ResearchNews
OpenAI models found up to 14% PPA gains on AMD's GPU
OpenAI says its models rewrote SystemVerilog RTL for AMD's flagship GPU IP and delivered up to a 14% improvement in power, performance and area. Some of the changes are already in a design AMD sent for manufacturing.
· 3 min read

What happened
On November 18, OpenAI detailed how it uses its models for chip design (OpenAI). The headline number comes from AMD: up to 14% improvement in PPA (power, performance, area) on the flagship GPU IP, relative to heavily optimized baselines. The model got there by rewriting the SystemVerilog RTL that describes the chip's circuits.
OpenAI calls it roughly a year of progress in under five months. AMD has already included some changes in a recent design sent for manufacturing and added others to production design files for future versions. AMD also verified single-digit percentage PPA gains across key blocks of its next-generation flagship CPU using model-written designs.
How it was checked
The model used Synopsys tools to verify that each rewrite preserved the block's behavior and to evaluate power, performance and area trade-offs. OpenAI says the model reasons about RTL at the level of an expert microarchitect, rewriting circuits that compute, control operations or move data.
Jalapeño: silicon-proven
OpenAI has used models on its own inference chip, Jalapeño, since early 2025. The team estimates that model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit. One multiplier redesign used about 15% less cell area while cutting latency from two clock cycles to one. OpenAI says these optimizations were silicon proven during bring-up.
A RISC-V test across model generations
To track progress between generations, OpenAI tested models on two open-source RISC-V cores, PicoRV32 and Ibex, across 9,881 scored runs. At about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores (earlier: GPT-6.2 Astra and Sol: cheaper, more token-efficient). On PicoRV32 its score is about twice GPT-6 Astra's. All five models scored higher when designing a core from a specification than when improving the existing one; OpenAI notes this may partly be because they could drop features such as performance counters.
Caveats
- 14% is a best case, not an average, and there is no split between power, performance and area.
- The AMD numbers come only from OpenAI's post; products and blocks are not named.
- A design sent for manufacturing is not measured silicon. Only the Jalapeño results are described as silicon proven.
- The RISC-V harness and hidden workloads are not public, so outsiders cannot rerun the evaluation.
What's next
OpenAI is working with Synopsys to train GPT-Synopsys, a model meant to use Synopsys EDA tools the way expert chip designers do. If the pattern from coding carries over to hardware (the model proposes, a deterministic tool verifies), the payoff could be shorter chip development cycles, not just better individual blocks.
Produced with the help of AI, reviewed by an editor. How we make our articles
Reviewed and edited by the Neural Daily desk. About


