The company described Frostbite in a research blog post as a general-purpose CPU with a new instruction set “built from the ground up.” Architect Labs said a four-tile Frostbite cluster in a simulated 7nm design delivered 4.59 times the performance per watt and averaged 8.77 times the performance per area of a similarly sized “silicon-proven” dual-issue, in-order CPU core. Architect Labs said it cannot name the comparison core for legal reasons.
Conventional CPUs have hardware that “rediscovers, at runtime, which instructions actually depend on each other.” Architect Labs calls its alternative “multithreading for free,” where the program itself exposes parallel work without threads, locks, or out-of-order hardware required for that discovery.
Programs for Frostbite are compiled to interaction nets, a 1990 model of computation in which a program is a graph of nodes and computation happens by rewriting pairs of connected nodes. Because pairs of nodes being rewritten never overlap, they can run in any order or simultaneously and arrive at the same result. Every value is also used exactly once, so data can stay in small local memories instead of large shared caches.
Each tile in Frostbite has its own small local memory, and four tiles plus four L3 nodes form the cluster that was benchmarked, though the full design has 16 tiles. The cluster reached 500 MHz in the final layout, down from 800 MHz in early synthesis.
The AI agents that designed Frostbite started with what the company calls “zero architectural priors,” though the team first selected the workloads the design had to perform well on. The agents read papers and benchmarks, built Python models, and then selected one concept for focus. A 141-page PDF output containing a programming model, runtime, 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.”
Once the agents picked an architecture, they moved away from English and wrote the specification in the company’s own language, ASL, which translates to the Lean proof checker so it can be machine-checked. The agents then built a software model of the whole chip called Arcmodel, which ran about 200 times faster than RTL simulation. After an automated loop adjusted the model’s power estimates until they matched the circuit design within 5%, swarms of agents proposed changes in parallel, with each change checked against the spec, tested in the fast model, built into the circuit design, and measured for power. That process resulted in 58 ideas tried, with 37 kept.
A behavior emerged that “was never specified by any human,” known as work stealing, where a tile with nothing to do can pull spare work from a neighbor or go idle to save power.
This is not Architect Labs’ first design; the company showed off Redwood, an AI inference accelerator for robots and drones, in August. That design has not been made in silicon, though the company says it is working toward having it built at TSMC. Tendrils Compute, a Cambridge, U.K.-based possible rival, raised funding earlier this month for a processor it says will have “thousands of cores operating in parallel” and is also developing Vine, the language Frostbite’s programs are written in.
Architect Labs says more benchmarks are coming, including comparisons against a broader set of CPUs on popular benchmark suites such as SPEC and PARSEC. It also wants an FPGA-hosted “software playground for anybody to run a program on Frostbite” along with a “preliminary operating system” and eventually plans to release Frostbite’s RTL and instruction set.
