Chip ManufacturingLOT 261122 · 4 min read

AI Agents Designed a New CPU From Scratch, Startup Claims 4.59x Efficiency

Architect Labs says AI agents built Frostbite, a 16-tile CPU on interaction nets, delivering 4.59x perf/watt over an in-order core in 7nm simulation — with no silicon yet.

By Tom Whitfield869 words

Key takeaways

  1. Frostbite delivered 4.59x perf/watt over an in-order core across 25 workloads in simulated 7nm, dropping to 3.83x on 15 held-out workloads.
  2. AI agents started with 'zero architectural priors' and deleted the original 141-page spec after generating the first RTL.
  3. Of 58 agent-proposed changes, 37 were kept; work stealing emerged without human specification.
  4. The 16-tile design runs at 500 MHz in final layout, down from 800 MHz in early synthesis.
  5. Predecessor Redwood, an AI inference accelerator, was generated in under two weeks and still runs only on FPGA.
AI agents design an entirely new CPU architecture from scratch, startup claims
AI agents design an entirely new CPU architecture from scratch, startup claims — Magnascan / pixabay

Palo Alto startup Architect Labs says its AI agents designed, verified and specified an entirely new general-purpose CPU — and in pre-silicon simulation, the chip delivers 4.59 times the performance per watt of a conventional in-order core. The figure is a geometric mean across 25 workloads, measured on a simulated 7nm design. No silicon exists yet, and the company says it cannot name the comparison core for legal reasons.

The processor, called Frostbite, uses a new instruction set its AI systems built, in the company's words, "from the ground up." Across the same 25-workload set, the four-tile cluster averaged 8.77 times the comparison core's performance per area.

How solid are the numbers?

The headline efficiency figure hides a spread. On the 10 workloads used to tune the design, Frostbite reached 6.02x the comparison core's performance per watt. On the 15 workloads held out until final evaluation, the lead fell to 3.83x. The best held-out workload, conv2d, hit 23.12x; the worst, recursive Fibonacci, managed only 1.15x.

The clock story is equally mixed. Frostbite reached 800 MHz in early synthesis but settled at 500 MHz in the final layout. The full chip contains 16 tiles — small identical processing units linked by an on-chip network — while the benchmarked cluster uses four tiles plus four L3 nodes.

Why throw out the CPU rulebook?

Conventional CPUs carry hardware that, as the team puts it, "rediscovers, at runtime, which instructions actually depend on each other." Multicore chips also need deep cache hierarchies and coherence protocols, which scale poorly and consume area and power.

Frostbite takes a different route the company calls "multithreading for free." Programs compile to interaction nets, a 1990 model of computation in which a program is a graph of nodes and computation rewrites pairs of connected nodes. Rewrites never overlap, so they can run in any order and reach the same result. Every value is used exactly once, so data stays in small local memories rather than large shared caches.

The team's example: a puzzle placing five queens on a 5x5 board. A typical program searches board by board. As an interaction net, the same search splits into 3,183 small operations with a critical chain of only 262 steps, leaving about 12 operations ready to run in parallel at each step. Nobody rewrote the program; the parallelism was already intrinsic.

How did the AI agents actually work?

The agents started from what Architect Labs calls "zero architectural priors" — no inherited assumptions about what a CPU should look like. They read papers and benchmarks, built Python models, and selected one concept to pursue. The original 141-page PDF spec, containing a programming model, runtime and high-level blocks, was deleted after the first RTL design was generated, because "many of the numbers and estimates were well off what was physically possible."

The agents then abandoned English, which the team says cannot capture every detail, and wrote the specification in the company's own language, ASL. ASL translates to Lean, a proof checker, so the spec is machine-checked — a step the company says blocks a large class of bugs and stops agents gaming their objectives. From this came Arcmodel, a full software model of the chip written before any circuit design existed.

Because RTL simulation is slow, the agents tested ideas on Arcmodel, which ran roughly 200 times faster. Its power estimates were initially off by up to nearly 250%; an automated loop corrected the model until it matched the circuit design within 5%.

Swarms of agents then proposed changes in parallel. Each idea was checked against the spec, tried in the fast model, built into the RTL and measured for power before being kept or dropped. Of 58 ideas tried, 37 survived. One behavior "was never specified by any human": work stealing, where an idle tile pulls spare work from a neighbor or powers down.

Where does the energy go?

The power breakdown explains the gains. On a four-tile cluster, mean of 20 workloads, memory takes 38% of Frostbite's energy versus 57% for the comparison core. Instruction fetch, decode and issue cost 6% versus 27%, though Frostbite adds a 26% share for its on-chip network and scheduler. Clock distribution consumes 41% of total energy versus 49%.

This is not Architect Labs' first silicon project. In August it showed Redwood, an AI inference accelerator for robots and drones, whose RTL, verification, firmware, drivers and kernels its AI generated in under two weeks from a spec by two human architects. Redwood runs only on an FPGA, with the company working toward TSMC fabrication.

A potential rival, Cambridge-based Tendrils Compute, raised funding this month for a processor it claims will have "thousands of cores operating in parallel." It is also developing Vine, the same language Frostbite programs use.

Architect Labs plans an FPGA-hosted "software playground for anybody to run a program on Frostbite," a preliminary operating system, and eventual release of the RTL and instruction set. Broader comparisons — larger mesh configurations, multicore comparison cores with big caches, and wider out-of-order CPUs on SPEC and PARSEC — are due in the coming months.

Tom's Hardware (Source)

Filed under

Share this article:

More from Tom Whitfield

Tom Whitfield

Show full bio

Staff writer covering companies, deals and strategy in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,200 articles

Read also

« Previous article