Cutting AI Datacenter Energy with Reinforcement Learning: Measured Power Control of LLM Training from One GPU to the Fleet
The episode analyzes a paper detailing how to cut AI datacenter energy consumption using Reinforcement Learning (RL) power control. The discussion covers RL's success on single GPUs, its failure on sharded models, and its successful reapplication using 'generation concurrency.' Ultimately, the paper suggests AI training can significantly reduce required grid capacity.