Frostbite: An Agent Swarm Designed a CPU, and Lean Kept It Honest
A team let AI agents design a CPU from the instruction set down to the final layout, and it claims 4.59 times the performance per watt of a conventional core. Architect Labs published Frostbite on November 18. It reached the Hacker News front page on Friday, November 20, with 578 points. It is a general-purpose CPU with a brand-new instruction set, and the company says its AI systems iterated autonomously over the compiler, ISA, architecture, microarchitecture, RTL, verification and final GDS. On 25 workloads, 15 of them held out until the final evaluation, Frostbite reached a geometric mean of 4.59x performance per watt over a "silicon-proven superscalar (dual-issue, in-order)" core: 6.02x on the tuning set, 3.83x on the held-out set. It also claims 8.77x performance per area. Post: https://architectlabs.com/blog/frostbite
The trick is interaction nets. Put simply, a program is a graph, and computing means rewriting connected pairs of nodes. Every step is local, every dependency is explicit, and every value has exactly one owner. That means there is no shared mutable state, so the chip does not need deep coherent caches to guard it. The energy breakdown shows the payoff. Leaving out clock distribution, the baseline spends 57% of its energy on memory and 27% on fetching, decoding and issuing instructions. Frostbite cuts those to 38% and 6%, and pays 26% for its on-chip network and scheduler. The agents also came up with work stealing between tiles: a tile runs its own work, takes spare work from a neighbor, or goes to sleep. The company says no human specified that behavior.
The guardrail is the part worth stealing. Instead of English design documents, the team wrote an in-house spec language called ASL that lowers to Lean, and kept those contracts as the single source of truth. By the end there were over 800,000 lines of Lean. In the company's words, a provable source of truth "prevents a large class of bugs, silly mistakes, and reward-hacking instances when deploying large agent swarms." The swarm tried 58 ideas and kept 37. The system model runs 660 times faster than RTL simulation and was realigned until it agreed with the RTL to within 1 to 5%.
Now the discount. These are pre-silicon power numbers, and the post says so: synthesized on 7nm, 800 MHz at synthesis and 500 MHz in final GDS, with no tapeout yet. The baseline is a modest dual-issue in-order core that cannot be named "for legal reasons," not a modern out-of-order design. The comparison is one Frostbite quad (0.792 mmΒ²) against one baseline core (0.895 mmΒ²), and some kernels were run 64 times smaller than standard sizes. The promised FPGA playground and the RTL and ISA release are what will settle it.
My read: this is the same story as the math claims of the same week. Agents get bold, and stay honest, when a formal checker sits underneath them. Lean is quietly becoming the shared trust layer for agent swarms in proofs and in hardware alike. Beating an unnamed in-order core is not beating Apple or AMD. The real product here is the process.
← Back to all articles
The trick is interaction nets. Put simply, a program is a graph, and computing means rewriting connected pairs of nodes. Every step is local, every dependency is explicit, and every value has exactly one owner. That means there is no shared mutable state, so the chip does not need deep coherent caches to guard it. The energy breakdown shows the payoff. Leaving out clock distribution, the baseline spends 57% of its energy on memory and 27% on fetching, decoding and issuing instructions. Frostbite cuts those to 38% and 6%, and pays 26% for its on-chip network and scheduler. The agents also came up with work stealing between tiles: a tile runs its own work, takes spare work from a neighbor, or goes to sleep. The company says no human specified that behavior.
The guardrail is the part worth stealing. Instead of English design documents, the team wrote an in-house spec language called ASL that lowers to Lean, and kept those contracts as the single source of truth. By the end there were over 800,000 lines of Lean. In the company's words, a provable source of truth "prevents a large class of bugs, silly mistakes, and reward-hacking instances when deploying large agent swarms." The swarm tried 58 ideas and kept 37. The system model runs 660 times faster than RTL simulation and was realigned until it agreed with the RTL to within 1 to 5%.
Now the discount. These are pre-silicon power numbers, and the post says so: synthesized on 7nm, 800 MHz at synthesis and 500 MHz in final GDS, with no tapeout yet. The baseline is a modest dual-issue in-order core that cannot be named "for legal reasons," not a modern out-of-order design. The comparison is one Frostbite quad (0.792 mmΒ²) against one baseline core (0.895 mmΒ²), and some kernels were run 64 times smaller than standard sizes. The promised FPGA playground and the RTL and ISA release are what will settle it.
My read: this is the same story as the math claims of the same week. Agents get bold, and stay honest, when a formal checker sits underneath them. Lean is quietly becoming the shared trust layer for agent swarms in proofs and in hardware alike. Beating an unnamed in-order core is not beating Apple or AMD. The real product here is the process.
Comments