OpenAI published details Wednesday of how its models are changing chip design, with the headline claim being up to 14% improvement in power, performance and area on AMD's flagship GPU IP, measured against heavily optimized baselines. Some of the model-proposed changes are already in a design AMD sent for manufacturing; others went into AMD's production design files for future versions. OpenAI says the results represent roughly a year of progress achieved in less than five months.
AMD separately verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU using designs written by OpenAI's models.

From Jalapeño to third-party silicon
OpenAI first reported using its models on chip work in August, on its own custom inference chip, Jalapeño. The new post gives specifics: work began with GPT-4o in early 2025, and model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit, optimizations confirmed as silicon-proven during bring-up. One multiplier redesign used about 15% less cell area while dropping latency from two clock cycles to one. An adder redesign saved a similar 15% in area while keeping two-cycle latency.
The mechanism is the models rewriting SystemVerilog RTL, the code that describes circuits, then verifying with Synopsys tools that each rewrite preserves behavior. OpenAI says the models reason about RTL at the level of an expert microarchitect.
The evals, with caveats OpenAI itself lists
To track capability across model generations, OpenAI tested successive models on two open-source 32-bit RISC-V cores: PicoRV32, a compact FPGA-oriented core, and Ibex, the embedded processor in the OpenTitan security chip. Given the original hardware description and an instruction to raise throughput, models were scored on hidden workloads, with designs zeroed if they failed correctness checks or exceeded 16 times the original core's area. Across roughly 112,000-token budgets and 9,881 scored runs total, GPT-6.2 Astra posted the highest average score on both cores, and GPT-6.1 Sol outscored GPT-5.6 Sol.
Two findings deserve the skepticism OpenAI attaches to them. First, models scored higher designing a core from written requirements than improving an existing design, which partly reflects freedom to drop features like performance counters that the test workloads never used. Second, OpenAI concedes that requiring those features, tightening the area limit and checking timing after physical synthesis would make the benchmark closer to real production design.
The Synopsys angle
The most durable part of the announcement may be the partnership: OpenAI and Synopsys are training GPT-Synopsys, a new model, to operate the EDA tools expert designers actually use. That is a distribution play as much as a capability play. If verification and trade-off analysis happen inside commercial EDA flows, the bottleneck moves from whether models can write RTL to whether chip teams will sign off on designs they did not draw. AMD's tape-out decisions suggest at least one customer is ready.
Sources
- OpenAI: Advancing chip design with OpenAI models (includes the August 2026 Jalapeño reference)
