
OpenAI Models Advance Chip Design for Jalapeño and AMD

OpenAI reports that its frontier models are now producing expert-level chip design improvements, with measured gains in its custom inference chip Jalapeño, third-party AMD hardware, and open-source RISC-V cores. Early work began with GPT-4o on Jalapeño in early 2025. In the vector unit, model-proposed changes reduced standard-cell area by about 13%; in the matrix unit, by about 10%. One multiplier redesign used about 15% less cell area in synthesis while reducing latency from two clock cycles to one, and a separate adder redesign cut cell area by about 15% while keeping two-cycle latency. The team says these optimizations were silicon proven during Jalapeño‘s silicon bring-up.
On AMD, the collaboration reports up to 14% improvement relative to heavily optimized baselines on flagship GPU IP. The model rewrote SystemVerilog RTL and reasoned at the level of an expert microarchitect, using Synopsys tools to verify behavioral preservation and evaluate power, performance, and area trade-offs. AMD has included some changes in a recent design sent for manufacturing and added others to production design files. AMD also verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU from model-written designs; OpenAI expects further gains on top of GPU and CPU baselines. The source describes roughly a year of progress in less than five months.
To track capability across model generations, OpenAI evaluated models on two open-source 32-bit RISC-V cores: PicoRV32, a compact core designed for FPGAs, and OpenTitan’s Ibex embedded processor. Each model received the original hardware description and was asked to increase throughput, measured as instructions executed per second. Designs were checked in simulation; failed checks or designs using more than 16 times the original core area scored zero. Passing designs were scored by geometric-mean instructions per clock cycle on hidden workloads multiplied by Yosys-estimated maximum clock frequency, normalized so the best observed throughput under constraints was 1.0. With comparable budgets around 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores, and GPT-6.1 Sol outscored GPT-5.6 Sol on both. On PicoRV32, GPT-6.2 Astra‘s mean score at that budget is about twice GPT-6 Astra’s; on Ibex, its mean score continues to climb at larger budgets.
In the second task, models designed a RISC-V core from written requirements and open-source templates, choosing data movement and stage partitioning. At roughly 112,000 tokens per run, all five evaluated models scored higher when designing from specification than when optimizing an existing design on both cores; GPT-6.2 Astra gained more on Ibex than on PicoRV32. The full evaluation included 9,881 scored runs across both tasks, cores, and token budgets. Starting from a PicoRV32 specification, GPT-6.2 Astra passed a mean score near 0.7 at about 112,000 tokens and scored higher at the next two observed budgets; on Ibex it averaged about 0.75 near that budget.
The source notes a possible reason for the specification-design advantage: models can omit features present in the original cores, such as performance counters and extra instructions, that may barely affect throughput on test workloads. Requiring those features, tightening circuit-area limits, and checking timing after physical synthesis would make the evaluation more representative of production chip design. To ground capability development in trusted evaluators and broaden the design challenges addressed, OpenAI is partnering with Synopsys to train GPT-Synopsys, a new model that uses Synopsys electronic design automation (EDA) tools the way expert chip designers do.

