Architect Labs, a startup based in Palo Alto, California, has announced the creation of a new central processing unit designed entirely by artificial intelligence. The company claims its AI agents discovered, designed, and verified the chip, which it calls Frostbite, from scratch without relying on existing hardware blueprints.
According to a research blog post published by the company, Frostbite features a new instruction set built from the ground up. In simulated tests on a 7nm design that has not yet been manufactured, a four-tile cluster of the chip delivered 4.59 times the performance per watt of a similarly sized conventional dual-issue, in-order CPU core. The startup noted that it cannot name the specific comparison core due to legal restrictions.
Frostbite Processor Architecture and Performance
Conventional CPUs typically use hardware to determine which instructions depend on each other at runtime. This approach requires deep cache hierarchies and complex coherence protocols to keep memory views consistent across multiple cores, which consumes significant area and power. Architect Labs describes Frostbite as offering “multithreading for free” by having the program itself expose parallel work, eliminating the need for threads, locks, or out-of-order hardware discovery.
The chip compiles programs into interaction nets, a computational model from the 1990s where a program is represented as a graph of nodes. Computation occurs by rewriting pairs of connected nodes. Because these pairs never overlap, they can execute in any order or simultaneously while producing the same result. Each value is used exactly once, allowing data to remain in small local memories rather than large shared caches.

To illustrate the efficiency, the team cited a chess-like puzzle involving five queens on a 5×5 board. A standard program searches for solutions sequentially. As an interaction net, the same search breaks into 3,183 small operations with a longest dependency chain of only 262 steps. This allows an average of 12 operations to run simultaneously at each step without rewriting the program.
Each tile in the Frostbite design contains its own small local memory. Four tiles plus four L3 nodes form the cluster benchmarked against the conventional core. The full layout consists of 16 tiles. Early synthesis reached 800 MHz, but the final layout settled at 500 MHz. The cluster also averaged 8.77 times the performance per area of the comparison core.
Architect Labs AI Design Process
The AI system that built Frostbite started with “zero architectural priors.” The team first selected the workloads the chip needed to handle well, then allowed the agents to read papers and benchmarks to build Python models. The agents selected one concept for focus, producing an original 141-page PDF containing a programming model and runtime. This document was deleted after the first register-transfer level design was generated because many estimates were physically unrealistic.

Once the architecture was selected, the agents switched from English to a proprietary language called ASL. ASL translates to Lean, a proof checker, which allows the design to be machine-checked. This process stops a large class of bugs and prevents agents from gaming their objectives. From this specification, the agents devised Arcmodel, a software model of the entire chip written in ASL before any circuit design existed.
Simulating the register-transfer level design is slow, so the agents tested ideas on the Arcmodel, which ran about 200 times faster. Initial power estimates were off by up to 250 percent, but an automated loop adjusted the model until it matched the circuit design within 5 percent. Swarms of agents then proposed changes in parallel, checking each against the spec and the fast model. Out of 58 ideas tried, 37 were kept. The process also revealed an unspecifed behavior called work stealing, where an idle tile pulls spare work from a neighbor to save power.
Architect Labs previously showcased Redwood, an AI inference accelerator for robots and drones, in August. That chip was generated in under two weeks from a human-written spec but still runs only on an FPGA. The company plans to release Frostbite’s RTL and instruction set eventually, alongside a software playground and a preliminary operating system. Further benchmarks against larger multicore CPUs are expected in the coming months.
Source: Tom’s Hardware

