Startup Claims AI Autonomously Designs Entirely New CPU Architecture, Simulation Results Outperform Traditional Cores
1 day ago / Read about 0 minute
Author:小编   

A U.S.-based AI startup, Architect Labs, claims that its AI system has designed from scratch a completely new general-purpose CPU architecture called Frostbite, based on an interactive network mathematical model. The design covers the entire process—from instruction set and microarchitecture to RTL code, verification, and GDS layout. Simulation results on a 7nm process show that this architecture achieves 4.59 times the performance per watt of a traditional reference CPU core on the same process, and 8.77 times the performance per unit area. These advantages stem primarily from a significant reduction in memory traffic and instruction overhead, with outstanding performance in parallel and branch-dynamic tasks. Currently, the architecture remains in the experimental stage, and no mass production plans have been announced. The company focuses on an AI end-to-end chip design and verification platform that can compress the cycle from chip concept to FPGA prototype down to two weeks. Its self-developed verification system greatly reduces reliance on traditional commercial EDA tools, with the goal of lowering the barrier to chip design and driving a new type of chip architecture transformation based on full-stack software-hardware co-design.