Sign in

Blog › Hacker News

Image: bodkins18 on Pixabay
Hacker News · 4 min read

Frostbite: An AI-Designed CPU Built on Interaction Nets

Published Nov 21, 2026

Architect Labs says its AI systems discovered, designed and verified a new CPU architecture, claiming 4.59x better performance per watt than a dual-issue in-order core.

Architect Labs has published details of Frostbite, a general-purpose CPU with a new instruction set that the company says its AI systems discovered, designed and verified from compiler to final GDS layout. The headline claim is a 4.59x gain in performance per watt over a modern, silicon-proven dual-issue, in-order core, measured on 25 workloads, 15 of which were held out until the final evaluation. The numbers are pre-silicon and come from the vendor itself, so they deserve scrutiny. The approach behind them is still unusual enough to be worth a closer look.

Why interaction nets

Conventional compilers hide the dependencies between parts of a program behind shared, mutable state. Hardware therefore assumes the worst case and rediscovers at runtime which instructions really depend on each other, using out-of-order machinery, deep cache hierarchies and coherence protocols. All of that costs area and power. Meanwhile, single-core performance has been largely flat since the early 2000s, which is why vendors keep adding cores.

Interaction nets, a model of computation dating back to the early 1990s, take a different route. A program becomes a graph of nodes, and computation proceeds by rewriting pairs of connected nodes. Each rewrite is local, and each value has exactly one owner, so independent rewrites can run in any order, or all at once, with the same result. Languages such as Vine compile ordinary-looking, Rust-like code to this representation. In theory that gives parallelism without threads, locks or race conditions, and lets data sit in small memories next to where it is used instead of in large coherent caches.

The idea is not new on the software side. Runtimes built on interaction nets already exist, but dedicated silicon for them has been rare. Frostbite is a bet that the model maps cleanly onto hardware if the design space is explored aggressively enough.

How the AI pipeline worked

The process started with architecture exploration. Agents read research papers and benchmarks, built Python models to compare ideas, and produced a 141-page document describing the programming model, runtime and major chip blocks. That document was later discarded once the first RTL showed that many of its estimates were far off physical reality.

The more interesting piece is the specification layer. Architect Labs built an in-house language called ASL and wrote a system model, Arcmodel, in it. ASL supports machine-checkable translations between RTL, system models, UVM and SVA, and lowers to Lean for theorem proving. By the end of the project the specification expanded to more than 800,000 lines of Lean. The company argues that a formal specification works better than natural-language documents as a single source of truth when many agents work in parallel, because it limits hallucinated details and reward hacking.

Arcmodel initially diverged from synthesized RTL power numbers by as much as roughly 250%. An automated recalibration step traced power hotspots to counters such as SRAM accesses and flits sent, then fitted Python functions until model and RTL agreed within 5%. After that, the model ran about 660 times faster than RTL simulation, which made rapid iteration possible. Across four searches and a final ablation round, the agent swarm tried 58 ideas and kept 37. One example the company cites is fine-grained clock gating together with a matching change to the software runtime, done in hours.

An emergent work-stealing design

Because an interaction net's parallelism is discovered at runtime and spread across space, Frostbite has no central scheduler. According to the write-up, the agents converged on three rules per tile: run your own work first, steal from a neighbor with spare work if idle, and otherwise power down. A frame manager parks continuations in local memory when a loop or branch splits into pieces of unknown size, and a hardware work queue distributes the new tasks. The instruction set follows from these constraints, with wide bundles, hardware loop counters, streams, and operations to take, peek and copy values from other parts of a running net.

What to watch

Several caveats apply. The benchmark set is small, and some kernels such as FFT, radixsort and LU were scaled down to 64 times smaller than their SPLASH-2 originals. The baseline is a dual-issue in-order core, not a high-end out-of-order design. Power figures are pre-silicon, and Architect Labs itself notes they would need correlation against post-silicon measurements over multiple tapeout loops. Programs also have to be written in, or compiled to, an interaction-net language, which is a significant ecosystem hurdle compared with existing ISAs.

Still, the pairing of a formally specified model with agent swarms that revise architecture, RTL and runtime together is the part developers should follow. If independent verification or silicon results appear, Frostbite could be an early data point that AI-driven co-design finds architectures humans were unlikely to explore.

Original reporting: architectlabs.com

Discussion

Loading comments …

The Repocharts newsletter

Once a week: the biggest repo clashes, the fastest risers and the AI and tech stories that matter. Free, no spam, unsubscribe with one click.

We send a confirmation link first (double opt-in). See our privacy policy.