INDEPENDENT PUBLICATION · AMERICAN AMBITION BECOME A FOUNDING SPONSOR NEWS & ANALYSIS / DATED AT PUBLICATION

CHIP DESIGN / LATEST COVERAGE

OpenAI and AMD find new gains in chip design.

OpenAI reports that its Super Intelligence (SI) tools helped AMD improve parts of a flagship graphics chip design that engineers had already optimized.

Improving an existing design.

A chip designer can make a circuit faster, but the redesign might take up more room or draw more power. Finding a better balance is especially valuable when the existing design has already been heavily optimized. In its November 18 engineering report, OpenAI says its work with AMD found improvements of up to 14% on the power, performance and area measures used to compare versions of AMD’s flagship graphics processor design. That figure is a reported design result, not a 14% speed or power claim for an entire chip.

The model suggested changes to the code that describes the circuits. The model used Synopsys design tools to check that the changes preserved the circuits’ behavior and to evaluate the trade-offs. AMD included some of the changes in a design it sent for manufacturing, according to OpenAI. Other changes went into design files for future versions. A design sent for manufacturing is an engineering milestone, not a result from chips already on the market.

What working silicon can tell us.

OpenAI describes a separate result on its own chip, Jalapeño, which is built to run models. The team estimates that model-suggested changes cut the circuit area used by standard components by about 13% in one unit that handles parallel arithmetic and 10% in another that handles matrix calculations. OpenAI says it verified those optimizations while testing the manufactured chip. These are area estimates for two parts of Jalapeño, not whole-chip savings or results from AMD hardware.

That distinction is what makes the two examples useful together. Jalapeño shows what the approach can contribute to a chip OpenAI built itself. AMD shows how the same idea can help engineers explore a design that was already heavily optimized by another team.

Tools engineers already use.

OpenAI is also working with Synopsys to train a model called GPT-Synopsys to use electronic design automation (EDA) software. That’s the software engineers use to turn circuit designs into something manufacturers can build and to check that those designs work. The report describes a collaboration, not a public release of the new model.

The useful part of this work is how ordinary the final question is: does the new design still do its job, and does it do it with a better use of time, power or space? The model gives engineers more ways to answer that question; the engineering checks tell them which ideas are worth keeping.

Keep exploring

The infrastructure behind the work.

Explore American infrastructure ↗