RL Controller Slashes GPU Power Violations, Boosts LLM Output by 18%
Instrumenting A100 GPUs with half‑second power telemetry, the authors trained a PPO meta‑controller that tweaks LLM generation parameters on‑the‑fly. On a 7B model it cut power‑limit violations by 89.8%, lifted token output 18.1% and improved energy efficiency 26.2%; a live 72B rollout saw 35% more output with far fewer violations.