跳转至

文章背景与核心概要

在生成式 AI 技术迅猛发展的当下，大语言模型 (Large Language Model, LLM) 训练已经成为现代数据中心电力消耗激增的关键源头，电网供电上限正成为限制算力集群扩张的核心瓶颈。为了在电网负荷高峰期实现动态削峰填谷，系统调度必须精确把握：当调低 GPU 功率墙时，模型训练的 Token 产出速度究竟会发生怎样的变化？本文首次系统性地提出了“任务电力弹性 (Job Power Elasticity)”这一度量体系，并构建了标准化的电力弹性指数 (Power Flexibility Index, PFI)。通过在 NVIDIA H200 与 H100 集群上完成的 131 次大规模实测，团队证实 PFI 感知调度在整体供电减少 30% 的极限约束下，能够弥补 63% 的吞吐量损失，为打造电网友好型与绿色低碳的 AI 基础设施提供了坚实支撑。

表征任务电力弹性：面向柔性电网的 AI 训练动态功率分配

Characterizing Job Power Elasticity for Power-Flexible AI Training

核心执行摘要

Executive Summary

大语言模型 (Large Language Model, LLM) 训练正迅速成为现代数据中心电力消耗增长的核心推手，供电容量的短缺已成为制约 AI 基础设施进一步扩张的瓶颈所在。本文首次对任务电力弹性 (Job Power Elasticity)开展了系统化表征——即研究在降低 GPU 功耗限制时，大模型训练吞吐量所产生的敏感度变化规律。为了量化这一特征，作者提出了电力弹性指数 (Power Flexibility Index, PFI)。这是一个用于评估降功耗性能代价的归一化指标，同时也可作为满足服务等级协议 (SLA) 的功率分配控制原语。通过在 H200 和 H100 GPU 集群上开展 131 次跨越不同模型规模的实测，该研究证明：基于 PFI 感知的智能功率分配机制能够在严苛的供电限制下最大化训练吞吐量，相比于传统的均匀功率削减方案，挽回了极其显著的算力性能损失。

Large language model (LLM) training is rapidly becoming a primary driver of electricity demand in modern data centers, pushing power availability to a critical bottleneck. This paper presents the first systematic characterization of job power elasticity—the sensitivity of training throughput to GPU power reductions. To quantify this behavior, the authors introduce the Power Flexibility Index (PFI), a normalized metric measuring performance cost that doubles as a control primitive for SLA-aware power allocation. Through empirical testing across 131 LLM training runs on H200 and H100 GPUs, the study demonstrates that intelligent, PFI-aware power allocation can maximize throughput under tight energy constraints, recovering a significant portion of performance compared to naive equal-weight allocations.

论文元数据

Paper Metadata

  • arXiv 编号： arXiv:2609.11542 [cs.AI]
  • 提交日期： 2026年9月10日
  • 所属领域： 人工智能 (cs.AI)
  • 论文作者：
  • Philip Colangelo
  • Charles Dawson
  • Shayan Sengupta
  • Ayse Coskun
  • Varun Sivaram
  • arXiv Identifier: arXiv:2609.11542 [cs.AI]
  • Submission Date: September 10, 2026
  • Subjects: Artificial Intelligence (cs.AI)
  • Authors:
  • Philip Colangelo
  • Charles Dawson
  • Shayan Sengupta
  • Ayse Coskun
  • Varun Sivaram

论文摘要

Abstract

大语言模型 (LLM) 训练是现代数据中心电力需求增长最迅猛的来源之一，供电容量已成为制约 AI 基础设施持续扩张的主要瓶颈。若能赋予此类计算工作负载灵活调节功耗的能力，不仅能为 AI 的算力增长释放额外的电力空间，还能平抑电价飙升并提高现有电网基础设施的利用率。然而，要实现这一电力灵活性，我们必须首先弄清：当降低 GPU 功耗时，训练任务的性能究竟会发生怎样的动态变化。

Large language model (LLM) training is among the fastest-growing sources of electricity demand in modern data centers, and power availability is a primary bottleneck to continued AI infrastructure growth. Making the power consumption of these workloads flexible could unlock additional power for AI growth, limit increases in electricity prices, and improve the utilization of existing grid infrastructure. However, to realize this flexibility, we must first understand how the performance of training workloads changes when GPU power is reduced.

