AI NewsInfrastructureAnnouncement

Architect Labs taped out Frostbite, a CPU AI swarms designed

Architect Labs says AI designed a new CPU that runs 4.59 times more work per watt than a silicon-proven baseline.

AI News

Editorial2 min read

LinkedInX
Die-shot layout of the Frostbite CPU, a tile-based design routed on a 7 nanometre process

Why it mattersChip design is one of the slowest engineering jobs. If AI can reach tape-out in weeks, the question for every hardware team changes.

A small team says it pointed AI at a blank sheet and got a chip. Architect Labs is a frontier AI research company in Silicon Valley. It published a post on 18 November about Frostbite, a new CPU. AI systems wrote the instruction set, the microarchitecture, the register-transfer code and the final layout. Humans set the task and reviewed the output. The post hit Hacker News the same week and sits at 651 points.

Frostbite is built on interaction nets. Architect Labs sums up the model in one line. A program is a graph of nodes, the company says. Each step rewrites a pair of nodes connected through their ports. That shape fits a tile-based chip. Many small processors steal work from each other. The common design inside most phones and laptops is one big out-of-order engine instead.

The numbers Architect Labs reports

Architect Labs measured Frostbite against what it calls the TCPU. The company calls the TCPU "a modern, silicon-proven superscalar (dual-issue, in-order) CPU core". It will not say which chip the TCPU is, for legal reasons.

The test used 25 workloads. 10 drove the search. 15 were held back. Frostbite reached 4.59 times the performance per watt on average, and 8.77 times the performance per area. The chip reached final layout on a 7 nanometre process. Synthesis ran at 800 megahertz. The final mask ran at 500 megahertz. The power figures come from pre-silicon models. No working chip is in a working board yet.

What the AI did, and what the humans did

Architect Labs says AI agents tried 58 ideas across four searches. 37 of them stayed in the design. The agents wrote a formal specification. That specification lowered to more than 800,000 lines of Lean, the theorem prover. Humans set the Lean framework and wrote the proofs each design had to pass. The write-up names every layer the AI touched. Compiler. Instruction set. Architecture. Microarchitecture. Register-transfer code. Verification. Final GDS.

The caveats, as the post states them: this is research, not a product. The baseline is unnamed. The power is pre-silicon. A working chip in a working board is still ahead.

The useful part for a team building software, where the output depends on a model, is the method. Architect Labs let AI rewrite its own target through a formal spec. A theorem prover then caught what review could not. That is the same loop evals aim at in application code. It just ran here on a job that used to take hundreds of engineers and years of work.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project