OpenAI AI optimized the AMD graphics chip by 14% – the neural network completed a year’s work of engineers in five months

Источник изображения: Milad Fakurian / unsplash.com

OpenAI talked about progress in using its artificial intelligence models in the development of processors. AI is helping to create new circuits and find new ways to speed up and reduce the footprint of existing chips. The models’ abilities were tested on our own OpenAI Jalapeño accelerator and on AMD chips.

  Image source: Milad Fakurian / unsplash.com

Image source: Milad Fakurian / unsplash.com

OpenAI began working on Jalapeño in early 2025, using the GPT-o3 mini model. In the vector block, which is responsible for performing parallel calculations, AI helped optimize adders and multipliers, data conversion blocks and special operations nodes. The model increased the efficiency of logic circuits and redistributed calculations between stages of the pipeline – as a result, the same hardware components began to be used for different operations, and the amount of data temporarily stored in memory was reduced.

By redesigning the multiplier, the model reduced its area by 15% and reduced the operation time from two cycles to one; The adder area was also reduced by 15% while maintaining the same two-cycle delay. The optimizations proposed by AI helped reduce the area of ​​standard cells of a vector block by an estimated 13%, and a matrix block by 10%.

As part of cooperation with AMD, OpenAI models helped improve the characteristics of key blocks of the company’s flagship graphics processor in three areas: power consumption, performance and area (PPA – Power, Performance, Area). According to OpenAI, the model was able to improve aggregate PPA performance by up to 14% relative to already carefully optimized original designs. The model proposed changes to the RTL code of SystemVerilog, a language for describing digital circuits at the hardware level. At the same time, she analyzed not only computing units, but also control and data transmission circuits. OpenAI estimates that in less than five months, it has achieved progress that would have taken about a year with a traditional approach.

  Image source: Thorium / unsplash.com

Image source: Thorium / unsplash.com

AMD has already included some of the proposed changes in one of the projects sent for production, while others have been added to the design documentation for use in future versions of the chips. In addition, the company confirmed multi-percentage point increases in PPA performance across key blocks of its next-generation flagship CPUs developed using OpenAI models. The circuits were tested using Synopsys tools to ensure that the blocks retained their original behavior after the changes, as well as to evaluate trade-offs between power consumption, performance and die area.

OpenAI compared the capabilities of the two generations of models by testing on two open processor cores with RISC-V architecture: PicoRV32, designed for programmable chips, and Ibex, which is used in OpenTitan. The models were tasked with increasing the performance of finished cores and then redesigning the processor based on the requirements. In tests with a comparable volume of calculations – about 112 thousand tokens per run – the best result on both cores was shown by GPT-6.2 Astra, and GPT-6.1 Sol surpassed GPT-5.6 Sol.

When building a processor from scratch, all five models scored higher than when trying to improve an existing design. A total of 9,881 trials were conducted in the study. The advantage of designing a processor from scratch is that the new design can omit certain features of the original processor that have little effect on the results of the given tests. On the other hand, this means that such tests do not fully reflect real-world design conditions: rigorous testing should take into account mandatory functions, set more stringent limits on chip area, and verify the design after physical design.

Next, OpenAI intends to cooperate with Synopsys. The two companies are developing GPT-Synopsys, an AI model that will be able to use computer-aided design and circuit verification tools in the same way as experienced chip designers.

If you notice an error, select it with the mouse and press CTRL+ENTER.