OpenAI published the technical report “Advancing chip design with OpenAI models” on Wednesday (November 18), showing that its AI models are now proposing design improvements that make it into real silicon. The headline is the partnership with AMD: by rewriting the code that describes the circuits (SystemVerilog RTL) of the company's flagship GPU, the models achieved up to 14% better PPA — the combined measure of power, performance and area — against a baseline that was already heavily optimized. According to OpenAI, the work represents roughly a year of optimization progress achieved in less than five months.
The result did not stay on paper: AMD has already included some of the model-proposed changes in a design sent for manufacturing and added the rest to its production design files for future versions. The company also verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU, using designs written by the models.
Jalapeño: the in-house chip that proved the method first
Before AMD, OpenAI tested the approach on Jalapeño, its first custom inference chip. Engineers started using GPT-o3 mini to help design the chip in early 2025, and the model-proposed changes reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit — optimizations proven in silicon during the chip's bring-up. Among the detailed examples, a matrix-unit multiplier redesign used about 15% less cell area while cutting latency from two clock cycles to one. As Geekiki reported in October, Jalapeño already runs on hosts with AMD EPYC Turin CPUs; OpenAI's August report said AI assistance took the chip from initial design to tapeout in nine months.
RISC-V tests track progress across generations
To measure how chip-design capability changed from one model generation to the next, OpenAI tested five models — 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 and Ibex, for a total of 9,881 scored runs. At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores when optimizing existing RTL.
There is an important caveat, made by OpenAI itself: when designing a core from written requirements, the models scored higher than when optimizing an existing design — but starting from a specification, they can omit features found in the original cores, such as performance counters and additional instructions, which could explain part of the advantage. Requiring those features and checking timing after physical synthesis would make the test closer to real chip design.
Next step: GPT-Synopsys
To connect the models' gains to the verification systems engineers already use, OpenAI announced a partnership with Synopsys to train GPT-Synopsys, a model able to use the company's electronic design automation (EDA) tools the way an expert chip designer does. The work builds on the multi-year agreement announced on September 30, 2026. Synopsys said customer design data would not be used to train the model.
The figures deserve some caution: the 14% and the CPU gains are results disclosed by OpenAI — AMD has not published its own numbers, no specific AMD product was named, and there are no measurements from finished AMD silicon. Even so, model-proposed changes already being in a design sent for manufacturing marks the move of AI chip design from demonstration to production.
Sources
- OpenAI: “Advancing chip design with OpenAI models” (November 18, 2026)
- Unite.AI: “OpenAI Reports Model-Driven Gains in AMD Flagship Chip Designs”
- i10X: “OpenAI Chip Design: 14% PPA Gains on AMD's GPU”
Featured image: AMD Ryzen 9 7900X processor; photo by CristoCalis, CC BY-SA 4.0, via Wikimedia Commons.
Transparency: this article was produced with the assistance of artificial intelligence, based on cross-checking public posts on X and sources from the internet. Please consult the original sources to confirm the information.
By Muse