AI & ComputeLOT 261120 · 4 min read
OpenAI reports 14% PPA gain for AMD’s upcoming flagship GPU
OpenAI reports up to 14% PPA gains for AMD’s upcoming flagship GPU, unspecified single-digit CPU gains, and a year of design effort compressed into under five months.
Key takeaways
- OpenAI reports up to 14% PPA improvement for an upcoming AMD flagship GPU and an unspecified single-digit gain for a new AMD processor.
- OpenAI says AI-assisted optimization compressed the equivalent of one year of conventional chip-design work into under five months.
- OpenAI’s PicoRV32 and Ibex tests used 112,000 tokens per run, with designs failing to work or exceeding 16x the prior area receiving zero.
- GPT-6.2 Astra achieved the lowest overall cost and highest optimization score in the processor-core test.
- Jalapeño engineering samples were running workloads by June 2026; small deployments were targeted for late 2026 and broad rollout for 2027.

By Tom Whitfield
OpenAI says AI models used with AMD produced up to a 14% improvement in power, performance and area for circuitry in an upcoming flagship GPU. The same work delivered an unspecified single-digit PPA gain for a new AMD processor.
These are design-stage results reported by OpenAI, not AMD product specifications or measured commercial-chip performance. The disclosure did not name the chips, disclose their process nodes or wafer sizes, identify the models used for AMD, or break down the 14% figure among the three PPA components.
What changed in AMD’s chip designs?
Engineers used undisclosed AI models to optimize register-transfer-level designs for AMD’s forthcoming GPU and CPU. OpenAI said one model found “further improvements by rewriting the code that describes the chip’s circuits.”
That code was written in SystemVerilog RTL, a language commonly used to describe high-end chip circuitry. The models could therefore propose changes to the design itself, rather than merely adjust compiler settings or software schedules.
OpenAI also said its models reasoned about RTL at what the company called an expert microarchitect level. It expects additional PPA improvements for both the GPU and processor designs from the improved baselines.
The practical result is code-level PPA improvement; silicon-level and commercial benefits remain unconfirmed.
How much development time did the models save?
OpenAI characterized the optimization effort as equivalent to condensing one year of conventional chip-design work into “under 5 months.” The company described that schedule as unprecedented for this type of work.
The comparison covers the claimed optimization effort, not the complete schedule for designing, fabricating and validating an AMD processor. OpenAI presented the figure as evidence that model-guided RTL work can shorten part of the design cycle, while providing no separate accounting for downstream verification.
The time claim matters only if future designs reproduce the same pace without transferring more effort to physical implementation and testing.
Where does this fit in OpenAI’s chip program?
OpenAI previously worked with Broadcom on OpenAI Jalapeño, an inference-targeted chip. According to the company, Jalapeño engineering samples were already running actual workloads as far back as June 2026.
OpenAI expected small initial deployments around late 2026 and set a broader rollout for 2027. Published benchmarks also showed what OpenAI described as strong power and latency savings against the market-leading Nvidia Blackwell GPUs in single-token prediction workloads, although the disclosure provided no numerical comparison.
Jalapeño gives OpenAI’s silicon effort a deployment path, while its benchmark results and rollout schedule remain company claims rather than demonstrated commercial volume.
How did GPT perform on PicoRV32 and Ibex?
OpenAI separately described how its high-end large language models evolved for chip-design work. Researchers asked the models to improve the PicoRV32 and Ibex processor cores, giving each run 112,000 tokens and setting throughput as the objective.
Throughput combined instructions per clock with clock-cycle duration. The scoring rules imposed strict limits on both function and design size:
- A design received 0 if it failed to work or used more than 16x the previous area.
- The best design meeting the constraints received a score of 1.0.
- The models received no information about the workloads on which the cores would eventually run.
GPT-6.2 Astra recorded the strongest result by a wide margin. It achieved both the lowest overall cost and the highest optimization score in the test.
A second test found that the models performed better when they effectively redesigned the chips from their original specifications rather than merely optimizing existing code. The core tests favor broad redesign by the strongest model, but they do not identify which model produced AMD’s reported gains.
Which models performed the work?
OpenAI did not state whether the AMD experiment used unmodified GPT models or specialized systems. Task-specific high-end models already exist for organizations able to afford them, but the disclosure did not place AMD’s work in either category.
The company is also working with Synopsys to train GPT-Synopsis, a model intended to operate chip-design tools. That partnership could connect language-model reasoning with the wider electronic-design automation workflow, though OpenAI gave no performance data for GPT-Synopsis.
Until OpenAI identifies the models used with AMD, the reported PPA gains cannot be tied to a specific GPT version or reproduced independently.
The next test is whether the claimed PPA gains survive implementation in a shipping AMD part and whether OpenAI can move Jalapeño from planned late-2026 deployments to a 2027 rollout. Validated gains would give AMD a shorter design cycle and a stronger efficiency case against Nvidia; without hardware results, 14% remains a tool-assisted design claim rather than a commercial advantage.
Tom's Hardware (Source)
More from Tom Whitfield
Show full bio
Staff writer covering companies, deals and strategy in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,150 articles