Back home

Frostbite

x+web 0.90
Visit source

Intro

Frostbite is the general-purpose CPU and new instruction set Architect Labs introduced on 18 November 2026. The official research post Introducing Frostbite: A new CPU architecture discovered, designed, and verified end-to-end by AI (labeled Research, November 18, 2026) says: “Today, we're introducing Frostbite, a general-purpose CPU with a new instruction set built from the ground up.” The company site is architectlabs.com.

The post frames this as a follow-on to the earlier Redwood work, in which its AI systems implemented, verified, and deployed an inference accelerator in about two weeks. Here the question is whether those systems, given more freedom to co-design software and hardware, can discover and build novel hardware that substantially outperforms existing human-made designs. Frostbite is built around interaction nets: the post summarizes that every step is local, every dependency is explicit, and every value has exactly one owner. Programs are written in Vine, a language with Rust-like syntax, and compiled to an intermediate representation called Ivy.

This entry records the new CPU name Frostbite. Do not confuse it with Electronic Arts’ Frostbite game engine, or with Architect Labs’ earlier public accelerator project Redwood. The post also names an in-house formal specification language, ASL, and a system model, Arcmodel; the authors say they will describe ASL in a later post, and this slot does not file those as standalone product names.

Advertisement

Explain

The post says the design started with zero architectural priors and was carried out by multiple agent swarms through architecture exploration, formal contracts and system modeling, RTL and power alignment, and then joint iteration. Exploration grouped programs by how they expose work—many independent pieces, rounds of parallel work, long sequential chains, data-dependent branches, and recursion—and built a search set plus a held-out eval set. Five more practical workloads were later added to the held-out set: a Kalman filter, PCA, finite element analysis, an LU solve, and LZ4 decompression. The system then wrote a full software model of the chip, called Arcmodel, in an in-house specification language named ASL. Officially, ASL supports machine-checkable translations and refinements among RTL, system models, UVM, SVA, and more, with power and performance estimates accurate to within 1–5%, roughly 200× faster than RTL simulation, and a native translation to Lean for theorem proving; by the end, Frostbite’s ASL is said to have lowered to 800k+ lines of Lean. Arcmodel is not physically aware at first, and its estimates can diverge from timing-converged, synthesis-clean RTL by almost 250% in the worst case, so an autonomous pipeline recalibrates counters in the formal spec from RTL power hotspots until the error falls into a 1–5% band. After that, contracts, Arcmodel, RTL, and power move together. The authors say four searches plus a final round of ablations tried 58 ideas and kept 37.

In the results section, Frostbite was synthesized on a 7 nm process, reaching 800 MHz timing at synthesis and 500 MHz in the final GDS. Each Frostbite quad (a two-by-two grid of tiles) plus four L3 nodes is 0.792 mm². The comparison point is a single core of a modern, silicon-proven superscalar CPU the post calls TCPU (“Traditional CPU”): 0.895 mm² at the same process, described as a dual-issue, in-order processor. A footnote says the exact CPU cannot be disclosed for legal reasons. For a similar-area “single-core” comparison, the post reports an average 8.77× performance-per-area across the workloads, and a geometric mean of 4.59× the TCPU’s performance-per-watt across all 25 workloads (6.02× on the 10 search-set workloads, 3.83× on the 15 held-out workloads). It also says the architecture is strongest on highly parallelizable tasks but still outperforms the TCPU on branching, dynamic code. These are company figures, and they are pre-silicon numbers, not independent reruns by this radar.

Planned next steps include a software playground so anyone can run a program on Frostbite hosted on FPGAs, a preliminary operating system, further performance-per-watt work and broader CPU comparisons, and a release of Frostbite’s RTL and ISA.

Usage

  1. Read the official post at https://architectlabs.com/blog/frostbite , and optionally the company homepage at https://architectlabs.com/ and the earlier Redwood post at https://architectlabs.com/blog/redwood so the two project names are not merged;
  2. The post’s next steps include an FPGA-hosted software playground and a later release of the RTL and ISA. There is no public chip purchase page or ISA download in this article; availability follows later official updates;
  3. When citing the 4.59× performance-per-watt, 8.77× performance-per-area, 7 nm, 800/500 MHz, or 0.792 mm² figures, attribute them to the 18 November 2026 official post, and note that the comparison point is the post’s TCPU (dual-issue, in-order, 0.895 mm² at the same process), whose exact model is not disclosed. These are pre-silicon results;
  4. Do not treat Frostbite as EA’s Frostbite game engine, as Architect Labs’ Redwood accelerator, or treat ASL, Arcmodel, Vine, or Ivy as newly introduced standalone product names from this post. This page only restates the official article.