Abstract:
Artificial intelligence startup Architect Labs claims that its experimental Frostbite CPU architecture has 4.59 times higher performance per watt and 8.77 times higher performance per unit area than traditional processor cores. Designed from scratch using artificial intelligence, Frostbite takes a radically different approach to computing, but its impressive results have yet to be verified on an actual chip.

Architect Labs, an American artificial intelligence startup, claims that its artificial intelligence system has designed a new general-purpose CPU architecture from scratch, including instruction sets, microarchitecture, RTL code, verification and final GDS layout. The company claims that under 25 workloads, its new design's performance per watt is 4.59 times that of traditional CPU cores, and its performance per unit area is 8.77 times that of traditional CPU cores. However, these results come from pre-chip fabrication simulations, not actual physical processors.
Unlike traditional CPUs, the Frostbite architecture is based on a mathematical model called interactionnets, which was first proposed in 1990. Rather than relying on a single large core, it uses a grid of small computing units that execute programs represented in the form of an interaction network: an interaction network is a graph in which computation is performed by local rewriting of connected nodes. Each computing unit processes its own tasks first and receives additional work from neighboring computing units when idle. If there is no work to do, it can go into an inactive state to save power. This distributed approach eliminates the need for a centralized scheduler and is designed to make the architecture more scalable.
Architect Labs says this approach also reduces the need for shared mutable memory, complex cache hierarchies and cache coherence mechanisms. Each value has only one owner, and computation is done through local interactions between connected nodes. Programs are written in the Vine language developed by the company, with a syntax similar to Rust, and compiled into an intermediate representation called Ivy.
The company said its AI system explored the architecture with "zero prior knowledge of the architecture," iterating across the compiler, instruction set, architecture, microarchitecture, RTL, verification and final GDS layout. The process used formal, machine-verifiable specifications and multi-agent clusters, with a total of 58 proposals considered over four rounds of search and a final round of testing. Architect Labs said 37 proposals were ultimately retained.

Frostbite's 7nm layout features a 4×4 grid, highlighting a 2×2 Frostbite quad and an L3 node
For performance evaluation, Architect Labs synthesized the Frostbite core using a 7nm process. The design reached a frequency of 800MHz during the synthesis process and the final physical layout was 500MHz. A Frostbite Quad consists of four computing units arranged in a 2×2 grid and four L3 cache nodes, with an area of 0.792 square millimeters. The comparison CPU core (which the company calls TCPU) uses the same process technology and has an area of 0.895 square millimeters. For legal reasons, Architect Labs is not disclosing the identity of the rival processor.
Across 25 workload tests, the Frostbite architecture achieved a geometric mean performance per watt of 4.59 times that of the reference CPU core. The advantage reached 6.02x across the 10 workload tests used during the optimization process and 3.83x across the 15 workload tests retained before final evaluation. The company also reported a performance advantage of up to 8.77 times per unit area.

How does the performance per watt of the Frostbite architecture compare to traditional CPUs under 25 workloads? The dashed lines represent the search set, the hold-out set, and the geometric mean of all workloads.
Architect Labs said the performance improvements come from reduced memory traffic and instruction overhead. Test results vary significantly depending on the workload. Frostbite performs particularly well on highly parallel tasks, and Architect Labs also noted that it also outperforms the reference core on branching and dynamic workloads. The company plans to expand the testing to include industry-standard SPEC and PARSEC benchmarks and compare them to larger, wider, out-of-order CPU designs.

While the results are encouraging, the Frostbite architecture is still in the experimental stage and is not a mass-produced processor. Architect Labs has not announced plans for mass production or launch of the Frostbite chip. The company plans to release RTL and instruction set architecture (ISA), build an FPGA-based software test platform, develop a preliminary operating system, and use industry-standard benchmarking tools such as SPEC and PARSEC to compare the performance of larger Frostbite configurations with multi-core CPUs.
Comments