Key Points
- OpenAI said its models delivered up to 14% PPA gains on AMD's flagship GPU IP.
- AMD has already sent some model-written design changes to manufacturing, with others filed for future chips.
- The work moves AI chip design from research demos into production silicon at two major chipmakers.
The latest:
OpenAI said its frontier models rewrote circuit code for AMD’s flagship GPU IP and produced up to 14% improvements in power, performance and area against heavily optimized baselines. The company described the result as roughly a year of design progress compressed into under five months. AMD has already shipped some of those changes into a recent design sent for manufacturing, according to the blog post.
Details:
- The AMD result: According to OpenAI, its model rewrote SystemVerilog RTL — the code describing a chip’s circuits — and found up to 14% PPA gains on AMD’s flagship GPU IP relative to baselines OpenAI described as already heavily optimized. The company called the pace unprecedented at the frontier of chip design.
- Into production: AMD has incorporated some of the model-generated changes into a recent design already sent for manufacturing, OpenAI said, and added others to its production design files for future versions. The post did not say which product line or when those chips ship.
- CPU gains too: OpenAI said AMD separately verified single-digit percentage PPA improvements across key blocks of its upcoming next-generation flagship CPU, using designs written by the model. OpenAI said it expects further gains on top of both the GPU and CPU work.
- OpenAI’s own chip: Engineers began using GPT-o3 mini in early 2025 on Jalapeño, the company’s first custom inference chip. Across the effort, the team estimates model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit.
- The circuit detail: Savings in the vector unit came from floating-point adders and multipliers, conversion functions, reduction logic and the special-function unit, OpenAI said. One multiplier redesign used about 15% less cell area in synthesis while cutting latency from two clock cycles to one.
- Silicon proven: OpenAI said the Jalapeño optimizations were validated during silicon bring-up, the stage where a fabricated chip is first powered on and tested, rather than remaining synthesis-stage estimates. The post gave no production timeline for Jalapeño itself.
- The benchmark: To track generational progress, OpenAI tested models on two open-source 32-bit RISC-V cores, PicoRV32 and Ibex. At about 112,000 tokens per run, GPT-6.2 Astra scored highest on both cores, with GPT-6.1 Sol outscoring GPT-5.6 Sol. The evaluation covered 9,881 scored runs.
- Verification layer: OpenAI said the model used Synopsys tools to confirm each rewrite preserved the block’s behavior and to weigh power, performance and area trade-offs. The company described the model as reasoning about RTL at the level of an expert microarchitect.
- The Synopsys deal: OpenAI is partnering with Synopsys to train a new model, GPT-Synopsys, to operate electronic design automation tools the way expert chip designers do. The company framed the move as grounding development in trusted evaluators. No launch date or commercial terms were disclosed.
Background:
PPA — power, performance, area — is the core trade-off in chip design: engineers spend years shaving cell area and clock cycles from blocks already near-optimal. Gains in the low double digits against a tuned baseline are unusual, which is why OpenAI benchmarks against heavily optimized designs rather than fresh ones.
Between the lines:
The AMD disclosure matters more than the numbers. Jalapeño is OpenAI’s own chip, so gains there are self-assessed end to end; the AMD work was verified by a third party that then committed it to manufacturing and production files. That is the line between an internal result and an industry one. The Synopsys partnership points the same direction — placing the models inside the verification tools chipmakers already trust.
What’s next
Watch for AMD to identify which shipped design carries the model-written changes, for disclosed PPA figures on its next-generation flagship CPU and GPU, and for a release timeline or availability terms for GPT-Synopsys.