OpenAI says its models improved AMD GPU design by up to 14 per cent
OpenAI says its models found improvements of up to 14 per cent to AMD’s heavily optimised GPU design in less than five months. AMD has already taken some of the changes into manufacturing, the company said.
OpenAI said in a blog post on 18 November that its models had rewritten hardware descriptions for AMD’s flagship GPU. A hardware description is code written in SystemVerilog that defines a chip’s logic. The model checked every change with Synopsys design software, OpenAI said. It used the tools to confirm that the circuit block still behaved the same and to estimate the change’s effect on power, performance and area (PPA).
The improvement of up to 14 per cent is the best result, measured against AMD’s already heavily optimised designs. OpenAI does not say which blocks or which of the three goals it applies to. The company estimates that the result represents roughly a year of progress achieved in less than five months. OpenAI also says AMD has already included some of the changes in a design sent for manufacturing. In addition, AMD has seen single-digit PPA gains in key blocks of its next-generation processor, the company said. All information about AMD comes from OpenAI, and The Decoder’s report is also based on OpenAI’s announcement.
OpenAI used the same method to design its own Jalapeño inference chip from early 2025, initially with GPT-4o. One multiplier redesigned by the model used about 15 per cent less cell area in synthesis, and its latency fell from two clock cycles to one. OpenAI’s team estimates that the model’s changes reduced standard-cell area by about 13 per cent in the vector unit and 10 per cent in the matrix unit. The improvements work in the finished chip, the company said.
OpenAI compared model generations on two open-source RISC-V processor cores, PicoRV32 and Ibex. The models had to increase the number of instructions the core executes per second. The figure was based on test workloads hidden from the models and on the maximum clock frequency estimated with Yosys. A design that failed the checks or had more than 16 times the original area scored zero. The best observed result was set to 1.0. The score therefore shows how close a model came to the best solution seen, not the speed-up over the original core. With a budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores. On PicoRV32, its score was about twice that of GPT-6 Astra.
In a second task, the models designed a core from a written specification. With the same budget of about 112,000 tokens, all five models scored higher than when improving an existing core. GPT-6.2 Astra averaged about 0.75 on Ibex at this budget. OpenAI itself notes that the models may have left out features of the original cores, such as performance counters and extra instructions. This may explain part of the gap. The test would better reflect production design if it required these features, a stricter area limit and timing checks after physical synthesis, the company said.
Next, OpenAI is working with Synopsys to train GPT-Synopsys, a model intended to use Synopsys’s chip design software (EDA tools) as an expert designer would.