Trusence Every claim has a source
Last updated 22 November 2026 Search Türkçe
Research

Startup says its AI agents designed a new CPU architecture, Frostbite, in simulation

The chip has not been fabricated, so the performance-per-watt gains rest on simulations against an undisclosed comparison core.

A brass clockwork automaton beside an unfinished chip in its mould, with a magnifying glass resting on an unrolled blueprint on a wooden workbench.
AI-generated illustration

Palo Alto startup Architect Labs says its AI agents designed Frostbite, a general-purpose CPU with a brand-new instruction set, and simulations show a four-tile cluster reaching 4.59 times the performance per watt of a comparable dual-issue, in-order core across 25 workloads. The figures come from a 7nm design that has not been fabricated, and the company declined to name the comparison core. Frostbite runs programs compiled to interaction nets, so parallelism is inherent in the code rather than being extracted by out-of-order hardware or managed through threads and locks. The agents wrote the specification in the company's own ASL language, which maps to the Lean proof checker so it can be machine-checked, and built a software model about 200 times faster than RTL simulation to test ideas. Gains ranged from 1.15x on recursive Fibonacci to 23.12x on conv2d, and the company's previous AI-generated accelerator, Redwood, still runs only on an FPGA.

Why it matters

If the claims hold, AI-driven design could shorten the path to new processor architectures, which have traditionally taken large teams and years of work. The result stays unproven until Frostbite is built in silicon and measured against a named baseline, and the company's earlier AI-generated accelerator, Redwood, still runs only on an FPGA. Readers should treat the figures as company-reported simulation results rather than established performance.

Sources · 1