OpenAI has shown that its language models can do more than just write text and programs—they can also improve chips themselves. On Wednesday, the company published the results of a collaboration with AMD, according to which its models improved the design of an upcoming flagship graphics chip by as much as 14 percent. Three parameters are tracked simultaneously: power consumption, performance, and chip area, known in the industry by the acronym PPA. OpenAI had previously tested the same approach on its own Jalapeño chip, and it also delivered smaller improvements to a future AMD processor.
A Year’s Work in Less Than Five Months
AMD engineers usually spend years fine-tuning their chip designs, with each additional percentage point of improvement becoming increasingly difficult to find. Nevertheless, OpenAI’s model managed to identify new opportunities. It rewrote SystemVerilog code, which is used to describe the circuits inside a chip, and compared the results with versions that AMD had already thoroughly optimized.
OpenAI claims that this approach achieved roughly a year’s worth of progress in less than five months, describing the result as unprecedented in cutting-edge chip design. According to OpenAI, AMD has already incorporated some of the changes into a recent design that it sent to manufacturing. It saved other modifications in its production files for future chip versions.
According to OpenAI, the model reasons about circuits at the level of an experienced microarchitecture designer. It can rewrite parts of a chip that perform calculations, control operations, or move data. To prevent errors, it used tools from Synopsys to verify that each modified component behaved in the same way as the original. It used the same tools to evaluate the benefits and trade-offs of each change in terms of power consumption, performance, and area.
The collaboration does not end with graphics chips. AMD confirmed that designs produced by OpenAI’s model delivered single-digit percentage improvements in key parts of its upcoming next-generation flagship processor. OpenAI expects to push both the graphics chip and the processor even further, although it is already starting from a very well-optimized baseline.
Jalapeño Was the First Test
OpenAI did not begin using artificial intelligence in chip development with AMD. It had already been using AI since the beginning of 2025 on Jalapeño, its first in-house chip, which it is developing with Broadcom and which is used to run completed models. At the time, it turned to GPT-4o, and even this older model found ways to noticeably improve the design.
In the unit that controls and executes parallel computations, savings came from floating-point adders and multipliers, conversion functions, and other computational circuits. The model tested different logic configurations and different ways of distributing computations among individual processing stages, finding places where hardware could be shared and cache memory reduced.
The matrix unit also delivered significant results. One newly designed multiplier occupied roughly 15 percent less area and produced its result in one clock cycle instead of two. A redesigned adder saved a similar amount of area while maintaining its original speed. Overall, the team estimates that the changes proposed by the model reduced the standard-cell area by roughly 13 percent in the first unit and by 10 percent in the matrix unit. These modifications have already been tested on actual silicon as OpenAI brought the first chips online.
According to earlier information, Jalapeño was already running on engineering samples in June 2026. OpenAI plans a limited deployment toward the end of this year and a broader rollout in 2027.
Tests on Open-Source Processors
To determine how the models’ capabilities change from one generation to the next, the company had them work on two freely available 32-bit RISC-V cores. PicoRV32 is a small core designed for FPGA programmable chips, while Ibex serves as an embedded processor in the OpenTitan project’s security chip.
The models were given the original core description and tasked with increasing its throughput, meaning the number of instructions processed per second. The rules were strict. If a design failed simulation checks or occupied more than sixteen times the original area, it received a score of zero. Performance was measured on tasks the models had not seen beforehand, and the best result achieved was assigned a value of 1.0.
With a similar budget of roughly 112,000 tokens per attempt, GPT-6.2 Astra won on both cores. On PicoRV32, it achieved approximately twice the score of its predecessor, GPT-6 Astra. The newer GPT-6.1 Sol outperformed the older GPT-5.6 Sol.
The second test produced a somewhat surprising result. The models were given only written requirements and examples of how the core connects to the rest of the system, and they determined for themselves how data would flow and how the work would be divided among individual stages. All five tested models performed better when designing the core from scratch than when improving an existing design. In this task, GPT-6.2 Astra achieved an average score above 0.7 on PicoRV32 and roughly 0.75 on Ibex. The entire test included nearly ten thousand evaluated attempts.
OpenAI acknowledges that the result may be somewhat skewed. When designing from scratch, the models may have omitted features present in the original cores, such as performance counters or certain instructions, that do not improve performance on the tested tasks. According to the company, a stricter area limit and timing verification after physical design would make the test more representative of actual manufacturing.
What OpenAI Did Not Tell Us
Many details about the collaboration with AMD are missing. OpenAI did not disclose which graphics chip was involved or when AMD sent it to manufacturing. Nor did it publish measurements from finished AMD silicon; verification on an actual chip has so far only been performed for Jalapeño. It is also unknown which specific model worked on the AMD design and whether it was a standard GPT version or a variant customized specifically for chip development.
Moreover, the 14 percent figure applies to one part of the graphics chip and to a combination of three metrics at once. It therefore cannot be interpreted as meaning that AMD’s new graphics card will be 14 percent faster in games.
OpenAI is now training a new model called GPT-Synopsys together with Synopsys. It is intended to learn how to operate electronic design automation tools in the same way experienced developers do.
Sources: openai.com, theinformation.com and tomshardware.com



