OpenAI says its artificial intelligence models contributed to the design of Jalapeño, its first custom inference chip, and proposed improvements to processor designs at Advanced Micro Devices (AMD). The company also measured progress across model generations using open-source processor cores and announced work with Synopsys on a model trained to use chip design tools.
Smaller circuits in Jalapeño
OpenAI engineers began using GPT-4o to help design Jalapeño in early 2025. The company says the model was already able to make meaningful suggestions for optimizing power, performance and area together, an industry measure known as PPA.
In the vector unit, which handles parallel computing tasks, the models examined components including floating-point adders and multipliers, conversion and reduction logic, and the special-function unit. They explored alternative logic implementations and ways to divide calculations among pipeline stages, finding opportunities to share hardware and reduce intermediate storage.
In the matrix unit, one redesigned multiplier required about 15% less cell area in synthesis while reducing latency from two clock cycles to one. A separate adder redesign also cut cell area by about 15% while retaining its two-cycle latency. Changes to low-precision dot-product logic and pipeline organization produced further savings.
OpenAI estimates that model-proposed changes reduced standard-cell area overall by about 13% in the vector unit and 10% in the matrix unit. The company says the optimizations were validated during the initial bring-up of Jalapeño silicon.
Work with AMD
OpenAI also worked with AMD on PPA improvements. It says the model achieved gains of up to 14% on flagship GPU design elements compared with heavily optimized baselines. It did so by rewriting SystemVerilog RTL code, which describes a chip’s circuits. OpenAI characterized the result as roughly a year of progress achieved in less than five months.
AMD has already incorporated some of the changes into a design recently sent for manufacturing and added others to production design files for future versions. OpenAI says the model was able to rewrite circuits that perform calculations, control operations and move data. It used Synopsys tools to check that the changes preserved each block’s behavior and to assess trade-offs among power, performance and area.
AMD also verified single-digit percentage PPA gains from model-written designs in key blocks of its upcoming next-generation flagship CPU.
Tests on two open processor cores
To compare chip design capabilities across model generations, OpenAI used two open-source 32-bit RISC-V processor cores. PicoRV32 is a compact core designed for reprogrammable FPGA chips, while OpenTitan uses the Ibex core as the embedded processor in its security chip. Both are complete processor cores for embedded, microcontroller-class workloads.
In the first task, models received the original hardware description of each core and were asked to increase throughput, measured in instructions per second. That measure accounts for the possibility that a change could reduce the number of cycles needed for an instruction while making each clock cycle longer.
The designs were checked in simulation. Those that failed a check or used more than 16 times the original core’s area received a score of zero. For designs that passed, OpenAI measured instructions per clock cycle on workloads hidden from the models, then multiplied the geometric mean by the maximum clock frequency estimated with Yosys. For each core, the best throughput observed under the same constraints received a score of 1.0; averages included runs that scored zero.
With a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra achieved the highest average score on both cores, while GPT-6.1 Sol also outperformed GPT-5.6 Sol. On the existing PicoRV32 design, GPT-6.2 Astra scored about twice as high as GPT-6 Astra at that budget. On Ibex, GPT-6.2 Astra’s average continued to rise at higher budgets, while GPT-6 Astra and GPT-6.1 Sol were closer together.
In the second task, models received written requirements and open-source templates showing how each core connects to the rest of the system. They chose how data would move through the core and how to divide the work into stages. Correctness checks, memory and cache settings, and throughput objectives remained the same as in the first task.
At about 112,000 tokens per run, all five models scored higher when designing a core from requirements than when improving the existing design, for both cores. GPT-6.2 Astra gained more on Ibex than on PicoRV32. The full evaluation included 9,881 scored runs across the two tasks, two cores and all token budgets.
Starting from a PicoRV32 specification, GPT-6.2 Astra’s mean score exceeded 0.7 near the 112,000-token budget and was higher at the next two observed budgets. On Ibex, it averaged about 0.75 near that budget and remained ahead of the other models at larger shared budgets.
OpenAI cautioned, however, that models starting from written requirements could omit features present in the original cores, such as performance counters and additional instructions. Those features may do little to improve throughput on the test workloads, potentially explaining part of the score advantage. The company says the test would better represent production chip design if it required those features, imposed a tighter area limit and checked timing after physical synthesis.
A new model for Synopsys tools
OpenAI and Synopsys are also collaborating to train a new model called GPT-Synopsys. The aim is for it to use Synopsys electronic design automation tools as expert chip designers do. OpenAI says the partnership will ground model development in trusted evaluators and broaden the range of design challenges addressed.



