OpenAI’s AI Improved AMD’s Flagship GPU Design by Up to 14% (2026): What Changed, What Shipped and What’s Unproven

Jump to comments

Close-up of a bare AMD silicon die, its grid of circuit blocks glowing green, blue and gold against a teal background
The bare input/output die of an AMD Epyc 7702 server processor, photographed up close.
Photo: Fritzchens Fritz, 2019, via Wikimedia Commons (CC0)

OpenAI says its AI models found up to a 14% improvement in power, performance and area on the design of AMD’s flagship graphics chip, a design AMD’s own engineers had already spent years tuning. The company published the results on Wednesday, Nov. 18, 2026, in a post titled "Advancing chip design with OpenAI models," and AMD confirmed that some of the model-written changes are already in a chip design it recently sent for manufacturing.

The short answer: OpenAI’s models rewrote parts of the code that describes AMD’s GPU circuits, and Synopsys design software checked that every rewrite still behaved exactly like the original. OpenAI calls the result roughly a year of chip-design progress in under five months. The same approach cut the cell area of two compute blocks in OpenAI’s own Jalapeño inference chip by about 13% and 10%, and those gains held up on real silicon. The AMD figures are company-reported, "up to" numbers, and nobody has yet measured them on a finished AMD chip.

What OpenAI and AMD announced

OpenAI posted the news on X at 9:05 a.m. Pacific time on Nov. 18 and laid out four separate results in its blog post. They carry very different levels of proof, so it helps to take them one at a time.

AMD’s flagship GPU: up to 14% improvement in PPA, the industry’s shorthand for power, performance and area, measured against what OpenAI called "heavily optimized baselines" on AMD’s flagship GPU IP. The model got there by rewriting SystemVerilog RTL, the code that describes a chip’s circuits. AMD has already included some of the changes in a recent design it sent for manufacturing and added others to its production design files for future versions.

AMD’s next CPU: AMD verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship processor, using designs written by OpenAI’s model.

OpenAI’s Jalapeño chip: model-proposed changes reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit. OpenAI says these optimizations were proven in silicon during the chip’s bring-up, the first power-on and testing of manufactured parts.

Two open-source processor cores: OpenAI tested five of its models on improving and designing two small RISC-V cores, and its newest model, GPT-6.2 Astra, scored highest on both.

Neither company named the AMD product that carries the changes. OpenAI only calls it "flagship GPU IP," and AMD’s posts did not add a product name or a ship date.

What PPA, RTL and tape-out mean

Chip announcements are full of jargon. These are the four terms that matter for this story.

PPA: power, performance and area. Designers trade these against each other: a faster circuit often draws more power or takes up more silicon. Less area means more chips per silicon wafer and lower cost, and less power means less heat and electricity in a data center.

RTL: register-transfer level code, written in hardware languages such as SystemVerilog. It describes what a chip’s circuits do, cycle by cycle, before software tools turn it into millions of logic gates and then a physical layout.

Verification: proving that a changed circuit still does exactly what the old one did. OpenAI says its model used Synopsys tools to check that each rewrite "preserved the block’s behavior" and to measure the trade-offs.

Tape-out: the moment a finished design is handed off for manufacturing. OpenAI president Greg Brockman wrote on X that some of the AMD results were "included in a design that’s recently been taped out."

This matters because RTL work is where chip designers spend years shaving percentages off designs that are already good. An AI model that keeps finding gains there, and whose changes survive formal checks, is doing work that used to belong only to senior engineers.

How the model found gains a human team had missed

OpenAI’s post is short on method, but it says the model reasoned about RTL "at the level of an expert microarchitect" and could rewrite circuits that "perform calculations, control operations, or move data."

The Jalapeño section gives the most concrete detail. OpenAI says it started using one of its earlier models, o3-mini, on the chip in early 2025. In the vector unit, which manages and executes parallel computing tasks, savings came from floating-point adders and multipliers, conversion functions, reduction logic and the special-function unit. The models tried different logic implementations and different ways of splitting work across pipeline stages, which let them share hardware and store fewer intermediate results.

