The Illusion of Power Capping in LLM Decode: A Phase-Aware Energy Characterisation Across Attention Architectures

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ma, Bole, Afzal, Ayesha, Eitzinger, Jan, Wellein, Gerhard
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909035898339328
author Ma, Bole
Afzal, Ayesha
Eitzinger, Jan
Wellein, Gerhard
author_facet Ma, Bole
Afzal, Ayesha
Eitzinger, Jan
Wellein, Gerhard
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2605_11999
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle The Illusion of Power Capping in LLM Decode: A Phase-Aware Energy Characterisation Across Attention Architectures
Ma, Bole
Afzal, Ayesha
Eitzinger, Jan
Wellein, Gerhard
Distributed, Parallel, and Cluster Computing
Artificial Intelligence
Machine Learning
Performance
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.
title The Illusion of Power Capping in LLM Decode: A Phase-Aware Energy Characterisation Across Attention Architectures
topic Distributed, Parallel, and Cluster Computing
Artificial Intelligence
Machine Learning
Performance
url https://arxiv.org/abs/2605.11999