AI·News & analysis
OpenAI says its AI found up to 14% gains in AMD's flagship GPU design
OpenAI says its models improved AMD's flagship graphics chip design by up to 14% and that AMD has already sent some of the changes to manufacturing. It also shared results from its own Jalapeño chip.

Tide
Ripple
Sci-fi
6/10
Reality
Shipping
AI that edits the blueprints of the chips it runs on.How we rate
OpenAI says its AI models improved AMD's flagship GPU design by up to 14% in power, performance and area, against designs that were already heavily optimized.
AMD has already put some of the changes into a design sent for manufacturing. OpenAI calls it a year of progress in under five months. The figures are self-reported, and OpenAI is next training GPT-Synopsys with Synopsys, with no launch date.
What to know
- OpenAI says its models improved power, performance and area by up to 14% on AMD's flagship GPU design, measured against designs that were already heavily optimized.
- AMD has put some of the changes into a recent design it sent for manufacturing and added others to its production design files for future versions.
- On AMD's next flagship CPU, AMD verified single-digit percentage gains across key blocks.
- On OpenAI's own Jalapeño chip, early silicon tests support area cuts of about 13% in the vector unit and 10% in the matrix unit.
OpenAI says its AI models found up to 14% in improvements on AMD's flagship graphics chip design, and AMD has already sent some of those changes to a factory. That is the news. Now the fun part: what does it mean when an AI edits the blueprint of the chip it will one day run on?
What did OpenAI actually say?
OpenAI shared the results in a blog post on Wednesday. It worked with AMD on its flagship GPU, the graphics chip that powers AI data centers. The company says the gain was up to 14% in power, performance and area.
Those three words are a bundle. Chip designers call it PPA, short for power, performance and area. Power is how much electricity the circuit burns. Performance is how fast it runs. Area is how much silicon it takes up.
The three pull against each other. A smaller circuit can use less power but miss its speed target. A faster one can need more energy or more room.
OpenAI says the gain was measured against designs AMD had already heavily optimized. That makes the baseline hard to beat. Human experts had already squeezed those circuits.
How did the AI change the chip?
The models rewrote the code that describes the chip's circuits. That code is written in SystemVerilog RTL. RTL, short for register-transfer level, is a text description of the chip's digital logic: what each block does and how data moves between blocks.
Think of it as the recipe, not the cake. Change the recipe, and the factory eventually bakes a different chip.
Every rewrite had to prove it still did the same job. OpenAI says its model used Synopsys software to check that each change kept the block's behavior. The tools then measured the power, speed and area trade-offs.
Rewrites that broke the behavior were thrown out. Only the ones that passed were scored.
In real life It is like a copy editor who shortens every sentence in a book, then runs a checker to confirm the plot did not change.
Is any of this in a real chip?
Yes, at least in part. OpenAI says AMD has included some of these changes in a recent design it sent for manufacturing. AMD also added others to its production design files, so future versions can use them.
The AMD CPU side is smaller but confirmed. AMD verified single-digit percentage PPA gains across key blocks of its next flagship CPU, using edits the models proposed. OpenAI did not give a single headline number for the processor.
Why it matters: A design sent for manufacturing is a commitment. It means AMD's engineers trusted the changes enough to put real money behind them.
What about OpenAI's own chip?
OpenAI used the same approach on Jalapeño, its first custom inference chip, built with Broadcom. An inference chip is one made to run a finished AI model and produce answers, rather than to train the model.
OpenAI started using its models on Jalapeño in early 2025. The chip went from its first circuit code to tape-out in about nine months. Tape-out is the moment a finished design is handed off for manufacturing.
Early silicon tests support estimated area cuts for two parts of the chip:
- Vector unit: about 13% less standard-cell area, the space taken by the basic building blocks of the circuit.
- Matrix unit: about 10% less area.
One example is a multiplier redesign that used about 15% less cell area and cut its delay from two clock cycles to one. A separate adder redesign also cut area by about 15% and kept its two-cycle delay.
OpenAI says the finished chips reproduced the gains that the models found on paper.
Where does Jalapeño stand today?
Jalapeño is not just a drawing. Engineering samples, the first test chips, were already running real workloads as far back as June 2026.
OpenAI expects small first deployments around the end of 2026. It has pinned a broader rollout to 2027.
What's next: The chip's AI-assisted circuits are meant to start doing real work around the end of this year, with the larger rollout after that.
How good are the models at this?
OpenAI also built a test. It asked five generations of its models to improve two small open-source processors, called PicoRV32 and Ibex. Both use RISC-V, an open design standard for processors that anyone can use for free.
The rules were strict:
- Budget: about 112,000 tokens per run. A token is a small chunk of text the model reads or writes.
- Failure: a design scored zero if it failed a check or used more than 16 times the original area.
- Score: designs that passed were measured on instructions per clock cycle, using workloads hidden from the models.
GPT-6.2 Astra, OpenAI's newest model, scored highest on both processors. GPT-6.1 Sol also beat GPT-5.6 Sol on both.
There is a catch. All five models did better when they designed from written requirements than when they improved existing code. OpenAI notes the models could leave out features that the originals had, which makes the easier test easier.
What is GPT-Synopsys?
OpenAI is training a model called GPT-Synopsys together with Synopsys. It is meant to drive Synopsys' chip design tools the way expert chip designers do. Those tools are known as EDA, for electronic design automation.
