Architect Labs Claims AI Agents Designed New CPU Architecture from Scratch

Palo Alto startup Architect Labs announced that its AI systems designed, verified, and discovered Frostbite, a general-purpose CPU featuring a new instruction set "built from the ground up."

According to a research blog post from the company, the Frostbite design utilizes a model of computation known as interaction nets. In this paradigm, programs are represented as graphs of nodes where computation occurs by rewriting connected pairs. Because these node pairs never overlap, they can be executed in any order or simultaneously. This approach allows data to remain in small local memories rather than large shared caches, as every value is used exactly once.

The company reported that a four-tile Frostbite cluster, based on a simulated 7nm design, delivered 4.59 times the performance per watt of a similarly sized, silicon-proven dual-issue, in-order CPU core across 25 workloads. The cluster also averaged 8.77 times the performance per area of the comparison core. Architect Labs stated it cannot name the specific comparison core due to legal reasons.

To develop the architecture, the AI agents began with "zero architectural priors," reading papers and benchmarks to build Python models. The team noted that an initial 141-page PDF output containing a programming model and high-level blocks was deleted after the first register-transfer level (RTL) design was generated because "many of the numbers and estimates were well off what was physically possible."

Following the initial design phase, the agents transitioned from English to a proprietary language called ASL to write the specification. ASL translates to Lean, a proof checker that allows for machine-checking to prevent bugs and stop agents from gaming their objectives. This process allowed the agents to create Arcmodel, a software model of the chip, which ran approximately 200 times faster than simulating the RTL.

The final Frostbite layout consists of 16 tiles, with each tile containing its own small local memory. While the design reached 800 MHz during early synthesis, it reached 500 MHz in the final layout. The development process involved swarms of agents proposing changes in parallel, resulting in 58 ideas being tested, 37 of which were kept.

FI Journal
Privacy Overview

This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.