OpenAI has given more details regarding its collaboration with AMD. Unnamed AI models have been used to improve designs in AMD’s next-generation flagship GPU and CPUs.
OpenAI highlights that “Further improvements were made by re-working the System Verilog(Register Transfer Level) chip description”. In under 5 months, improvements equal to a year’s work can be made in traditional chip design. This has resulted in “unprecedented” performance gains in chip design.
The model was able to reason at “an expert micro-architecture level” in RTL. Further improvements in GPU and CPU PPA are expected.
In testing the large language models, the PicoRV32 and Ibex cores were optimized. For each model, 112,000 tokens were used. Throughput (Instructions per clock and clock cycle times) was measured. If the design failed, or was larger than 16x the original design, then it received a score of 0. Designs in keeping with the constraints of the test, and performing best, received a score of 1.0. The models were not informed what type of workloads would be carried out on the chip. GPT-6.2 Astra performed best in both metrics. It had the lowest total cost and the best optimization score.
However, in another test, it was found that better results were achieved when models were used to re-design chips, rather than optimize existing code. It is unclear if researchers used vanilla GPT models in this test. OpenAI is collaborating with Synopsys to train a new model called GPT-Synopsis for utilizing chip design tools.
OpenAI has also collaborated with Broadcom on an inference-oriented chip, Jalapeño. Engineering samples are being run with real workloads as early as June 2026. Initial small-scale rollouts are expected in late 2026. Large-scale rollouts are expected in 2027. Benchmarks from tests have shown improvements in energy efficiency and low latency in single-token prediction tasks over Nvidia Blackwell GPUs.
Reported by Tom’s Hardware.

