OpenAI revealed that its AI models contributed significant improvements to AMD chip designs, achieving up to 14% PPA (power, performance and area) gains relative to already heavily optimized baselines on the company's flagship GPU IP.
What happened?
In an official post dated November 18, 2026, OpenAI detailed its partnership with AMD. The models rewrote parts of the SystemVerilog RTL code that describes the chip circuits, identifying optimizations that AMD's engineering team had already refined over years. The company estimates this equates to roughly a year of conventional progress in less than five months.
Some of these changes have already been incorporated into a design recently sent for manufacturing, while others were added to production design files for future versions. AMD also verified single-digit percentage PPA gains on key blocks of its upcoming flagship CPU.
Why this matters
This result shows AI acting in the development cycle of the very hardware that runs it. The improvements did not come from early designs but from mature circuits. On OpenAI's own Jalapeño inference chip, model-proposed changes reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit, validated in silicon during bring-up.
The partnership also includes work with Synopsys to train a GPT-Synopsys model capable of using EDA tools the way expert designers do.
What changes in practice
For the semiconductor industry, this indicates that language models can accelerate RTL iterations and verification, shortening design cycles. PPA gains can translate into more energy-efficient chips or relatively better performance, although OpenAI does not detail the exact distribution among power, performance and area. Final products have not yet been named.
Primary source: OpenAI publication.
Image credit: OpenAI | Source: openai.com
By Grok