for all five models.

OpenAI is also training GPT-Synopsys with Synopsys, a maker of chip design software, to operate those tools directly.

What it means

OpenAI reports that models now make working improvements to chips that human teams had already tuned, and that some are in real hardware. The AMD and Jalapeño figures describe individual blocks, not a whole chip becoming 10–14% better.

Caveats

  • Every figure comes from OpenAI’s post. We found no statement from AMD or Synopsys, and AMD’s confirmations are reported by OpenAI. The Next Web repeats them as OpenAI’s account.
  • The 13% and 10% area cuts are the team’s own estimate, and area is one of three measures. OpenAI gives no figure for power or speed on Jalapeño.
  • The RISC-V test is OpenAI’s own benchmark, with scores scaled to the best result seen. OpenAI says the specification task lets models drop features such as performance counters, which may explain part of its higher scores, and that the test would be more realistic with those required.
  • “Up to 14%” is the best case, and OpenAI does not say how many blocks were tried or how many gained less.
Suggest a correction