In the matrix unit, one multiplier redesign used about 15% less cell area in synthesis and cut its latency from two clock cycles to one. A separate adder redesign cut cell area by about 15% while keeping a two-cycle latency.

OpenAI’s hardware lead, Richard Ho, described the same work more candidly in a September interview with More Than Moore. Some of the models’ source-code optimizations worked before the team could explain them.

We didn’t quite understand conceptually what it was doing. We said we would probably get there, we’d need to spend some time to think about it. But we were running fast, so we weren’t able to spend the time to understand. (Richard Ho, OpenAI, More Than Moore interview, Sept. 30, 2026)

That is why the verification step is the real story. Ho said in the same interview that he still does not fully trust the models, and that standard design tools still sign off every chip. "If the model is 99.99% correct, you can’t tape that out, it’s not accurate enough," he said.

How AMD and OpenAI reacted on Nov. 18

The announcement turned into a small parade of executive posts on X within three hours.

OpenAI, 9:05 a.m. PT: "Working with @AMD, our models identified up to 14% PPA improvement on flagship GPU IP in less than five months—building on designs AMD’s engineers had already spent years optimizing." The post passed 950,000 views by early Thursday.

Synopsys, 9:40 a.m. PT: the design-software maker called the work "compelling results" from frontier intelligence "combined with best-in-class EDA."

Greg Brockman, 11:05 a.m. PT: "roughly an additional year of progress on @AMD’s chip achieved in less than 5 months using our models."

AMD, 11:11 a.m. PT: the company quoted Brockman at The New York Times DealBook event: "Working with AMD, we’re able to get a year’s worth of full generational improvements in five months, just using our models."

Lisa Su, 11:48 a.m. PT: AMD’s chair and chief executive wrote: "Proud of these fantastic results – @OpenAI + @AMD engineers pushing the frontiers of chip design together!"

AMD chief executive Lisa Su, in a gray jacket and glasses, speaking at a podium with Frontier and Oak Ridge National Laboratory banners behind her
AMD chief executive Lisa Su speaking at Oak Ridge National Laboratory in Tennessee in 2019.
Photo: Genevieve Martin, OLCF at ORNL, 2019, via Wikimedia Commons (CC BY 2.0)

AMD shares rose 5.3% in regular trading on Wednesday and edged up another 0.3% overnight, Stocktwits reported. The same report noted a second boost for AMD’s mood that day: Rockstar told Wccftech that GTA VI, which launched this week, uses AMD’s FSR 4 upscaling on the Xbox Series S. Our earlier piece covers the GTA 6 reviews and release times.

The two companies were already tied together. In October 2025, AMD and OpenAI signed a 6-gigawatt agreement for OpenAI to deploy multiple generations of AMD Instinct data-center GPUs, starting with 1 gigawatt of MI450 chips in the second half of 2026. AMD also issued OpenAI a warrant for up to 160 million AMD shares that vests as deployments and share-price targets are hit. Better AMD chips are, in a direct sense, better chips for OpenAI.

The RISC-V test: how much better each model generation got

To track progress from one model to the next, OpenAI used two open-source 32-bit RISC-V processor cores: PicoRV32, a compact core built for reprogrammable FPGA chips, and Ibex, the embedded processor in the OpenTitan security chip.

The first task asked each model to make the existing core faster. Designs scored zero if they failed a simulation check or used more than 16 times the original core’s area. For designs that passed, OpenAI multiplied instructions per clock cycle, measured on workloads the models never saw, by the maximum clock speed estimated with the open-source Yosys tool.

The result: at a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores. On PicoRV32, its score at that budget was about twice that of GPT-6 Astra. Our guide to GPT-6.2 Astra and Sol covers what else changed in that model.

The second task gave each model only written requirements and asked it to design the core from scratch. All five models scored higher this way on both cores. OpenAI’s evaluation covered 9,881 scored runs in total.

