Vendors & EquipmentAmdOpenaiSynopsysAI Chip Design
OpenAI's models cut chip circuit area by up to 14 percent for AMD
OpenAI says its models delivered up to 14 percent gains on AMD's flagship GPU IP and shrank Jalapeño's vector unit 13 percent, with AMD already manufacturing AI-modified designs.
Vendors & EquipmentWhy it matters
- OpenAI reports up to 14 percent improvements on AMD's flagship GPU IP relative to heavily optimized baselines.
- GPT-o3 mini cut standard-cell area in Jalapeño's vector unit by 13 percent and matrix unit by 10 percent in early 2025.
- AMD has included AI-written changes in a chip design now in manufacturing and plans more in future production files.
- All five tested GPT models scored higher designing RISC-V cores from written requirements than improving existing designs.
- OpenAI will partner with Synopsys to train GPT-Synopsys to mimic chip engineers' EDA workflows.
The story
OpenAI says its language models have delivered up to 14 percent improvements on AMD's flagship GPU IP, on top of measurable gains on its own in-house inference chip, Jalapeño. The figures come from a company blog post, so treat them as vendor-marketing numbers until a silicon partner confirms them independently. AMD, notably, has already moved some of the AI-written changes into a design now in manufacturing.
The disclosure pulls back a curtain on a race that has so far been mostly rumour: can large language models actually do chip engineering work, or just write convincing emails about it? OpenAI's answer, with numbers attached, is the former — at least on specific, well-scoped blocks.
What did GPT-o3 mini do for Jalapeño?
OpenAI used GPT-o3 mini in early 2025 to help design Jalapeño, its custom inference chip. The model produced what the company calls non-trivial PPA improvements — power, performance and area, the three currencies of chip design.
The specifics:
- Standard-cell area in the vector unit dropped 13 percent.
- The matrix unit shrank by 10 percent.
An area saving of 13 percent on one unit is not decoration. On a modern accelerator, area is money, yield and power budget all at once. The catch, as ever, is that OpenAI graded its own homework here.
What happened with AMD?
The AMD collaboration produced the headline numbers. OpenAI reported gains of up to 14 percent relative to heavily optimized baselines on AMD's flagship GPU IP — emphasis on "heavily optimized," because squeezing single digits out of a mature design is normally a slow, human-priced activity.
AMD also saw single-digit PPA percentage gains across key blocks of its upcoming CPU using designs written by an OpenAI model. An unspecified model went further, rewriting SystemVerilog RTL — the code that describes a chip's circuits — and found additional improvements. OpenAI claims this equals a year's worth of progress achieved in less than five months. That claim deserves an asterisk the size of a wafer, but the direction of travel matters.
The commercial reality is concrete: AMD has reportedly included these changes in a recent chip design now in the manufacturing phase, and plans to fold further changes into production design files for future versions. Chips entering fab is the point where vendor claims stop being claims.
OpenAI said: "Looking ahead, we expect further PPA improvements on top of GPU and CPU relative to the most optimized baseline."
Why the RISC-V benchmark matters
Separately, OpenAI tested five models — GPT-5.5, GPT-5.5 Sol, GPT-6.1 Sol, GPT-6 Astra and GPT-6.2 Astra — on two open-source 32-bit RISC-V cores. Each model received an original hardware description for each core and a task: increase throughput. Then came a second task — design a core from written requirements and templates.
The result cuts against intuition. All five models scored higher designing a core from written requirements than when trying to improve an existing design. In other words, the models are better architects than renovators. Greenfield work, where the specification is clean, suits them more than surgical edits to legacy RTL.
What does the Synopsys partnership change?
The most strategically interesting item sits at the end of the post. OpenAI will partner with Synopsys to train a new model, GPT-Synopsys, to use Synopsys electronic design automation tools the way chip engineers use them — mimicking the actual workflow rather than just generating code fragments.
Synopsys dominates EDA software, so a model fluent in its toolchain would sit at the choke point of nearly every chip design on the planet. If GPT-Synopsys works, it changes what a design team of ten can attempt. If it doesn't, expect the announcement to age quietly.
For now, the evidence is a blog post with unusually specific numbers and one partner that has put the output into silicon. Watch AMD's production design files — and any benchmark OpenAI publishes without grading itself — for the next signal.
Also reported
Source: datacenterdynamics.com
More from Amara Osei
Show full bio
News editor covering technology and product news in Carrier business, 5G and 6G, fiber and broadband at Telecom Gazette.
583 articles