# Startup says AI designed Frostbite, a CPU with an entirely new architecture

> **TL;DR:** Architect Labs says its AI independently discovered, designed, and verified the Frostbite CPU, which uses a completely new instruction set A four-tile cluster reaches 4.59 times the performance per watt of a reference in-order core, or 3.83 times when counting only the 15 workloads that were not tuned by the AI The figures come from a simulation of an unfabricated 7 nm design, and the reference core is not named There is no comparison against out-of-order CPUs yet, and the company promises SPEC and PARSEC results within the next few months

- Original: https://migovi.com/en/frostbite-cpu-ai-thiet-ke-kien-truc/
- Author: [Sky Shepard](https://migovi.com/en/author/batman11689/)
- Published: 2026-11-22T21:02:25+07:00
- Categories: AI, Semiconductors, News
- Source: [Tom's Hardware](https://www.tomshardware.com/tech-industry/semiconductors/ai-agents-design-an-entirely-new-cpu-architecture-from-scratch-startup-claims-frostbite-posts-4-59x-the-performance-per-watt-of-an-in-order-core-in-pre-silicon-tests)

---

Frostbite delivers 4.59 times the performance per watt of an in-order core, but the result comes from a 7 nm simulation of a design that has never been built as a real chip.

## What Frostbite is and how it differs from a conventional CPU

Architect Labs, a startup based in Palo Alto, says its AI system independently discovered, designed, and verified Frostbite, a general-purpose CPU built on an instruction set created from the ground up. Across 25 workloads, a four-tile Frostbite cluster delivered 4.59 times the performance per watt of a dual-issue, in-order CPU core of comparable size that has already been proven in silicon. That number comes with several caveats: it is based on a 7 nm design that has only been simulated, the reference core is not named for legal reasons, and when only the 15 workloads without AI tuning are counted, the advantage drops to 3.83 times.

According to Architect Labs, traditional CPUs need hardware that "rediscovers" at runtime which instructions depend on which. Multicore designs also require multi-level cache systems and coherence protocols so that every core sees the same shared memory, an approach that is hard to scale and costs both die area and power. The company's approach is what it calls "multithreading for free," in which the program itself exposes the work that can run in parallel, so there is no need for threads, locks, or out-of-order hardware to hunt for that work.

## Interaction nets and how the chip runs in parallel

Frostbite programs are compiled into interaction nets, a computation model dating to 1990 in which a program is a graph of nodes and computation proceeds by rewriting pairs of connected nodes. Because the pairs being rewritten never overlap, they produce the same result whether they run in any order or all at once. Each value is also used exactly once, so data can live in small local memories instead of a large shared cache. The team gives the example of placing five queens on a 5 x 5 chessboard so that none attacks another: run as an interaction net, the search splits into 3,183 small operations, and the longest chain that must run sequentially is only 262 steps. That works out to roughly 12 operations ready to run at the same time at each step, with no changes to the program.

On the hardware side, each tile is a small, identical processing unit, and the tiles are connected by an on-chip network and have their own local memory. The cluster used for benchmarking consists of four tiles plus four L3 nodes, while the full chip layout has 16 tiles. The design hit 800 MHz in the first synthesis pass but settled at 500 MHz in the final layout. Normalized by area, the cluster delivers an average of 8.77 times the performance of the reference core.

![Diagram of the Frostbite chip showing a 4 x 4 grid of tiles with one group of four tiles and an L3 node](https://migovi.com/wp-content/uploads/2026/11/frostbite-cpu-ai-thiet-ke-kien-truc-1.jpg)

*Diagram of the Frostbite chip, with one four-tile cluster and an L3 node outlined (Image: Architect Labs)*

## How the AI designed the chip

According to Architect Labs, the AI started from a state of "zero architectural priors," meaning no built-in architectural assumptions, although the team did choose in advance the workloads the chip had to handle well. The agents read research papers and benchmarks, built models in Python, and then picked one idea to pursue. The first 141-page PDF, covering the programming model, runtime, and high-level blocks, was deleted once the first register-transfer level (RTL) design existed, because many of the figures and estimates in it were far off from physical limits.

After settling on an architecture, the agents dropped English because, in the team's view, it cannot express every detail precisely or catch every error. They then wrote the specification in the company's own language, called ASL. ASL is translated into Lean, a proof assistant, so a machine can check the specification. The company argues this blocks a large class of errors and also limits the agents' ability to game their assigned goals. From there, the agents built Arcmodel, a software model of the entire chip written in ASL, before any circuit design existed. The model runs about 200 times faster than RTL simulation. Its initial power estimate was off by nearly 250%, but an automated loop tuned it until it matched the circuit design within 5%.

Next, multiple groups of agents proposed changes in parallel. Each change was checked against the specification, tested on the fast model, put into the circuit design, and measured for power before being kept or discarded. In total, 58 ideas were tried and 37 were kept. The process also produced a behavior that the company says no human specified, called work stealing: a tile that runs out of work takes surplus work from a neighboring tile, and if there is nothing to take, it goes idle to save power.

## What Architect Labs plans next

Frostbite is not Architect Labs' first design. In August, the company introduced Redwood, an AI inference accelerator for robots and drones. For Redwood, the company says AI produced the RTL, verification, firmware, drivers, and kernels in under two weeks from a specification written by two human architects. Redwood still runs only on an FPGA rather than as a real chip, though the company says it is looking into manufacturing at TSMC. In a separate effort, Tendrils Compute, a company based in Cambridge, UK, recently raised funding for a processor it says will have thousands of cores running in parallel. It is also developing Vine, the language used to write programs for Frostbite.

Architect Labs says it will publish more benchmarks, but that is only part of the plan. The company wants to build a software playground running on FPGAs so anyone can try running programs on Frostbite, along with a rudimentary operating system, and eventually to release the chip's RTL and instruction set. In the near term, it plans to expand the set of workloads and compare against more types of CPUs, including a larger Frostbite mesh configuration set against multicore versions of the reference core with large caches. It also plans to test against larger, wider out-of-order CPUs, with results on popular benchmark suites such as SPEC and PARSEC expected in the next few months.

**MIGOVI VIEWPOINT**

Migovi finds the idea notable because Frostbite follows an entirely different computation model rather than further tweaking an old architecture, but the available numbers are not enough to conclude anything about real-world performance. The chip has only been simulated, the reference core is an unnamed in-order design, and the advantage fell from 4.59 times to 3.83 times once the AI-tuned workloads were excluded, which suggests part of the edge comes from the choice of test material. A more meaningful comparison would be against out-of-order CPUs and multicore designs with large caches, which the company has promised to publish on SPEC and PARSEC. For users in Vietnam, this is not something to buy or expect soon, so the sensible move is to wait for those results and the open RTL before putting faith in the 4.59 times figure.

Source: [Tom's Hardware](https://www.tomshardware.com/tech-industry/semiconductors/ai-agents-design-an-entirely-new-cpu-architecture-from-scratch-startup-claims-frostbite-posts-4-59x-the-performance-per-watt-of-an-in-order-core-in-pre-silicon-tests)