OpenAI flagged its own caveat: designing from a specification let models leave out features the original cores have, such as performance counters and extra instructions, which may explain part of the higher scores. Timing after physical synthesis was not checked either, so these are not production-grade results.

AI has helped design chips before. What is different now

AI in chip design is not new. Google’s AlphaChip, first described in a 2020 preprint and later in the journal Nature, has generated chip layouts used in every generation of Google’s TPU AI accelerator since 2020, according to a 2024 Google DeepMind post. Synopsys also sells its own AI design tools under the Synopsys.ai name.

The difference is where in the process the AI works. AlphaChip plays a placement game: it decides where blocks sit on the silicon. OpenAI’s models rewrite the RTL itself, the logic that engineers write by hand, and then rely on standard tools to prove the rewrite is correct. That is closer to what a human microarchitect does all day.

An AI model that keeps finding gains there, and whose changes survive formal checks, is doing work that used to belong only to senior engineers.

The next step is a product. In September, OpenAI and Synopsys announced GPT-Synopsys, a specialized model trained to operate Synopsys design tools the way expert engineers do. OpenAI licenses the tools to train it, the two companies share revenue when customers use it, and it will run on OpenAI-hosted infrastructure. Synopsys says customer design data will not be used to train the model and that early engagements with leading chip companies are underway. Neither company has announced pricing or a general release date.

What we know

The results are real enough for AMD to ship them: AMD put some model-written changes in a design sent for manufacturing and others into production design files, and verified CPU gains itself.

Jalapeño’s gains are proven in silicon: the 13% and 10% cell-area reductions held up during bring-up of OpenAI’s manufactured chip, which OpenAI co-developed with Broadcom and plans to deploy in its data centers by the end of 2026.

Every change is checked: Synopsys tools verified that each rewrite kept the block’s behavior, and OpenAI’s hardware lead says standard design flows still sign off its chips.

What is still unclear

What the 14% is made of: OpenAI gave a maximum, not an average, and did not split it between power, performance and area. A 14% area cut in one block is a very different result from a 14% speedup across a whole chip.

The baseline: "heavily optimized" is OpenAI’s description. The design stage, process and measurement conditions behind the comparison are not public.

Which chip, and when: no AMD product has been named. There is no finished-silicon measurement of the AMD changes yet, and no public date for when a chip carrying them reaches customers.

How much was the model, and how much the engineers: OpenAI describes the work as a partnership with AMD’s team and does not say how many proposals the model made, how many failed verification, or how much human steering it needed.

Bottom line

OpenAI’s Nov. 18 results are the strongest public evidence so far that a general AI model can improve production chip designs at the RTL level, not just in lab tests. The Jalapeño gains are proven in silicon, AMD is shipping some of the GPU changes, and AMD itself verified the CPU gains. The headline 14% is still a company-reported best case with no published breakdown. The next proof point is a finished AMD chip, and how quickly GPT-Synopsys reaches other chipmakers.

Further reading

AMD Ryzen 7 9800X3D desktop processor box
AMD Ryzen 7 9800X3D Desktop Processor

AMD’s 8-core gaming chip with stacked 3D V-Cache, one of the most popular desktop processors for building a gaming PC.

AMD Ryzen 7 9700X desktop processor box
AMD Ryzen 7 9700X Desktop Processor

An 8-core, 16-thread Zen 5 processor for everyday work, coding and gaming on AMD’s AM5 platform.

Logitech MX Keys S low-profile wireless keyboard in graphite
Logitech MX Keys S Wireless Keyboard

A quiet, low-profile wireless keyboard that switches between up to three computers, a favorite of engineers who write code all day.

Cover of Chip War by Chris Miller
Chip War: The Fight for the World’s Most Critical Technology

Chris Miller’s history of the semiconductor industry, from the first transistors to the companies that design and build today’s most advanced chips.

Cover of Code, second edition, by Charles Petzold
Code: The Hidden Language of Computer Hardware and Software (2nd Edition)

Charles Petzold builds a working computer from simple switches and logic gates, the clearest way to understand what RTL code actually describes.

Aglena Best Of product picks

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top
Aglena