Skip to content
Week 47News, and what it means for you

AI models

AI designs chips: better blocks for AMD and a whole new processor

On one day, OpenAI reported that its models improved parts of AMD’s flagship chips by up to 14%, and the start-up Architect Labs unveiled a processor designed end to end by AI agents. It works because software checks every step.

On 18 November, AI designing chips made the news twice. OpenAI reported that its models had improved parts of AMD’s flagship chips by up to 14%, and the start-up Architect Labs unveiled a processor that AI agents designed from start to finish.

OpenAI had its models rewrite RTL: the SystemVerilog code that describes what each part of a chip does. These were already heavily optimised blocks from AMD’s GPU designs. Synopsys software checked that every new version still did exactly the same, and only then did the gain in power, performance and area count. The best blocks improved by up to 14%. AMD put some of the changes into a design that has already gone to manufacturing, and verified single-digit gains itself on its next flagship processor.

The models also made proposals for Jalapeño, OpenAI’s first chip of its own. According to OpenAI, the first tests on real chips back up estimates that this made the logic in two compute units about 13% and 10% smaller.

Architect Labs, which raised $24 million in June, went further. Swarms of AI agents designed Frostbite, a processor with a new instruction set: the basic commands a processor understands. They worked on that instruction set, the compiler and the circuits at once, down to the final layout at 7 nanometres. Of the 58 ideas they tried, they kept 37. The company reports on average 4.59 times as much work per watt as an existing processor core, and 3.83 times on programs kept out of sight during the design.

Both claims come from the makers themselves. OpenAI gives no figure for a whole chip and has a financial stake in AMD: it may buy up to 160 million of its shares. For legal reasons, Architect Labs cannot name the core it compares against, and Frostbite has not been made yet: its figures are calculated.

Roughly a year of progress in under five months.

OpenAI on its work on AMD’s GPU designs, 18 November 2026, in short

From code to chip

A chip starts as code. In RTL, designers describe, clock tick by clock tick, what each block does: which numbers it adds, stores or passes on. Software turns that code into logic gates, often hundreds of millions of them, and then into a layout. At tape-out the layout goes to the factory, and a fault that only shows up in the actual chip costs another round there. According to the Wilson Research Group’s 2022 verification study, chip designers spend 49% of their time on verification, and processor projects can have five verification engineers for every designer.

How you know a block still does the same

A rewritten block must do exactly what the old one did, only more efficient, faster or smaller. The usual check is testing: run the same tests before and after the change and compare the results. That only covers what the tests touch. Formal equivalence checking goes further: software proves that both versions behave the same in every situation. In 2014, Intel engineers described how this let them change designs faster and more safely.

The reports we read do not say exactly which checks AMD and OpenAI used. Architect Labs captured Frostbite in a single specification, translated into more than 800,000 lines of Lean, a language in which a computer checks proofs, and checked the agents’ work against it.

AI in chip design is not new

Chip designers have used AI for years, but for other jobs. Google DeepMind’s AlphaChip decides where the components go on a chip, in hours rather than weeks or months of human work, and has been used for the last three generations of Google’s own AI chips. What is new is that language models now write the chip’s code itself.

The question is alive in the Netherlands too. At the opening of the ChipNL Competence Center in Nijmegen last autumn, a panel including NXP, Axelera AI and Eindhoven University of Technology discussed not whether AI will change chip design but how deeply, Bits&Chips reported.

Our analysis

AI broke through in chip design because the field already checks everything

It is no accident that AI gets this far here. In chip design, checking already takes almost half the work, because a fault after tape-out is so costly. A model can propose many variants of a block, and the software says which ones behave the same and which are better. Richard Ho, OpenAI’s head of hardware, said this autumn that he still does not fully trust the models’ changes: with hundreds of millions of gates, being right 99.99% of the time is not enough. So every change goes through the full check, which is exactly what makes the model’s work usable.

The figures deserve nuance: 14% is the best result on individual blocks, not the gain for a whole chip, and Frostbite is compared with a core the company may not name. Those are open questions, not reasons to wave it away. The strongest evidence stands apart: AMD put changes into a chip on its way to the factory, and a chipmaker does not do that lightly.

The lesson reaches beyond chips. AI writes software quickly too, and there too it only becomes usable when something tireless checks it: tests that pin down what the software must do. Our expectation: organisations with good tests will get the most out of AI, because they can judge every proposal straight away.

What stands out

  • Improvements nobody could explain On Jalapeño, the engineers did not always understand why a change from the models helped. Because everything was checked, they could still go ahead.
  • Behaviour no human came up with In Frostbite, a compute block with nothing to do takes over work from a busy neighbour, or goes to sleep to save power. No human ever specified that, says Architect Labs.
  • The design document went in the bin The first plan, 141 pages long, was deleted because many of its numbers were physically impossible. From then on, the checkable specification was the chip’s only description.

What it means for you

AI pays off most in work whose results you can check automatically.

  • Pin down what good looks like first Tests and rules that describe what software must do are the foundation. Then AI may propose, and the check decides.
  • Measure the gain on your own work Someone else’s best result is a pointer, not a promise. Look at what AI delivers on your own software and processes.
  • Your colleagues keep the final say At OpenAI, too, the engineers remain the final arbiter, says Richard Ho. AI lets them try out more in the same time.

Where we come in · Development

We develop with AI, and tests decide

Flaire develops custom software, such as ERP, apps and integrations, within a few weeks and for a fixed price up front, faster thanks to AI. Whatever AI proposes only goes through once the tests that pin down what the software must do have passed. You get software that is tested, secured and documented, and the source code is yours.

How do you know whether software written by AI really does what it should?

Tell us which software you would like developed or improved. We’ll think it through with you and always reply within one working day.

More from week 47

All news

Search our services, the news and the rest of the site.