LLM解码时功率限制形同虚设，真实节能需锁频而非限功。
The Illusion of Power Capping in LLM Decode: A Phase-Aware Energy Characterisation Across Attention Architectures
- 用锁频替代限功，精准控制解码能耗
- 解码仅耗137–300W，远低于700W上限，限功无效
- 适配新注意力架构的节能策略，适合生产级部署
功率限制是LLM服务中标准的GPU能耗调控手段，看似有效：吞吐量下降、功耗读数降低、能耗预算达标。但我们发现，在主导生产环境的自回归解码阶段，这种效果是假象。在NVIDIA H200上，四种注意力范式（GQA、MLA、Gated DeltaNet、Mamba2）的解码仅消耗137–300W（700W GPU），从未触发限功，因内存密集型解码已饱和HBM带宽，计算未达瓶颈，留有大量功率余量。固件主动降频进一步掩盖问题，使吞吐下降被误归因于限功。通过锁定SM频率，可消除上述干扰。锁定频率在所有架构上均优于限功，恢复最多32%解码能耗，且吞吐损失极小。我们识别出三类依赖架构的动态电压频率调节行为模式，并发现新注意力结构共有的能量特征：预填充成本高，但解码高效，最终在生产批大小下总请求能耗较GQA减半。
原文摘要 · Abstract (English)
Power capping is the standard GPU energy lever in LLM serving, and it appears to work: throughput drops, power readings fall, and energy budgets are met. We show the appearance is illusory for the phase that dominates production serving: autoregressive decode. Across four attention paradigms -- GQA, MLA, Gated DeltaNet, and Mamba2 -- on NVIDIA H200, decode draws only 137--300\,W on a 700\,W GPU; no cap ever triggers, because memory-bound decode saturates HBM bandwidth rather than compute and leaves power headroom untouched. Firmware-initiated clock throttling compounds the illusion: these deviations can corrupt any throughput measurement that attributes them to the cap. SM clock locking dissolves both confounds. By targeting the lever that is actually on the critical path, clock locking Pareto-dominates power capping universally, recovering up to 32\% of decode energy at minimal throughput loss. We identify three architecture-dependent DVFS behavioural classes and characterise a common energy pattern across novel attention replacements: a heavy prefill cost recouped by efficient decode, eventually halving total request energy relative to GQA at production batch sizes.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。