The model didn't tune a chip. It rewrote the circuit — and the change is already in a design headed for tape-out.
OpenAI disclosed on November 18 that its models directly rewrote SystemVerilog RTL — the code that describes how a chip's circuits behave — and squeezed up to 14% PPA (power-performance-area) improvement out of AMD's flagship GPU IP. The number is not the story. The story is that the changes shipped: AMD put some of them into a design recently sent for tape-out, and wrote the rest into production design files for later versions. This is not a demo. It is a delivered change.
How it works: three steps, not parameter tuning
Step one: the model reasons about RTL like an expert microarchitect. It reads the code describing circuit behavior, then restructures computation, reuses hardware, and reroutes data paths. It is not adjusting a few hyperparameters.
Step two: every rewrite has to survive Synopsys EDA verification. The rewritten circuit must behave identically to the original module, and the tools then measure the three-way trade-off between power, performance, and area. A rewrite that fails verification is discarded.
Step three: OpenAI is training a model with Synopsys — GPT-Synopsys — to drive EDA tools directly, the way an expert engineer would. The goal is to move from "can write circuits" to "can operate the full chip-design toolchain."
The numbers: four figures, four different weights
| Scenario | Number | How much it means |
|---|---|---|
| AMD flagship GPU IP | Up to 14% PPA improvement | Against a baseline already heavily optimized by AMD engineers; OpenAI's words: roughly a year's optimization progress in under 5 months |
| AMD next-gen flagship CPU key blocks | Single-digit percentage PPA improvement | Verified, but OpenAI gave no exact number |
| OpenAI's own inference chip, Jalapeño | Vector-unit standard-cell area down ~13%, matrix-unit area down ~10% | Silicon-measured, not a simulation estimate; design to tape-out in 9 months |
| Open-source RISC-V core evaluation | GPT-6.2 Astra scored highest at equal compute | On two open-source cores; designing from a written spec worked better than modifying existing code |
Jalapeño is OpenAI's first self-designed inference chip, designed with GPT-4o since early 2025. It matters because the same method worked both on someone else's flagship GPU and on OpenAI's own chip that already taped out.
Four honest limits, stated up front
First, 14% is a maximum, on one specific GPU IP. It is not "the whole chip is 14% faster." Using that number to claim full-card performance gains would be a category error.
Second, designing from a written spec works better than modifying existing code. OpenAI itself admits that when modifying an existing design, the model can drop functionality that was already in the original.
Third, the next-gen CPU's "single-digit percentage" has no exact number, so its weight can't be judged independently.
Fourth, all of this is vendor-reported, with no independent reproduction yet. The 14% is measured by OpenAI and accepted by AMD — but no third party has run the same flow.
The question worth arguing about
In the chip design flow, which role gets eaten by models first: the verification engineer, or the microarchitect?