Osegi
AI·1 min read

OpenAI models cut AMD GPU chip design by up to 14%

OpenAI says its models squeezed up to 14% more out of AMD GPU circuit designs that were already heavily tuned, in under five months. Some changes are headed for manufacturing.

Close-up of a GPU chip die on a circuit board with abstract glowing circuit traces suggesting AI-assisted design.
Image: البوابة التقنية

OpenAI says its AI models improved the circuit design of AMD's flagship GPUs by up to 14%, and that AMD has already folded some of the changes into a design sent off for manufacturing. The comparison point was not a rough draft. It was designs that had already been heavily optimised.

In an announcement published on 18 November 2026, OpenAI said the work took less than five months. It estimates that is roughly a year of engineering progress. The target was PPA: power consumption, performance and chip area.

How did OpenAI's models improve the AMD chip design?

The model rewrote the code that describes the circuits, which is written in a hardware description language called SystemVerilog. It then used Synopsys tools to check that the rewritten circuits still worked the same way, and to weigh the trade-offs between power, speed and area.

According to OpenAI, AMD put some of the changes into a design it recently sent for manufacturing, and added others to its design files for future versions. The announcement does not name the products that will carry the improvements or say when they will ship.

OpenAI also said AMD verified single-digit percentage gains in PPA across key blocks of its upcoming flagship CPU, using designs the model wrote.

What does the 14% figure actually mean?

Less than it might sound. The number refers to improvements in circuit design. The announcement does not show that a whole GPU runs 14% faster, and it does not break down how much each of the three PPA measures improved.

These are also figures from OpenAI describing the collaboration, not independent benchmarks of finished products you can buy. For now, the interesting part is the method: a model editing chip-design code, with established verification software checking its work.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:openaiamdgpuchip designsynopsyssystemverilog

More from this day

See all →

More in: AI

See all →