本文首次系统化地表征了 LLM 训练中的任务电力弹性 (Job Power Elasticity)（即吞吐量对功率降低的敏感度）。为了对这种弹性进行量化，我们提出了电力弹性指数 (Power Flexibility Index, PFI)，这是一个用于量化功率削减带来的性能代价的归一化指标，并为兼顾服务等级协议 (SLA) 的电力灵活性调度提供了控制原语。

This paper presents the first systematic characterization of job power elasticity (the sensitivity of throughput to power reductions) in LLM training. To quantify elasticity, we introduce the Power Flexibility Index (PFI), a normalized metric that quantifies the performance cost of power reductions and provides a control primitive for SLA-aware power flexibility.

我们从 131 次基于 H200 的 LLM 训练任务（外加 24 次 H200 验证任务以及 34 次匹配的 H100 任务）中收集了完整数据，涵盖稠密模型与混合专家 (Mixture-of-Experts, MoE) 架构、预训练与微调任务，集群规模最高达 32 块 GPU。我们发现，LLM 训练任务表现出显著但差异较大的电力弹性，并识别出了能够在运行时精准预测 PFI 的遥测信号指标。最终，我们证明了基于 PFI 感知的功率分配能够在电力受限条件下最大化集群总 Token 吞吐率 (Tokens/s)。在整体电力压减 30% 的情况下，基于 PFI 的功率分配方案为每个任务追回约 1.5k Tokens/s 的吞吐量，弥合了均匀分配基准与具备完美信息的理论最优解 (Oracle) 之间 63% 的性能差距。我们的研究成果确立了电力弹性作为训练任务一项可测物理特性的地位，并为构建感知电力、响应电网的现代化 AI 基础设施奠定了理论基础。

We collect data from 131 LLM training runs on H200 (plus 24 H200 validation runs and 34 matched H100 runs), including both dense and mixture-of-experts models, pretraining and fine-tuning tasks, and up to 32 GPUs. We find that LLM training jobs exhibit substantial but variable power elasticity, and we identify telemetry signals that predict PFI at runtime. Finally, we demonstrate that PFI-aware power allocation maximizes total tokens/second throughput under power constraints. Under a 30% power reduction, PFI-aware power allocation recovers ~1.5k tokens/s per job, 63% of the performance gap between an equal-weight allocation and an oracle with perfect information. Our results establish power elasticity as a measurable property of training jobs and provide a foundation for power-aware, grid-responsive AI infrastructure.

核心贡献与关键发现

Key Contributions & Findings

  1. 首度系统性表征： 全面测定了涵盖稠密模型、混合专家 (MoE) 架构、预训练及微调任务在内的 LLM 训练工作负载的任务电力弹性。
  2. 电力弹性指数 (PFI)： 提出了一个标准归一化指标来衡量削减功率输入所带来的性能代价，可作为高效的运行时控制原语。
  3. 运行时遥测预测： 发现了能够在大模型运行期间精准预判 PFI 指标的具体硬件与系统遥测信号。
  4. 优化的功率分配策略： 提出了 PFI 感知的电力管理方案，在严苛电力约束下成功最大化全局吞吐率 (Tokens/s)，在削减 30% 电力时单任务挽回约 1.5k Tokens/s 的性能。
  1. First Systematic Characterization: Establishes job power elasticity for LLM training workloads across dense models, mixture-of-experts (MoE) architectures, pretraining, and fine-tuning tasks.
  2. Power Flexibility Index (PFI): Introduces a normalized metric to measure the performance cost associated with reduced power inputs, acting as an effective runtime control primitive.
  3. Runtime Telemetry Prediction: Identifies specific telemetry signals capable of accurately predicting PFI during active execution.
  4. Optimized Power Allocation: Proposes PFI-aware power management that successfully maximizes throughput (tokens/second) under strict power constraints, recovering ~1.5k tokens/s per job under a 30% power reduction.

论文获取与资源

Access & Resources