Cooling Architecture Optimization and PUE Modelling for Hyperscale AI Data Centres in High-Altitude Low-Pressure Environments: A Case Study on the Qinghai–Tibetan Plateau

{"Preprint":[0],"of":[1,32,39,99,105,111,193,202,299,362,385],"a":[2,65,157,268,376],"modelling":[3],"study":[4,410],"on":[5,115,166,302],"cooling":[6,57,116,140,174,221,242,295,343],"architecture":[7,117],"selection":[8,118],"for":[9,156,281,415],"hyperscale":[10],"AI":[11,163,417],"data":[12,164,418],"centres":[13,419],"located":[14],"at":[15,214,307,311,344],"high":[16],"altitude,":[17],"where":[18,86,97],"atmospheric":[19,103],"pressure":[20],"is":[21,244,296,398],"only":[22,106,365],"70-80":[23],"kPa":[24],"and":[25,59,90,144,147,187,205,208,226,262,286,309,355,388,420,427],"air":[26,89,136,241],"density":[27,137,179,403],"falls":[28,223,328],"to":[29,138,150,233,250,257,329],"about":[30],"80%":[31,260],"its":[33,324],"sea-level":[34],"value.":[35],"The":[36,109,170,409],"rapid":[37],"scaling":[38],"artificial-intelligence":[40],"(AI)":[41],"computing":[42],"clusters":[43],"has":[44,119],"pushed":[45],"rack":[46,379,401],"power":[47,61,228,402],"densities":[48],"from":[49,331],"5–10":[50],"kW":[51,378],"toward":[52],"100":[53],"kW,":[54],"making":[55,60],"liquid":[56,342,396],"mandatory":[58],"usage":[62],"effectiveness":[63],"(PUE)":[64],"first-order":[66],"design":[67,270,413],"objective.":[68],"China's":[69],"\\"East":[70],"Data,":[71],"West":[72],"Computing\\"":[73],"(东数西算)":[74],"initiative":[75],"encourages":[76],"siting":[77],"such":[78,83],"facilities":[79],"in":[80,124,390],"western":[81],"provinces":[82],"as":[84,177,185,189],"Qinghai,":[85],"cold":[87],"ambient":[88],"abundant":[91],"renewable":[92],"electricity":[93],"are":[94],"attractive,":[95],"but":[96],"elevations":[98],"2,000–3,000":[100],"m":[101,216],"imply":[102],"pressures":[104],"70–80":[107],"kPa.":[108],"effect":[110],"this":[112,290],"low-pressure":[113,335],"environment":[114],"received":[120],"little":[121],"quantitative":[122,412],"treatment":[123],"the":[125,152,167,178,182,194,198,247,253,300,303,313,316,334,338,348,356,359,363,394,422],"literature.":[126],"This":[127],"paper":[128],"develops":[129],"an":[130,190],"analytical":[131],"model":[132,171],"that":[133,213,375,389],"couples":[134],"altitude-dependent":[135,199],"air-side":[139,173,200,220,293,326,428],"capacity,":[141],"fan":[142,227],"energy":[143,386,407],"hence":[145],"PUE,":[146,354],"uses":[148],"it":[149],"determine":[151],"optimal":[153],"liquid-cooling":[154,195,248],"fraction":[155,196,249,397],"10,000-accelerator":[158],"(12":[159],"MW":[160],"IT)":[161],"green":[162],"centre":[165],"Qinghai–Tibetan":[168],"Plateau.":[169],"expresses":[172],"capacity":[175,222],"degradation":[176],"ratio":[180],"ρ/ρ₀,":[181],"fan-energy":[183],"penalty":[184,256,336],"ρ⁻²,":[186],"PUE":[188,237,255,264,319],"explicit":[191],"function":[192],"f,":[197],"coefficient":[201],"performance":[203],"(COP),":[204],"fixed":[206],"distribution":[207],"auxiliary":[209],"losses.":[210],"Results":[211],"show":[212],"2,200":[215],"(ρ/ρ₀":[217],"=":[218,265,346],"0.805)":[219],"by":[224,230,238,277,400,406],"19.5%":[225],"rises":[229],"54.3%":[231],"relative":[232],"sea":[234],"level,":[235],"increasing":[236],"+0.068":[239],"when":[240],"alone":[243],"used;":[245],"raising":[246],"0.8":[251,347],"reduces":[252],"altitude-induced":[254],"+0.014":[258],"(an":[259],"mitigation)":[261],"yields":[263],"1.154,":[266],"satisfying":[267],"≤1.15":[269],"target.":[271],"A":[272,370],"full-year":[273],"hourly":[274],"simulation":[275],"driven":[276],"measured":[278],"TMYx":[279],"weather":[280],"Xining":[282],"(2,266":[283],"m),":[284],"Beijing":[285,308],"Shanghai":[287],"then":[288],"sharpens":[289],"conclusion:":[291],"although":[292],"free":[294],"available":[297],"91.7%":[298],"year":[301],"plateau":[304,314,357,392],"against":[305],"65.3%":[306],"54.7%":[310],"Shanghai,":[312],"records":[315],"worst":[317],"air-only":[318],"(1.230":[320],"versus":[321],"1.200)":[322],"because":[323],"mean":[325],"COP":[327],"6.47":[330],"8.60":[332],"—":[333],"outweighs":[337],"climate":[339],"benefit.":[340],"With":[341],"f":[345,367,381],"three":[349,364],"sites":[350],"converge":[351],"within":[352],"0.006":[353],"becomes":[358],"most":[360],"efficient":[361],"above":[366],"≈":[368],"0.96.":[369],"rack-density":[371],"constraint":[372],"analysis":[373],"shows":[374],"45":[377],"requires":[380],"≥":[382],"0.56–0.67":[383],"irrespective":[384],"price,":[387],"low-electricity-price":[391],"regions":[393],"optimum":[395],"governed":[399],"rather":[404],"than":[405],"cost.":[408],"provides":[411],"guidelines":[414],"high-altitude":[416],"identifies":[421],"trade-off":[423],"between":[424],"free-cooling":[425],"availability":[426],"heat-transfer":[429],"degradation.":[430]}

