OpenAI models design AMD chips with 14% efficiency gain

OpenAI says its artificial intelligence models have helped design AMD's chips, improving the company's flagship graphics processor designs by up to 14 percent and condensing roughly a year of engineering work into less than five months.
The claim comes from a research update published by OpenAI on 18 November, titled "Advancing chip design with OpenAI models". In it, the company says its models rewrote the code that describes a chip's circuits, known as SystemVerilog RTL, to find improvements that AMD's engineers had not already squeezed out of designs that were already heavily optimised.
"We have partnered with AMD on PPA improvements and seen up to 14% improvement relative to heavily optimized baselines on the flagship GPU IP," OpenAI wrote. PPA stands for power, performance and area, the three-way trade-off that governs how good a chip design is. AMD has already included some of the model-written changes in a design it has sent for manufacturing, and added others to its production design files for future versions.
CPU gains verified too
The graphics work is not the whole story. OpenAI says AMD also verified single-digit percentage gains in PPA across key blocks of its upcoming next-generation flagship CPU, using designs written by the company's model. OpenAI says its models can now reason about circuit code "at the level of an expert microarchitect", rewriting circuits that calculate, control operations or move data.
Every proposed change was checked with tools from Synopsys, which verified that each rewrite preserved the block's behaviour and measured the resulting trade-offs between power, performance and area. That verification step matters, because a smaller or faster circuit is only useful if it still does exactly what the original design did.
AMD's chief executive, Lisa Su, welcomed the results on social media on 18 November, saying she was "proud of these fantastic results" and that OpenAI and AMD engineers were "pushing the frontiers of chip design together".
The same approach built OpenAI's own chip
OpenAI first applied the technique to Jalapeño, its first custom inference chip. Engineers began using GPT-o3 mini to help design the chip in early 2025, and OpenAI estimates the model-proposed changes cut standard-cell area by about 13 percent in the vector unit and 10 percent in the matrix unit, with the savings confirmed once the manufactured silicon was tested.
The company also tested successive model generations on open-source RISC-V processor cores, with its newest model, GPT-6.2 Astra, leading the evaluations. OpenAI cautioned that models working from written requirements alone could leave out features present in the original cores, such as performance counters, so human engineers remain in the loop.
Why it matters
If the results hold up, AI-assisted design could shorten one of the slowest steps in building new chips, at a time when demand for AI hardware is outstripping supply. The gains reported so far are measured against already-optimised designs rather than whole chips, so they do not translate directly into a 14 percent faster graphics card. But changes already in a design sent for manufacturing suggest the work is moving from research into real products, and OpenAI says it expects further improvements on top of the GPU and CPU results.