Distributed LLM Pretraining During Renewable Curtailment Windows: A Feasibility Study

TL;DR

This study explores the feasibility of training large language models during renewable energy curtailment periods by leveraging geo-distributed GPU clusters, aiming to reduce emissions while maintaining training quality.

Contribution

It introduces a system for distributed LLM training during curtailment windows, combining local and federated training, with a prototype demonstrating reduced emissions.

Findings

Curtailment-aware scheduling maintains model quality.

Operational emissions reduced to 5-12% of baseline.

Prototype successfully trains a 561M-parameter model across clusters.

Abstract

Training large language models (LLMs) requires substantial compute and energy. At the same time, renewable energy sources regularly produce more electricity than the grid can absorb, leading to curtailment, the deliberate reduction of clean generation that would otherwise go to waste. These periods represent an opportunity: if training is aligned with curtailment windows, LLMs can be pretrained using electricity that is both clean and cheap. This technical report presents a system that performs full-parameter LLM training across geo-distributed GPU clusters during regional curtailment windows, elastically switching between local single-site training and federated multi-site synchronization as sites become available or unavailable. Our prototype trains a 561M-parameter transformer model across three clusters using the Flower federated learning framework, with curtailment periods derived from real-world marginal carbon intensity traces. Preliminary results show that curtailment-aware scheduling preserves training quality while reducing operational emissions to 5-12% of single-site baselines.