ECO-AI – Energy-Conscious Optimization for AI Training
Автор: VTES
Загружено: 2025-12-18
Просмотров: 14
Описание: ECO-AI: Energy-Conscious Training for Generative Models (CVPR 2025) — This video summarizes the CVPR 2025 paper “ECO-AI – Energy-Conscious Optimization for AI Training,” which presents practical ways to reduce energy use and CO₂ emissions when training diffusion-based image generators and multimodal text-to-image systems. Through controlled experiments on models from ~200M to 3B parameters (run on 8× and up to 64× NVIDIA A100 GPUs) and under two grid scenarios (70% vs 30% renewable), the authors show that mixed precision (FP16) cuts energy by ~15%, emissions can more than double depending on the electricity mix, energy scales almost linearly with training time, and batch-size efficiency improves up to ~64–128 before diminishing returns. The paper concludes with a sustainability playbook: use FP16 by default, pick efficient batch sizes, schedule or locate training on cleaner power, consider optimizer choice as an energy trade-off, reduce wasted compute with early stopping and adaptive schedules, and rely on fine-tuning/transfer learning to avoid full retraining whenever possible.
Повторяем попытку...
Доступные форматы для скачивания:
Скачать видео
-
Информация по загрузке:
