PUE is blind to the schedule: energy and carbon trade-offs of GPU training workloads in a high-altitude AI data centre

{"Power":[0],"Usage":[1],"Effectiveness":[2],"(PUE)":[3],"is":[4,35,63,94,127,130,140,180,230],"the":[5,23,26,40,88,91,102,113,131,144,198,203,209,213,238,242,251,273,286],"dominant":[6],"efficiency":[7],"metric":[8],"for":[9],"data":[10,262],"centres,":[11],"and":[12,77,112,139,202,205,250,261,266,270,285],"China's":[13],"\\"East":[14],"Data,":[15],"West":[16],"Computing\\"":[17],"programme":[18],"uses":[19],"it":[20],"to":[21,65,78,96,109,122,136,143,152,220],"rank":[22],"greenness":[24],"of":[25,38,43,52,87,133,174,178,208,237],"western":[27],"hub":[28],"clusters.":[29],"This":[30],"paper":[31],"shows":[32],"that":[33,128,229],"PUE":[34,93,129],"structurally":[36],"incapable":[37],"ranking":[39],"carbon":[41,81,103,172,223,247],"performance":[42],"AI":[44],"training":[45,57],"workloads.":[46],"A":[47],"compute-phase":[48],"/":[49],"communication-phase":[50],"model":[51,69],"a":[53,66,71,190,194],"100":[54],"000-step":[55],"distributed":[56],"job":[58,105],"on":[59,189],"10":[60],"000":[61],"accelerators":[62],"coupled":[64],"temperature-dependent":[67],"cooling":[68,210],"at":[70,163,232],"2":[72],"266":[73],"m":[74],"plateau":[75],"site":[76],"time-varying":[79],"grid":[80,114,204],"intensity.":[82],"Across":[83],"ten":[84],"feasible":[85],"schedules":[86,156],"same":[89],"job,":[90],"mean":[92],"identical":[95,231],"four":[97],"decimal":[98],"places":[99],"(1.1490)":[100],"while":[101,157],"per":[104,116,187],"spans":[106,119],"13.1x":[107,249],"(4.91":[108],"64.25":[110],"tCO2e)":[111],"energy":[115],"delivered":[117],"step":[118],"15.7x":[120],"(0.142":[121],"2.229":[123],"kWh/step).":[124],"The":[125,170,279],"mechanism":[126],"ratio":[132],"facility":[134],"power":[135,160,201],"IT":[137,159],"power,":[138],"therefore":[141],"invariant":[142],"compute":[145],"duty":[146],"cycle,":[147],"which":[148],"ranges":[149],"from":[150,272],"0.685":[151],"0.935":[153],"across":[154],"those":[155],"peak":[158],"stays":[161],"fixed":[162,196],"10.40":[164],"MW.":[165],"Two":[166],"further":[167],"results":[168,280],"follow.":[169],"marginal":[171],"cost":[173],"buying":[175],"one":[176],"accelerator-hour":[177],"makespan":[179],"exactly":[181],"P_comm":[182],"x":[183],"CI":[184],"(0.096":[185],"kgCO2e":[186],"accelerator-kWh":[188],"0.30":[191],"kg/kWh":[192],"grid),":[193],"quantity":[195],"by":[197,224,248,256],"accelerator's":[199],"idle-phase":[200],"completely":[206],"independent":[207],"design.":[211],"And":[212],"entire":[214],"liquid-cooling":[215],"range":[216],"f":[217],"=":[218],"0":[219],"1":[221],"changes":[222],"9.11":[225],"%,":[226],"an":[227],"amount":[228],"all":[233],"3":[234],"840":[235],"points":[236],"parameter":[239],"sweep,":[240],"whereas":[241],"synchronisation":[243],"schedule":[244],"alone":[245,255],"moves":[246],"job's":[252],"start":[253],"hour":[254],"54":[257],"%.":[258],"All":[259],"code":[260],"are":[263,281,288],"openly":[264],"available":[265],"reproduce":[267],"every":[268],"figure":[269],"table":[271],"equations,":[274],"with":[275],"no":[276],"digitised":[277],"data.":[278],"model-derived,":[282],"not":[283],"measured,":[284],"limitations":[287],"stated":[289],"explicitly.":[290]}

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-09-15
DOI
https://doi.org/10.5281/zenodo.22763084
Primary Topic
Cloud Computing and Resource Management
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
preprint

PUE is blind to the schedule: energy and carbon trade-offs of GPU training workloads in a high-altitude AI data centre

Xian Cai
Zenodo (CERN European Organization for Nuclear Research)
Cloud Computing and Resource Management
preprint

PUE is blind to the schedule: energy and carbon trade-offs of GPU training workloads in a high-altitude AI data centre

Xian Cai
preprint en

Abstract

Power Usage Effectiveness (PUE) is the dominant efficiency metric for data centres, and China's "East Data, West Computing" programme uses it to rank the greenness of the western hub clusters. This paper shows that PUE is structurally incapable of ranking the carbon performance of AI training workloads. A compute-phase / communication-phase model of a 100 000-step distributed training job on 10 000 accelerators is coupled to a temperature-dependent cooling model at a 2 266 m plateau site and to time-varying grid carbon intensity. Across ten feasible schedules of the same job, the mean PUE is identical to four decimal places (1.1490) while the carbon per job spans 13.1x (4.91 to 64.25 tCO2e) and the grid energy per delivered step spans 15.7x (0.142 to 2.229 kWh/step). The mechanism is that PUE is the ratio of facility power to IT power, and is therefore invariant to the compute duty cycle, which ranges from 0.685 to 0.935 across those schedules while peak IT power stays fixed at 10.40 MW. Two further results follow. The marginal carbon cost of buying one accelerator-hour of makespan is exactly P_comm x CI (0.096 kgCO2e per accelerator-kWh on a 0.30 kg/kWh grid), a quantity fixed by the accelerator's idle-phase power and the grid and completely independent of the cooling design. And the entire liquid-cooling range f = 0 to 1 changes carbon by 9.11 %, an amount that is identical at all 3 840 points of the parameter sweep, whereas the synchronisation schedule alone moves carbon by 13.1x and the job's start hour alone by 54 %. All code and data are openly available and reproduce every figure and table from the equations, with no digitised data. The results are model-derived, not measured, and the limitations are stated explicitly.

Zenodo (CERN European Organization for Nuclear Research)
Qinghai University (CN)
Cloud Computing and Resource Management
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.