OpenAI Says Models Found 14% PPA Gains in AMD GPU Designs
TL;DR
- OpenAI says its models found up to 14% power, performance and area gains on AMD's flagship GPU, measured against already heavily optimized baselines.
- AMD folded some model-written edits into a design sent for manufacturing and verified single-digit PPA gains on key blocks of a next-generation flagship CPU.
- Model-proposed edits cut standard-cell area about 13% and 10% in two blocks of OpenAI's Jalapeño inference chip, which moved design to tapeout in nine months.
OpenAI says its models found improvements of up to 14 percent across power, performance and area in selected circuits for AMD's flagship GPU, measured against designs the chipmaker had already heavily optimized. On its own site, the company characterized the result as 'roughly a year of progress achieved in less than five months.'
AMD did not treat the output as a demo. Some of the model-written changes have already been folded into a design sent for manufacturing, with others added to production design files for use in future versions, as TheNextWeb noted. On its next-generation flagship CPU, AMD separately 'verified single-digit percentage PPA gains across key blocks' using edits proposed by OpenAI's models.
The pattern repeats on OpenAI's own silicon. On Jalapeño, its first custom inference chip, 'initial silicon tests support estimated standard-cell area reductions of about 13 percent in its vector unit and 10 percent in its matrix unit.' AI assistance moved the part from initial design to tapeout in nine months. The models rewrote the circuits in SystemVerilog; Synopsys design tools checked that each rewrite still behaved like the original, then measured the tradeoffs.
The Synopsys relationship is the other half of the announcement. OpenAI is training a model called GPT-Synopsys to drive Synopsys' electronic design automation tools 'the way expert chip designers do,' under a multi-year agreement that covers joint development, go-to-market and revenue sharing. The internal workflow becomes a product other designers can buy.
The caveat is OpenAI's own. 'All five tested models scored better when designing from written requirements than when improving existing code,' the company writes, acknowledging that 'models could leave out features present in the originals.' A PPA number is only as useful as what it excludes, and the AMD baseline is the harder test because it was written by experienced engineers. This is one of 19 AMD-tagged stories we've logged in 90 days, and the first where OpenAI-authored RTL is shipping to a fab inside an AMD design.
Originally reported by openai.com
Read the original article →Original headline: OpenAI Says Its Models Find 14% PPA Gains on AMD GPU IP, Partners With Synopsys for GPT-Synopsys