Frostbite is a real CPU architecture. The 4.59x is not yet a checkable number
Architect Labs says its agent swarms designed Frostbite, a new CPU instruction set, at 4.59x performance per watt. The rival core and the RTL stay private.

Architect Labs put Frostbite up on 20 November: a general-purpose CPU with an instruction set built from scratch, where agent swarms iterated over the compiler, the ISA, the architecture, the microarchitecture, the RTL, verification and the final GDS. The headline number is a 4.59x performance-per-watt multiple against a silicon-proven superscalar core, a geometric mean across 25 workloads, 15 held out until the final evaluation. The RTL and the ISA are not published. Neither is the name of the CPU it beat.
The step up from Redwood is the architecture itself
The previous post, Redwood, made a narrower claim: given an architecture, the company's systems could implement, verify and deploy an inference accelerator quickly. Frostbite moves the claim one level up: the architecture and instruction set are now searched rather than handed in.
The method is described in unusual detail, to the company's credit. Agents wrote a 141-page design PDF covering programming model, runtime and high-level blocks, then deleted it once the first RTL existed because the estimates were "well off what was physically possible". It was replaced by an in-house formal specification language, ASL, that lowers to Lean and keeps RTL, system model, UVM and SVA agreed on one machine-checkable source of truth: more than 800,000 lines of Lean across 31 blocks by the end. Four searches plus a final ablation tried 58 ideas and kept 37.
The mechanism is a 1990 result, and the post says so
None of the hardware theory is new, and the post does not pretend otherwise. Interaction nets come from Lafont's 1990 POPL paper, which is cited in the references. A program is a graph, computation is the local rewriting of connected node pairs, and every value has exactly one owner. Because each node has one principal port, active pairs cannot overlap, so rewrites commute and can run in any order.
The contribution is the claim that this property is worth silicon: parallelism comes out of the program instead of out of a reorder buffer, and data can sit in small local memories rather than a coherent cache hierarchy. Programs are written in Vine, which compiles ordinary-looking code to the net. On the post's five-queens example the net runs 3,183 interactions on a critical path of 262 steps, roughly twelve independent interactions ready at any moment.
That argument is checkable in principle, and it is the strongest thing in the release: a compiler either produces those nets or it does not. Without the ISA, nobody outside can check it end to end.
Twenty-five workloads, ten of them tuned on
Every architecture multiple rests on a suite, so the suite matters more than the multiple. Frostbite's 25 workloads are grouped by how a program exposes work — independent pieces, staged rounds, sequential chains, branches, recursion — each group split into a search set and a held-out set. The split is disclosed, and their own chart prints the spread: 23.12x on conv2d at the top, 1.15x on recursive Fibonacci at the bottom, with geometric means of 6.02x on the ten search workloads and 3.83x on the fifteen held out. The 3.83x is the number that matters. The 4.59x blends the tuned set into it.
| Item | Figure | Status |
|---|---|---|
| Workloads | 25 (10 search, 15 held out) | split disclosed |
| Perf/W, all 25 | 4.59x vs TCPU | self-reported, pre-silicon |
| Perf/W, held-out 15 | 3.83x | never tuned on |
| Perf/area | 8.77x | 0.792 vs 0.895 mm² at 7 nm |
| RTL, ISA, comparator | not released | promised, or withheld |
One line in the method undercuts the phrase "general-purpose" more than any benchmark number. The FFT, radixsort and LU kernels follow SPLASH-2 "although 64 times smaller, because at regular sizes they are not suitable for exploratory work". A 64x-smaller kernel fits in the local memory that is the design's main advantage, and exercises a different part of the machine. Shrinking benchmarks for exploration is normal; presenting the result as evidence about the workload class is a separate claim, and this post does not separate them.
The comparator has no name
The footnote is one sentence: "For legal reasons, we cannot disclose the exact CPU." What is disclosed is that it is dual-issue and in-order, 0.895 mm² at the same 7 nm node against 0.792 mm² for a Frostbite quad plus its L3 node, and silicon-proven. A dual-issue in-order core is a modest baseline; the wide out-of-order cores in Apple, AMD and Arm parts are absent, and the post concedes as much by promising later runs against "larger, wider, out-of-order CPUs" and against multi-core arrangements with large caches, on SPEC and PARSEC. That is the right comparison. It is also the one this post does not contain.
Naming the part would cost nothing, and would let anyone with a synthesis license check the area and power figures.
Pre-silicon watts, at 500 MHz
Performance per watt is a ratio, and both halves are modelled here. Synthesis closed at 800 MHz; the final GDS closed at 500 MHz. Power comes from Arcmodel, recalibrated against RTL until the two agreed within 5% after starting near 250% worst-case error. The authors are direct: these are "still pre-silicon power numbers", and correlating them to production silicon takes tapeout loops they have not run.
The energy breakdown explains where the savings sit. Excluding clock distribution, the traditional core spends 57% of its energy on memory and 27% on fetch, decode and issue; the Frostbite quad spends 38% and 6%, and pays 26% for its on-chip network and scheduler. That is a coherent story about a real design trade. It is also, at 500 MHz, an energy-per-operation claim rather than a latency claim.
What would turn the number into a result
The closing list is short and specific. "Deliver a software playground for anybody to run a program on Frostbite, hosted on FPGAs." "Build a preliminary operating system capable of managing hardware resources, multiple programs, virtualization, and more." Scale the swarms and benchmark "on a much larger set of workloads against a broader set of CPUs". And: "Release the RTL and ISA for Frostbite."
All four are future work. Until they land, the accurate headline is narrower: a group that has already delivered one AI-designed accelerator says it can now search an instruction set from a workload description, argues coherently for why that should be efficient, and reports a multiple nobody outside can reproduce. The Hacker News thread drew 578 points and spent most of them on interaction nets rather than on the anonymous baseline — the right instinct about the architecture, the wrong one about the benchmark.
Three levers normally make a hardware claim independent: a workload suite the vendor did not choose, a comparator the reader can name, and a design the reader can inspect. Frostbite ships none of the three on day one. The company says it will ship two of them, which is more than most announcements offer, and less than a step-function claim needs.


