AI agents design an entirely new CPU architecture from scratch, startup claims
Reported by Tom's Hardware1 source ↓
Architect Labs says its AI systems designed and verified Frostbite, an interaction-net CPU. In simulations across 25 workloads, a four-tile cluster delivered 4.59 times the performance per watt of an unnamed in-order core; Frostbite has not been built in silicon.
Key points
- The 4.59x performance-per-watt result is from a simulated 7nm design, not a silicon chip
- Frostbite’s four-tile cluster averaged 8.77x the comparison core’s performance per area
- Performance-per-watt gains ranged from 1.15x on recursive Fibonacci to 23.12x on conv2d
Sources
Details
Frostbite programs use interaction nets, a computation model in which programs are graphs and computation rewrites pairs of connected nodes. Architect Labs says this exposes parallel work in the program itself, avoiding conventional out-of-order hardware and shared-cache coherence. Each tile has local memory, and the full design has 16 tiles; the reported benchmark used four tiles and four L3 nodes.
The simulated design reached 500 MHz in its final layout, after reaching 800 MHz in early synthesis. Across 25 workloads, the company reports a 4.59x geometric-mean performance-per-watt advantage over a similarly sized dual-issue, in-order core; it says the comparison core cannot be named for legal reasons. Gains averaged 6.02x on 10 workloads used to tune the design and 3.83x on 15 held out for final evaluation.
Architect Labs says its agents developed a machine-checkable specification and a software model before producing the circuit design. The company reports that the model ran about 200 times faster than RTL simulation, and that an automated process reduced its power-estimate error from as much as nearly 250% to within 5% of the circuit design. Agents tried 58 design changes, of which 37 were kept.
Written from coverage as of
Images
- Tom's Hardware