OpenAI says its latest AI models are now good enough to make real improvements to the chips that power AI, and it has the results to back that up. The company reports that its models helped reduce circuit area in Jalapeño, its first custom inference chip, and in AMD's production designs. AMD has already put some of those changes into a chip sent for manufacturing.
OpenAI engineers began using GPT-4o to help design Jalapeño in early 2025. Even that early model found ways to improve power, performance and area, often called PPA in chip design. Most of the savings came in two compute blocks. In the vector unit, which handles parallel computing tasks, the model suggested changes to floating-point adders and multipliers, conversion functions, reduction logic and a special-function unit. It also looked for ways to divide work across pipeline stages and share hardware to reduce intermediate storage.
Gains in Jalapeño and AMD's chips
In the matrix unit, an arithmetic component, the gains added up across several circuits. One multiplier redesign used about 15% less cell area during synthesis and cut latency from two clock cycles to one. A separate adder redesign reduced cell area by about 15% while keeping its two-cycle latency. Across the full Jalapeño effort, the team estimates the model's proposed changes reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit. OpenAI says those optimizations were proven in silicon during Jalapeño's bring-up.
The work with AMD went further. OpenAI says its models delivered up to a 14% improvement over heavily optimized baselines on AMD's flagship GPU intellectual property. The model did this partly by rewriting the SystemVerilog register-transfer-level code that describes the chip's circuits. OpenAI describes the result as roughly a year of progress achieved in less than five months. AMD has included some of the changes in a recent design sent for manufacturing and added others to its production design files for future versions.
OpenAI says the model reasoned about the code at the level of an expert microarchitect. It can rewrite circuits that perform calculations, control operations or move data. Synopsys tools checked that each rewrite preserved the block's behavior and weighed the trade-offs in power, performance and area. AMD also reported single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU, using designs written by the model.
Testing on open-source processor cores
To measure how quickly the models are improving, OpenAI tested each one on two open-source 32-bit RISC-V cores. The first, PicoRV32, is a compact core designed for reprogrammable chips known as FPGAs. The second, Ibex, serves as the embedded processor in the security chip for OpenTitan. Both are complete processor cores aimed at embedded, microcontroller-class workloads.
In the first task, the models were given the original hardware description of each core and asked to increase throughput, meaning the number of instructions the core can execute per second. Designs scored zero if they failed a correctness check or used more than 16 times the original core's area. OpenAI then measured performance on workloads hidden from the models and scored each core against the best result observed under the same constraints.
With a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores. GPT-6.1 Sol also outscored GPT-5.6 Sol on both. In the second task, the models designed a core from written requirements and templates rather than improving an existing one. All five models scored higher on that task than on the first. GPT-6.2 Astra reached a mean score of about 0.7 on PicoRV32 and about 0.75 on Ibex near 112,000 tokens. The full evaluation included 9,881 scored runs.
OpenAI cautioned that the design-from-specification test may overstate the results. Starting from requirements lets the models leave out features found in the original cores, such as performance counters and extra instructions, which may do little for throughput on the test workloads. The company said requiring those features, setting a tighter limit on circuit area and checking timing after physical synthesis would make the test more representative of production chip design.
OpenAI is also working with Synopsys, the electronic design automation company, to train a new model, referred to as GPT-Synopsys, to use Synopsys tools the way expert chip designers do. The company said the goal is to ground model development in trusted evaluators and expand the range of design problems the models can handle.
Este artigo foi escrito com a assistência de IA.
News Factory APP - notícias agênticas para impulsionar seu SEO e AEO.