OpenAI says its models boosted AMD flagship GPU design by 14 percent in power performance area
OpenAI disclosed that its large language models helped AMD optimise the circuitry of an upcoming flagship graphics processor, achieving a 14 percent improvement in power, performance and area metrics. The same models also delivered single-digit gains on a new CPU design.

Key takeaways
- OpenAI models delivered a 14 percent PPA improvement on an AMD flagship GPU design
- Single-digit PPA gains were also reported on a new AMD CPU design
- The AI work condensed an estimated year of traditional design effort into under five months
- OpenAI is developing specialised chip design models with Synopsys and has prior experience with Broadcom
In this article
OpenAI details chip design collaboration with AMD
OpenAI has revealed new details of its partnership with AMD, stating that its AI models were used to optimise the hardware maker's forthcoming flagship GPU and CPU designs. According to Tom's Hardware, the models worked on SystemVerilog RTL, the language used for high-end chip design, and identified further improvements by rewriting the code that describes the chip's circuits. OpenAI claims the optimisation effort is equivalent to condensing a year of conventional chip design work into under five months, a timeframe it describes as unprecedented in the field.
Measured gains on graphics and processor designs
The company reports a 14 percent improvement in power, performance and area (PPA) for the GPU circuitry. PPA is a combined metric that chip engineers use to evaluate the efficiency of a design across power consumption, speed and silicon area. For the CPU design, OpenAI says it achieved unspecified single-digit PPA gains. The firm adds that its models were reasoning on RTL at the level of an expert microarchitect and expects to squeeze further improvements from the new baseline.
Background on AI-assisted chip development
This is not OpenAI's first foray into chip design. The company previously worked with Broadcom on an inference-targeted processor codenamed Jalapeño, which was running actual workloads on engineering samples as early as June 2026. Small initial deployments were expected around late 2026 with a broader rollout pinned for 2027. Test benchmarks have shown purported power and latency savings versus Nvidia Blackwell GPUs in single-token prediction workloads. OpenAI also tested its high-end models on open-source RISC-V cores such as PicoRV32 and Ibex, tasking them to optimise for throughput under strict area constraints.
Evolution of model capabilities and tooling

In those tests, the models were given 112,000 tokens per run and scored on a scale where zero meant a non-functional or excessively large design and 1.0 represented the best result. The GPT-6.2 Astra model achieved the lowest overall cost and highest optimisation score. A second test showed that models produced better results when redesigning chips from initial specifications rather than optimising existing code. OpenAI notes it is also working with Synopsys to train a specialised model called GPT-Synopsis for use with chip design tools, though it has not confirmed whether unmodified public models were used for the AMD work.
Sources
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



