Frostbite CPU: AI-Designed Chip Claims 4.59x Gain

Frostbite CPU architecture designed by AI using interaction nets

The Frostbite CPU is a general-purpose processor with a new instruction set that Architect Labs says its AI systems designed, verified and took to a final GDS layout with little human input. On 25 workloads, the company reports 4.59x better performance per watt than a modern, silicon-proven superscalar CPU.

These are the company’s own pre-silicon numbers from its blog post, not results from a manufactured chip. Fifteen of the 25 workloads were held out until the final evaluation, so the design wasn’t tuned to them.

The idea behind it

Frostbite is built around interaction nets. In this model a program is a graph of nodes, and computation rewrites connected pairs of nodes. Each rewrite touches only a small part of the graph, and pairs never overlap, so they can run in any order or all at once with the same result. Every value is used exactly once, so there is no shared mutable state to guard.

Conventional compilers hide dependencies between program parts, so CPU hardware rediscovers them at runtime with out-of-order machinery and deep cache hierarchies. Architect Labs argues that an interaction net keeps the dependencies explicit. On a five-queens search, it says about twelve interactions were ready at every step with no change to the program. Programs are written in Vine, a Rust-like language, and compiled to an intermediate form called Ivy.

The company expects two benefits: parallelism with no threads, locks or race conditions, and small local memories next to where data is used, which cuts the area and power spent on large caches.

Timeline of the Frostbite program showing key events, stages from architecture exploration through alignment and iteration, and the agent worktrees created in each stage.
Image: Architect Labs

How the AI worked

The system iterated over the compiler, instruction set, architecture, microarchitecture, RTL, verification and layout, using agent swarms. It started with zero architectural priors and produced a 141-page exploration document, later deleted once the first RTL existed because many estimates were off.

Architect Labs built a specification language, ASL, and a system model called Arcmodel. It says Arcmodel’s power and performance estimates are within 1-5%, and it runs about 200x faster than RTL simulation. The spec lowers to over 800,000 lines of Lean for proof checking. At first, Arcmodel diverged from the RTL by almost 250% in the worst case, so an automated pipeline recalibrated it to within 5%.

Across four searches and a final ablation round, agents tried 58 ideas and kept 37, including clock gating on a running tile and smaller L3 tables.

Table of 25 workloads in five groups, each split into search set and eval set. Parallel: gemm, chacha20 and spmv in the search set; conv2d, radixsort and pca in the eval set. Staged: fft and binomial in the search set; lcs, bfs, emeans, lu solve and fea in the eval set. Sequential: biquad and sha256 in the search set; huffman and kalman in the eval set. Branching: json in the search set; randomforest and lz4 in the eval set. Recursive: mergesort and nqueens in the search set; quicksort, kdtree and fib in the eval set.
Image: Architect Labs

Work stealing without a scheduler

Because parallelism is spread across tiles, there is no central scheduler. Each tile runs its own work first, takes work from a neighbor with spare tasks if idle, and powers down if none exists. A hardware work queue and a small frame manager track tasks that finish in any order and wake the next section when the last piece arrives.

A web of 31 clusters, one per block of the chip, each made of the signals, events, state and clauses its Lean contract names, linked by blue curves where blocks share signals; below, the frame manager's contract drawn out, with its signals on the left feeding events, state and 49 clauses on the right.
Image: Architect Labs

Why Architect Labs says this matters

The company says single-core CPU performance has largely stalled since the early 2000s, which pushed the industry toward large multicore chips such as Intel Xeon, AMD Zen and NVIDIA Vera. Those chips must keep the illusion that every core sees the same memory, and that takes deep cache hierarchies and coherence protocols that cost area and power.

It sees Frostbite as a proof of concept. In its view, every class of chip, from datacenter CPUs to AI accelerators and network switches, will go through similar architectural shifts as AI systems improve. That is a forecast, not a result.

The post follows an earlier one on Redwood, where the company says its AI built and deployed an inference accelerator in record time from a given architecture. Frostbite goes further because the AI also chose the architecture.

Architect Labs says the agents also found hardware features, such as work stealing between tiles and fine-grained clock gating, that would have been hard to specify up front.

What to watch

The claim is large, and independent checking will matter. The numbers are pre-silicon, and the company acknowledges power figures would need correlating to real silicon over tapeout loops. Architect Labs says it will publish more on ASL and the architecture. Until a Frostbite CPU is manufactured and tested by others, 4.59x is a design-stage result.

Source: Architect Labs

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.