DS-LLM: Leveraging Dynamical Systems to Enhance Both Training and Inference of Large Language Models

Traditional techniques fail to address the fundamental bottleneck. • Emerging techniques like Quantum computing, optical computing, and computing-inmemory are promising but facing significant technical barriers. • Is it feasible to rely on mature CMOS-based technology to accelerate LLM training from 10 million hours to 10,000 hours while reducing energy consumption from 20 TJ to 200 MJ? -> Yes! By training LLMs on DS-machines! Motivation Inference Acceleration: Mapping existing LLM onto DS-machines