OpenAI has published a detailed report on the application of its artificial intelligence models during the semiconductor design phase. The company demonstrated that neural network models are capable not only of generating alternative digital circuit designs but also of making targeted optimizations to already fine-tuned processor blocks, reducing their area, power consumption, and latency. The results were verified on OpenAI's own Jalapeño accelerator, on AMD chips, and on two open-source RISC-V processor cores. According to the company, the optimization work that would traditionally take an engineering team about a year was completed by the neural network in less than five months.

Jalapeño: Optimizing Vector and Matrix Blocks

Development of the Jalapeño accelerator began in early 2025, and initially, the GPT-o3 mini model was used to work with the circuits. In the vector block, responsible for parallel computing, the neural network optimized adders, multipliers, data conversion blocks, and special operation nodes. The model redistributed computations between pipeline stages so that the same hardware components were used for different operations, and the volume of data temporarily stored in memory was reduced. Specific figures: the redesigned multiplier took up 15% less area, and execution time dropped from two clock cycles to one; the adder's area also decreased by 15% while maintaining a two-cycle latency. Combined, these optimizations reduced the standard cell area of the vector block by approximately 13% and the matrix block by 10%.

Collaboration with AMD: Up to 14% PPA Gains

As part of its partnership with AMD, OpenAI models improved the characteristics of key blocks in the company's flagship GPU across three metrics: power, performance, and area (PPA). According to OpenAI, the combined PPA improvement reached up to 14% relative to already thoroughly optimized baseline designs. The neural network proposed changes directly in the SystemVerilog RTL code, analyzing not only compute blocks but also control and data transfer circuitry. AMD has already incorporated some of the proposed changes into a production-bound design, while others were added to project documentation for future chip versions. The company also confirmed a multi-percent PPA improvement in key blocks of next-generation flagship CPUs. The correctness of the modified circuits was verified using Synopsys tools, ensuring that the original behavior of the blocks was preserved and evaluating trade-offs between power, performance, and die area.

Generational Comparison: GPT-6.2 Astra vs. GPT-5.6 Sol

To objectively assess progress between model generations, OpenAI conducted tests on two open-source RISC-V processor cores: PicoRV32, designed for programmable logic devices, and Ibex, used in the OpenTitan project. The task was to improve the performance of pre-existing cores and then design a processor from scratch based on a requirements specification. The volume of computation per run was about 112,000 tokens. GPT-6.2 Astra showed the best result on both cores, with GPT-6.1 Sol outperforming GPT-5.6 Sol. When designing a processor from scratch, all five tested models scored higher than when attempting to improve an existing layout, as a new architecture can omit features that have little impact on benchmark results. A total of 9,881 tests were conducted during the study.

Methodology Limitations and Plans for GPT-Synopsys

OpenAI emphasizes that tests on open-source cores do not fully reflect real-world industrial development conditions: strict verification requires accounting for mandatory features, imposing tighter chip area constraints, and performing verification after physical design. Nevertheless, the company views the data as an indicator of future direction. The next step will be deep collaboration with Synopsys: the two companies are developing a specialized GPT-Synopsys model that will be able to use electronic design automation (EDA) and verification tools like experienced chip designers, closing the loop from RTL code generation to physical verification.