OpenAI says its models delivered up to 14% gains in power, performance and area (PPA) on AMD flagship GPU designs, measured against heavily optimized baselines, and some of the model-written circuit changes are already in a chip design AMD sent for manufacturing. OpenAI described the GPU work as roughly a year of engineering progress compressed into under five months.
The models rewrote SystemVerilog RTL, the hardware description code that defines how a chip's logic blocks behave. Synopsys tools checked that every rewrite preserved the block's original behavior and then measured the power, performance and area trade-offs. AMD also folded additional model-proposed changes into production design files for future chip versions.
The gains extend beyond graphics. AMD verified single-digit percentage PPA improvements across key blocks of an upcoming next-generation flagship CPU using model-written designs, according to OpenAI's report.
Key facts
- 01OpenAI's models achieved up to 14% PPA gains on AMD flagship GPU IP versus heavily optimized baselines.
- 02OpenAI calls the GPU work roughly a year of progress done in under five months.
- 03Model-proposed changes cut standard-cell area about 13% in Jalapeño's warp engine and about 10% in its matrix unit.
- 04AMD verified single-digit percentage PPA gains on key blocks of an upcoming flagship CPU.
- 05OpenAI and Synopsys are training GPT-Synopsys under a multi-year deal announced September 30, 2026.
OpenAI also detailed results on its own silicon. The company began applying GPT-4o to Jalapeño, its first custom inference chip, in early 2025. Model-proposed changes cut standard-cell area by about 13% in the chip's warp engine and about 10% in its matrix unit.
Inside the matrix unit, one multiplier redesign reduced cell area by about 15% in synthesis and cut latency from two clock cycles to one. A separate adder redesign trimmed about 15% of cell area while keeping two-cycle latency. OpenAI said the optimizations were proven in silicon during Jalapeño bring-up, and the company has previously said AI assistance helped take the chip from initial design to tapeout in nine months.
To compare model generations, OpenAI ran five models, GPT-6.2 Astra, GPT-6 Astra, GPT-6.1 Sol, GPT-5.6 Sol and GPT-5.5, on two open-source RISC-V cores, PicoRV32 and the Ibex core from the OpenTitan project, across 9,881 scored runs. The tasks covered optimizing existing RTL and building a core from a written specification. A design scored zero if it failed simulation or exceeded 16 times the original core's area.
GPT-6.2 Astra posted the highest average on both cores for RTL optimization at about 112,000 tokens per run, and GPT-6.1 Sol beat GPT-5.6 Sol on both. On spec-to-RTL work, GPT-6.2 Astra averaged above 0.7 on PicoRV32 and about 0.75 on Ibex, on a normalized scale where the best observed design scores 1.0. All five models scored higher on spec-to-RTL than on RTL optimization.
OpenAI flagged a caveat in its own benchmark: spec-based scores may be inflated because models sometimes left out features such as performance counters and extra instructions that barely affect throughput on the test workloads. Its suggested fixes are requiring those features, tightening the area limit and checking timing after physical synthesis. The 14% AMD figure is a best-case, self-reported result for GPU blocks OpenAI did not name, and the block-level baselines have not been published.
Next, OpenAI is working with Synopsys to train a model called GPT-Synopsys to operate electronic design automation tools the way expert chip designers do. The effort builds on a multi-year agreement the two companies announced on September 30, 2026, covering joint development, go-to-market and revenue sharing.
Synopsys said engineers would set goals ranging from PPA optimization through timing and verification closure, with agents running the tools and iterating for engineer review. The company said customer design data would not be used to train GPT-Synopsys and would be encrypted at rest and in transit. Early technical work with customers is underway, with no launch date or pricing announced.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.




