AI-designed CPU architecture, Digimart Global: US AI startup Architect Labs says its AI systems designed a new general-purpose CPU architecture from scratch, covering the instruction set, microarchitecture, RTL, verification and final GDS layout. The company claims the design, called Frostbite, delivers 4.59 times the performance per watt and 8.77 times the performance per unit of area of a conventional CPU core across 25 workloads.
The results are based on pre-silicon simulations rather than a physical processor. Architect Labs has not announced a manufactured Frostbite chip or any commercial availability, so the performance claims have not been tested in silicon.
How the architecture differs
Unlike conventional CPUs, Frostbite is built around interaction nets, a mathematical model first introduced in 1990. Instead of a single large core, it uses a grid of small computing tiles. Each tile processes its own tasks, can take on work from neighboring tiles when idle, and can enter an inactive state to save power.
According to the company, this distributed approach removes the need for a centralized scheduler and is intended to make the design easier to scale.
Architect Labs also says the approach reduces the need for shared mutable memory, complex cache hierarchies and cache-coherence mechanisms, because each value has a single owner. Programs are written in the company’s Vine language, which has syntax similar to Rust, and are compiled into an intermediate representation called Ivy.
Testing and reported results
The company says its AI systems explored the design with zero architectural priors, using formal, machine-checkable specifications and multiple agent swarms. Architect Labs reports that 58 proposed ideas were considered across four searches and a final round of tests, and that 37 were retained.
For evaluation, Architect Labs synthesized Frostbite on a 7 nm process, reaching 800 MHz during synthesis and 500 MHz in its final physical layout. One Frostbite Quad, with four computing tiles in a 2×2 grid paired with four L3 nodes, occupies 0.792 mm².
The comparison core, which the company calls TCPU, occupies 0.895 mm² on the same process. Architect Labs has not disclosed the identity of the competing processor, citing legal reasons.
Across 25 workloads, the company reports a geometric mean of 4.59 times the performance per watt of the reference core. The advantage was 6.02 times across the 10 workloads used during optimization and 3.83 times across 15 additional workloads held back until final evaluation.
Results varied considerably by workload, with the strongest gains in highly parallel tasks. Architect Labs says it also outperformed the reference core in branching and dynamic workloads.
Next steps
Architect Labs plans to release the RTL and instruction set architecture, build an FPGA-based software playground, develop a preliminary operating system, and expand testing to industry-standard SPEC and PARSEC benchmarks. It also plans comparisons against larger, wider and out-of-order CPU designs.
Until those tests are published and a physical chip is built, the performance figures remain the company’s own simulation-based claims.














