Skip to main content

AMD and Intel, Step Aside: AI-Designed CPU Delivers Nearly 5x Performance Boost

AI startup Architect Labs claims that its experimental Frostbite CPU architecture delivers 4.59x better performance per watt and 8.77x better performance per unit area than traditional processor cores. Designed from scratch using AI, Frostbite employs a radically different computational approach, but its impressive results have yet to be validated on actual silicon.

AMD英特尔请让路 AI设计的全新CPU性能提升近5倍

US AI startup Architect Labs claims that its AI system designed a completely new general-purpose CPU architecture from scratch, including the instruction set, microarchitecture, RTL code, verification, and final GDS layout. The company claims that its new design delivers 4.59x better performance per watt and 8.77x better performance per unit area than traditional CPU cores across 25 workloads. However, these results come from pre-fabrication simulations, not actual physical processors.

Unlike traditional CPUs, the Frostbite architecture is based on a mathematical model called interaction nets, first proposed in 1990. Instead of relying on a single large core, it uses a grid of small compute units that execute programs represented as interaction nets: graphs where computation is achieved through local rewriting of connected nodes. Each compute unit first processes its own tasks and receives additional work from neighboring units when idle. If there is no work to do, it can enter an inactive state to save power. This distributed approach eliminates the need for a centralized scheduler, aiming to make the architecture more scalable.

Architect Labs states that 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 completed through local interactions between connected nodes. Programs are written in Vine, a language developed by the company with syntax similar to Rust, and compiled into an intermediate representation called Ivy.

The company stated that its AI system explored the architecture with "zero architectural prior knowledge," iterating through the compiler, instruction set, architecture, microarchitecture, RTL, verification, and finally the GDS layout. The process employed formal, machine-verifiable specifications and multi-agent clusters, considering 58 proposals across four rounds of search and a final round of testing. Architect Labs said that 37 proposals were ultimately retained.

AMD英特尔请让路 AI设计的全新CPU性能提升近5倍

The 7nm layout of Frostbite uses a 4×4 grid, highlighting a 2×2 Frostbite Quad and an L3 node

For performance evaluation, Architect Labs synthesized Frostbite cores using a 7nm process. The design achieved a frequency of 800MHz during synthesis, with a final physical layout frequency of 500MHz. A Frostbite Quad consists of four compute units arranged in a 2×2 grid along with four L3 cache nodes, occupying an area of 0.792 mm². For comparison, the reference CPU core (referred to by the company as TCPU) uses the same process technology and occupies 0.895 mm². Due to legal reasons, Architect Labs did not disclose the identity of the competitor processor.

Across 25 workload tests, the geometric mean of Frostbite's performance per watt was 4.59 times that of the reference CPU core. In the 10 workloads used during optimization, the advantage reached 6.02x; in the 15 workloads held back for final evaluation, the advantage was 3.83x. The company also reported a performance-per-area advantage of up to 8.77x.

AMD英特尔请让路 AI设计的全新CPU性能提升近5倍

How does Frostbite's performance per watt compare to traditional CPUs across 25 workloads? The dashed lines represent the geometric means for the search set, holdout set, and all workloads, respectively.

Architect Labs attributes the performance gains to reduced memory traffic and instruction overhead. Test results varied significantly depending on the workload. Frostbite excelled particularly in highly parallel tasks, and the company noted that it also outperformed the reference core in branch-heavy and dynamic workloads. The company plans to expand its testing scope to include industry-standard SPEC and PARSEC benchmarks and compare against larger, wider, out-of-order CPU designs.

AMD英特尔请让路 AI设计的全新CPU性能提升近5倍

Despite the encouraging results, the Frostbite architecture remains experimental rather than a mass-producible processor. Architect Labs has not announced any plans for mass production or commercial release of Frostbite chips. The company intends to release the RTL and instruction set architecture (ISA), build an FPGA-based software testing platform, develop a preliminary operating system, and use industry-standard benchmarks like SPEC and PARSEC to compare the performance of larger-scale Frostbite configurations against multi-core CPUs.