AI NewsModels & agentsAnnouncement
OpenAI says GPT-6.2 cut chip area 13 percent on its Jalapeño silicon
OpenAI says GPT-6.2 Astra cut standard-cell area by 13 percent on its Jalapeño vector unit. The matrix unit fell 10 percent.

Why it mattersOpenAI is now using its own models on its own chips. The team puts a number on what the model changed, and names the parts it did not change.
OpenAI is designing chips with its own models, on its own chip. The company published a write-up of what GPT-6.2 Astra changed. The work covers Jalapeño silicon, AMD GPU circuits, and open RISC-V cores. Each result is OpenAI's own measurement.
Jalapeño is the custom inference chip OpenAI co-developed with Broadcom. The team says it reached tape-out in nine months. The first silicon came back this year. The model-assisted work sits inside that project.
What the model changed on Jalapeño
OpenAI says model-proposed changes cut standard-cell area in two places. The vector unit fell by about 13 percent. The matrix unit fell by about 10 percent. The vector unit runs parallel math. The matrix unit runs the arithmetic that matrix multiplications map to on this chip.
A single redesigned multiplier used 15 percent less cell area, and cut latency from two clock cycles to one. A redesigned adder used 15 percent less cell area and kept its two-cycle latency. OpenAI says these optimisations were silicon-proven during Jalapeño's first bring-up.
What the model changed on AMD
OpenAI also ran the same approach on circuits inside AMD GPUs. The company reports up to 14 percent gains in power, performance and area, across the cases it published. AMD has been working with OpenAI on hardware since the two signed a multi-year compute deal earlier in 2026.
RISC-V, from specification to RTL
The team used GPT-6.2 Astra to generate hardware code for two open RISC-V cores. The cores are PicoRV32 and Ibex. The input was a plain-language specification alone. OpenAI says the generated designs met the behaviour of the originals. It does not claim they are better. The exercise was a test of whether a model can follow a chip specification end to end.
What is still missing
Every number in the post is OpenAI's. No independent synthesis run has been published, and the vendor's own benchmark is marketing until a third party reproduces it. The company has a separate multi-year partnership with Synopsys, announced 30 September 2026, to train GPT-Synopsys on EDA tools. That work is how any outside team would try the same loop.
Chip design is a long-cycle job. Jalapeño took nine months from first design to tape-out, which is fast for custom silicon but slow for software. A 13 percent area win on one unit does not save a whole tape-out. It does change what one engineer can explore in one week.
Source
OpenAI: Advancing chip design with our models and the Synopsys press release on GPT-Synopsys.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

