Frostbite is a general-purpose CPU with a new instruction set and architecture built around interaction nets, a model in which computation is represented as local graph rewrites. The design’s central argument is that conventional CPUs spend hardware resources rediscovering dependencies and maintaining shared-memory consistency, while interaction nets make dependencies explicit, give each value a single owner, and expose independent work for parallel execution. This could reduce reliance on out-of-order execution, cache hierarchies, and coherence machinery. The team reports 4.59 times the performance per watt of a silicon-proven dual-issue, in-order CPU core across 25 workloads. Fifteen workloads were held back from design tuning; the tests span parallel, staged, sequential, branching, and recursive computation, with practical additions including Kalman filtering, PCA, finite-element analysis, and LZ4 decompression.
AI systems explored the architecture and iterated across the instruction set, compiler, runtime, microarchitecture, RTL, verification, and physical-design output. A key part of the process was ASL, a machine-checkable specification language used as a shared contract for the software model, hardware, and verification. The resulting specification lowered to more than 800,000 lines of Lean, and a rapid system model enabled faster design exploration than RTL simulation. Because its early power and performance estimates diverged sharply from synthesized hardware, an automated calibration process matched the model to RTL measurements within 5%. The account presents Frostbite as evidence that AI-guided software-hardware co-design can discover architectures beyond conventional CPU approaches, while reporting pre-silicon results rather than deployed-chip performance.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.