Architect Labs introduced Frostbite, a general-purpose CPU architecture discovered, designed, and verified end-to-end by AI agent swarms using interaction nets as its computational model. Programs written in the Vine language compile to Ivy, an intermediate representation encoding interaction net graphs where every rewrite is local, dependencies are explicit, and values have single ownership, eliminating shared mutable state. This enables fine-grained spatial parallelism without threads, locks, or out-of-order hardware; work-stealing emerges dynamically across a grid of tiles, each running ready interactions from local input ports, pulling from neighbors when idle, and sleeping otherwise. The design flow used a formal specification language, ASL, translated to Lean for theorem proving, maintaining a single provable source of truth across ISA, system model (Arcmodel), RTL, and verification. Arcmodel was autonomously calibrated against post-synthesis RTL to within 5% power accuracy using scalar functions over formal counters (SRAM accesses, flits, frame misses). Synthesized on 7 nm, a Frostbite quad (2×2 tiles + four L3 nodes) occupies 0.792 mm² at 500 MHz GDS (800 MHz synthesis) versus 0.895 mm² for a modern dual-issue in-order superscalar core (TCPU). Across 25 workloads (10 search-set, 15 held-out including Kalman filter, PCA, FEA, LU solve, LZ4), Frostbite achieves a geometric mean 4.59× perf/W (6.02× search, 3.83× held-out) and 8.77× perf/area. Energy breakdown shows TCPU spends 57% on memory and 27% on fetch/decode/issue; Frostbite reduces these to 38% and 6%, spending 26% on on-chip network/scheduler. Scaling to 4×4 tile arrays maintains power distribution without coherence overhead.
Read original
hackernews