The News, Explained
On November 18, OpenAI reported how it used AI models to refine chip circuits. It published early results for its Jalapeño inference chip in August; this update explains changes to individual compute blocks. OpenAI estimates model proposals reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit, and says it verified the changes on fabricated silicon. Those percentages refer to the basic circuit cells in each unit. OpenAI gave no corresponding percentage for the entire chip. Source Source
For example, one multiplier used about 15% less cell area in synthesis, the step where design code is mapped to basic circuits. It also took one clock cycle to return a result instead of two. A clock cycle is one tick of the chip’s timing signal. At the same clock speed, a result that takes one tick arrives sooner than one that takes two. These figures describe that particular circuit. Source
For AMD, OpenAI reports up to 14% improvement on design measures that cover power, performance and area, against an already optimised GPU design. OpenAI says AMD included some of the changes in a design sent for manufacturing. In a separate test on two open-source RISC-V cores, throughput came from simulations and estimated clock speeds. A design sent for manufacturing, behaviour verified on a physical chip and a simulation score represent different stages of evidence. Source
OYOPICK’s Take
For a design team, proving that a proposed circuit still gives the same results matters as much as generating lots of options. A smaller multiplier can be useful, but the team still needs to measure how the change affects power and overall speed on the chip. If future reports show power and throughput across several workloads under the same conditions, it will be easier to tell how much of the verified improvement carries over to the product.