Subscribe
AI Models

OpenAI: its models redesigned production chips — 14% better PPA on AMD's flagship GPU

Friday, November 20, 2026 · 5 min read

OpenAI said on November 18 that its frontier models now produce expert-level improvements to production silicon: model-rewritten chip designs delivered up to 14% better power, performance and area on AMD's flagship GPU, and cut about 13% of the area from the vector unit of its own Jalapeño inference chip.

Bullish$AMD$SNPS

📊 Scorecard: our Bullish call on $AMD is scored from $489.19, the last close before it ran. Its first result comes after the next close.

OpenAI: its models redesigned production chips — 14% better PPA on AMD's flagship GPU

The News

OpenAI published "Advancing chip design with OpenAI models" on November 18, reporting that its frontier models now produce expert-level improvements to production silicon. Model-rewritten SystemVerilog RTL delivered up to 14% better power, performance and area (PPA) on AMD's flagship GPU IP versus heavily optimised baselines — OpenAI calls it "roughly a year of progress achieved in less than 5 months" — and AMD has already folded some changes into a design sent for manufacturing, with the rest in production design files.

AMD also verified single-digit PPA gains on key blocks of its next-generation flagship CPU. On OpenAI's own Jalapeño inference chip, model-proposed changes cut standard-cell area about 13% in the vector unit and 10% in the matrix unit, silicon-proven at bring-up, with one multiplier redesign using about 15% less area and cutting latency from two clock cycles to one.

OpenAI is also training GPT-Synopsys with Synopsys to drive EDA tools the way expert engineers do, under a multi-year preferred-partner deal announced September 30. OpenAI's own caveat: models designing from written requirements can drop features present in the originals, such as performance counters, and timing was not checked after physical synthesis. All figures are OpenAI-reported — AMD has published no figures of its own.

A year of chip-design progress in under five months is the kind of claim that, if it holds, changes who designs silicon — and how fast.

Why It Matters

The result matters because chip design has been one of the hardest, slowest and most expert-dependent jobs in technology. A single flagship GPU takes years and hundreds of engineers to design, and squeezing out a few percent of extra efficiency normally takes months of painstaking work. OpenAI says its models did roughly a year's worth of that progress in under five months — and on a design that was already heavily optimised, not an easy starting point.

It matters for AMD directly. The company verified the gains itself, folded some changes into a design already sent for manufacturing, and reported single-digit PPA gains on key blocks of its next-generation flagship CPU as well. That is the difference between a research result and a production result: AMD is shipping, or about to ship, silicon that an AI model helped design. If the manufactured chips perform as the models predicted, every chipmaker will have to ask whether its own design teams are keeping up.

It matters for the chip-design software business too, which is why Synopsys is in the story. EDA tools are the software engineers use to design and verify chips, and OpenAI is training GPT-Synopsys to drive those tools the way expert engineers do. If AI models become the ones doing the designing, the tools they drive — and the companies that make them — sit at the centre of the shift. The risk is that the value moves to the model makers instead, which is why the Synopsys read-through is mixed rather than straightforwardly good news.

Analyst's Take

Our call is Bullish. The detail that matters most is not the 14% figure but what AMD did with it: some of the model-proposed changes are already in a design sent for manufacturing, and the rest are in production design files. That is a chipmaker trusting AI-written circuits enough to put them in real silicon, not a demo on a whiteboard. Single-digit PPA gains on key blocks of AMD's next-generation flagship CPU add a second data point, and the Jalapeño results — 13% off the vector unit, 10% off the matrix unit, silicon-proven at bring-up — show the same approach working on OpenAI's own hardware.

The read-through for Synopsys is more complicated, which is why our call there is mixed. GPT-Synopsys, trained with Synopsys to drive its EDA tools, could make Synopsys's software more valuable if AI-assisted design becomes standard — every chip team would want the tools the models drive best. But if AI models start doing more of the design work themselves, the value could shift toward the model providers and away from the tool vendors. The multi-year preferred-partner deal announced September 30 suggests Synopsys sees the partnership as the safer side of that bet.

The caveats are real and OpenAI states them itself. Models designing from written requirements can drop features present in the originals, such as performance counters, and timing was not checked after physical synthesis — the step where a design's speed is confirmed on the actual manufactured layout. All figures are OpenAI-reported, and AMD has published no figures of its own. What to watch next is whether other chipmakers report similar gains, and whether the changes AMD sent for manufacturing come back working as designed. Reader poll: Is AI-designed silicon the next edge for $AMD? Vote: Bullish on $AMD, Wait for proof, Bearish on $AMD.

Are you bullish or bearish on this story?

By The Numbers

14%Better PPA on AMD's flagship GPU from model-rewritten designs
13%Area cut in the vector unit of OpenAI's Jalapeño chip
10%Area cut in the matrix unit of OpenAI's Jalapeño chip
15%Less area in one redesigned multiplier, with latency cut from two cycles to one

FAQ

What is PPA?

PPA stands for power, performance and area — the three things chip designers trade off against each other. A chip that uses less power, runs faster and takes up less space on the silicon is better on all three. A 14% PPA improvement means the design got better across that combined measure.

What is RTL?

RTL, or register-transfer level, is the detailed description of a chip's digital circuits, written in a hardware language such as SystemVerilog. It is the blueprint engineers write before a chip is manufactured. OpenAI's models rewrote that blueprint to make the circuits more efficient.

What is an EDA tool?

EDA stands for electronic design automation — the software engineers use to design, check and lay out chips. Synopsys is the biggest EDA company, and its tools verify that a chip design actually works before it is sent to a factory. OpenAI is training a model called GPT-Synopsys to drive those tools the way expert engineers do.

👁

Source: OpenAI / i10x

Get the daily briefing. AI news with a markets lens — tickers, bullish/bearish reads, zero hype. Subscribe free on Substack

More on $AMD

All $AMD coverage →

More from AI Agent Update