OpenAI: AI improved AMD GPU design by up to 14% in power, speed and area
OpenAI says its AI improved the design of AMD's flagship GPUs by up to 14% in power consumption, speed and die area. Think of it as a successful refactor: the chip does exactly the same thing, but uses less power or takes up less space. The model rewrote the circuit descriptions, and Synopsys software checked every edit. The takeaway: AI is useful where a machine verifies the result.
AI-processed from TNW; edited by Hamidun News
OpenAI says its models improved the design of AMD's flagship graphics processors by up to 14% across three metrics that chip engineers are always trading off against each other: power consumption, performance and die area (together known as PPA). Put simply, it resembles a successful code refactor: the chip does exactly the same thing, but uses less power or takes up less space on silicon, which means more chips fit on a single wafer. "Up to 14%" is the best result, not the average.
For developers and anyone building products on AI, something else matters more: the model did not edit a demo example but carefully optimized designs, and some of its edits are already in a chip that has been sent to production.
What exactly OpenAI did
The model worked with the code that describes a chip's circuits. For the model, this is much like source code is for a programmer: it proposes rewriting a section so the circuit comes out more compact or more efficient. Synopsys tools, which engineers use even without AI, then check that the chip still behaves the same way after the edit. If the check fails, the edit is discarded.
OpenAI estimates this produced roughly a year of progress in less than five months. Beyond the GPUs, AMD confirmed, according to OpenAI, a single-digit gain (that is, within a few percent) in key blocks of its next flagship processor. That is more modest than the gains on the graphics chips, but still notable for mature designs.
OpenAI's own Jalapeño chip and a model comparison
OpenAI applied the same approach to its own chip, Jalapeño, which went from design to production in nine months. The models' edits shrank two key parts of the chip by about 13% and 10%. According to the company, these changes held up in finished silicon, not only in simulation.
OpenAI also tested five generations of its models on two small open-source processors. The newest, GPT-6.2 Astra, delivered the best result both when improving an existing design and when building a processor from written requirements. The company itself adds a caveat: the second test is easier, because the models could skip implementing functions that were present in the originals.
What comes next and where the limits are
The next step is GPT-Synopsys, a model OpenAI is training together with Synopsys. It is meant to run inside the engineering tools chip designers already use, rather than in a separate environment.
Startups are drawn to the idea too: in June, Architect Labs raised $24 million to help companies with similar tasks.
Not everyone is fully convinced. Richard Ho, head of OpenAI's chip effort, said last month that he is not yet ready to take AI's changes on faith, which is why each one is verified. Also, the figures in this story come from OpenAI itself and are relayed from the company's statements: the article contains no independent measurements.
What this means in practice
The first lesson: AI works best where there is a fast, reliable automatic checker. In chips, that is Synopsys's verification tools; in ordinary software development, it is tests and CI. If a machine can check "correctness," autonomous edits from a model pay off in practice; if it cannot, a human pays the price of errors.
The second lesson: the gain is measured in schedule. Five months instead of a year, and nine months from design to production, matter to any company building its own hardware for AI workloads. If GPT-Synopsys really does fit into familiar tools, the barrier to entry for small teams will drop.
For now, only one thing can be said with confidence: AI speeds up engineers' work but does not replace them. Responsibility for every edit stays with people and with the verification software.
- Nov 20, 2026GitLab makes Duo Agent Platform credits last up to 2x longer
- Nov 20, 2026Superhuman opens Go to 20 third-party agents as the assistant passes 1 million users
- Nov 20, 2026AWS launches managed NVIDIA Dynamo for models that don't fit on a single server
- Nov 20, 2026Kimi Code CLI now works with Amazon Bedrock without a separate API key
Key Questions Answered
What exactly OpenAI did?
The model worked with the code that describes a chip's circuits. For the model, this is much like source code is for a programmer: it proposes rewriting a section so the circuit comes out more compact or more efficient. Synopsys tools, which engineers use even without AI, then check that the chip still behaves the same way after the edit. If the check fails, the edit is discarded. OpenAI estimates this produced roughly a year of progress in less than five months.
What are the key details on OpenAI's own Jalapeño chip and a model comparison?
OpenAI applied the same approach to its own chip, Jalapeño, which went from design to production in nine months. The models' edits shrank two key parts of the chip by about 13% and 10%. According to the company, these changes held up in finished silicon, not only in simulation. OpenAI also tested five generations of its models on two small open-source processors. The newest, GPT-6.2 Astra, delivered the best result both when improving an existing design and when building a processor from written requirements.
What comes next and where the limits are?
The next step is GPT-Synopsys, a model OpenAI is training together with Synopsys. It is meant to run inside the engineering tools chip designers already use, rather than in a separate environment. Startups are drawn to the idea too: in June, Architect Labs raised $24 million to help companies with similar tasks. Not everyone is fully convinced. Richard Ho, head of OpenAI's chip effort, said last month that he is not yet ready to take AI's changes on faith, which is why each one is verified.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.