The companies have a multi-year agreement covering joint development, go-to-market work and revenue sharing. We covered the start of that deal in our earlier story on GPT-Synopsys. No launch date has been given.
The idea is that the method OpenAI used on AMD and Jalapeño could become a product that other chipmakers buy. Other startups chase the same idea. Architect Labs raised $24 million in June to help companies design custom chips with AI.
What should we keep in mind?
The numbers are self-reported. They come from OpenAI's own post, and OpenAI does not say exactly how it calculated the 14% or how the gain splits across power, performance and area. The 14% is a block-level measure, not a whole-GPU speedup.
There is also a business link. OpenAI holds warrants for up to 160 million AMD shares. That does not make the results wrong, but it is worth knowing.
AMD's own actions help here. It confirmed the CPU gains and used some of the GPU changes in a real design.
And OpenAI's own chip chief, Richard Ho, said last month that he still does not fully trust AI design changes, so every one is checked. That is the right instinct when the output ends up in silicon.
What it means for you
- Faster chip cycles: If AI can compress a year of design work into under five months, new chips could arrive sooner, though this result covers selected blocks only.
- No instant speedup: The 14% is a circuit-design gain, not a promise that your next graphics card is 14% faster.
- Checks stay central: The method works because every rewrite is verified, so the human review step is not going away.
- Watch GPT-Synopsys: If it launches, the same approach could reach chipmakers beyond AMD and OpenAI.
The bottom line
OpenAI says its models improved AMD's flagship GPU design by up to 14%, and AMD has already used some of the changes. The numbers come from OpenAI itself, so the full picture will take more evidence. Still, a design sent to the factory with AI-written edits inside is a real milestone for AI in chip design.
Key facts
- Headline gain
- Up to 14% PPA on AMD's flagship GPU
- Time
- About a year of progress in under 5 months
- AMD CPU
- Single-digit PPA gains on key blocks
- Jalapeño
- About 13% and 10% area cuts
- Best model
- GPT-6.2 Astra
Got questions?
Quick answers, plain wordsWhat did OpenAI announce?
OpenAI published results showing that its AI models improved the design of AMD's flagship GPU by up to 14% in power, performance and area. The models rewrote the code that describes the chip's circuits. OpenAI says this was roughly a year of progress achieved in less than five months.
What does PPA mean?
PPA stands for power, performance and area. It is the set of three measures chip designers trade off against each other: how much electricity a circuit uses, how fast it runs, and how much silicon it takes up.
Is the AMD GPU now 14% faster?
No. The 14% figure is a gain in the power, performance and area trade-off on selected circuit blocks. It is not a claim that the whole graphics chip runs 14% faster, and the split between power, speed and area is not broken out.
Are these changes in a real chip?
Partly. AMD has included some of the changes in a recent design it sent for manufacturing and added others to its production design files for use in future versions. On OpenAI's own Jalapeño chip, the gains showed up in finished silicon.
What happened with AMD's CPU?
AMD verified single-digit percentage PPA gains across key blocks of its next flagship CPU, using edits proposed by OpenAI's models. OpenAI has not published a single headline number for the processor.
What is Jalapeño?
Jalapeño is OpenAI's first custom inference chip, built with Broadcom. An inference chip is one made to run a finished AI model and produce answers, rather than to train it. It went from first circuit code to the handoff for manufacturing in about nine months.
What is GPT-6.2 Astra?
GPT-6.2 Astra is OpenAI's newest model in the tests it published. It scored highest on both small open-source processor designs OpenAI used to compare five generations of its models.
What is GPT-Synopsys?
GPT-Synopsys is a model OpenAI is training with Synopsys to operate Synopsys' chip design tools the way expert chip designers do. The two companies have a multi-year agreement covering joint development, go-to-market work and revenue sharing. No launch date has been given.
Can we trust the numbers?
The figures come from OpenAI's own blog post, so they are self-reported. OpenAI also holds warrants for up to 160 million AMD shares. AMD did confirm the single-digit CPU gains and has used some of the GPU changes, which gives the claims real weight.
Does this replace chip engineers?
Nothing in the results says so. Each change was checked with Synopsys verification tools, and OpenAI's own chip chief Richard Ho said last month that he still does not fully trust AI design changes, so every one is checked.
SourcesOpenAI
Topics and tagsOpenAI, AI chips, Industrial robots, openai
Go deeper
- ExplainerWhat is a semiconductor chip?A semiconductor is a material, usually silicon, that can either conduct or block electricity. Chips pack billions of...
- ExplainerWhat is a GPU, and why does AI need them?A GPU (graphics processing unit) is a chip that does huge numbers of simple calculations at the same time. Built to...
Related stories

OpenAI built an AI that can actually run the software chip engineers use
OpenAI and Synopsys announced GPT-Synopsys, an AI model that can operate Synopsys' chip design software directly, reasoning through design and verification work engineers currently do by hand.

OpenAI says its agents now finish 46% of 2-to-4-hour research tasks alone
OpenAI's new report says coding agents now complete 46% of its researchers' two-to-four-hour tasks without help, up from 23% in March, while daily agent use hit $2,400 per researcher.

ChatGPT will retire Advanced and Standard voice on December 11
OpenAI will remove Advanced and Standard voice from ChatGPT on December 11, 2026, and move everyone to GPT-Live, the voice model it launched in July.
More in brief
- ASELSAN's EJDERHA electromagnetic weapon disables 51 drones with one commandNov 21
- Microsoft To Do will shut down on June 30, 2027Nov 21
- Medtronic's Hugo surgical robot cleared for general surgery in the USNov 21
- Google's Gemini now turns noisy fire radio into text for Watch DutyNov 21
- Internet Archive staff move to unionize with CWANov 20
- Etsy sellers are leaving over new terms that allow AI trainingNov 20