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-09-14
DOI
https://doi.org/10.5281/zenodo.22743561
Primary Topic
Heat Transfer and Optimization
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
preprint

Cooling Architecture Optimization and PUE Modelling for Hyperscale AI Data Centres in High-Altitude Low-Pressure Environments: A Case Study on the Qinghai–Tibetan Plateau

Xian Cai
Zenodo (CERN European Organization for Nuclear Research)
Heat Transfer and Optimization
preprint

Cooling Architecture Optimization and PUE Modelling for Hyperscale AI Data Centres in High-Altitude Low-Pressure Environments: A Case Study on the Qinghai–Tibetan Plateau

Xian Cai
preprint en

Abstract

Preprint of a modelling study on cooling architecture selection for hyperscale AI data centres located at high altitude, where atmospheric pressure is only 70-80 kPa and air density falls to about 80% of its sea-level value. The rapid scaling of artificial-intelligence (AI) computing clusters has pushed rack power densities from 5–10 kW toward 100 kW, making liquid cooling mandatory and making power usage effectiveness (PUE) a first-order design objective. China's "East Data, West Computing" (东数西算) initiative encourages siting such facilities in western provinces such as Qinghai, where cold ambient air and abundant renewable electricity are attractive, but where elevations of 2,000–3,000 m imply atmospheric pressures of only 70–80 kPa. The effect of this low-pressure environment on cooling architecture selection has received little quantitative treatment in the literature. This paper develops an analytical model that couples altitude-dependent air density to air-side cooling capacity, fan energy and hence PUE, and uses it to determine the optimal liquid-cooling fraction for a 10,000-accelerator (12 MW IT) green AI data centre on the Qinghai–Tibetan Plateau. The model expresses air-side cooling capacity degradation as the density ratio ρ/ρ₀, the fan-energy penalty as ρ⁻², and PUE as an explicit function of the liquid-cooling fraction f, the altitude-dependent air-side coefficient of performance (COP), and fixed distribution and auxiliary losses. Results show that at 2,200 m (ρ/ρ₀ = 0.805) air-side cooling capacity falls by 19.5% and fan power rises by 54.3% relative to sea level, increasing PUE by +0.068 when air cooling alone is used; raising the liquid-cooling fraction to 0.8 reduces the altitude-induced PUE penalty to +0.014 (an 80% mitigation) and yields PUE = 1.154, satisfying a ≤1.15 design target. A full-year hourly simulation driven by measured TMYx weather for Xining (2,266 m), Beijing and Shanghai then sharpens this conclusion: although air-side free cooling is available 91.7% of the year on the plateau against 65.3% at Beijing and 54.7% at Shanghai, the plateau records the worst air-only PUE (1.230 versus 1.200) because its mean air-side COP falls to 6.47 from 8.60 — the low-pressure penalty outweighs the climate benefit. With liquid cooling at f = 0.8 the three sites converge within 0.006 PUE, and the plateau becomes the most efficient of the three only above f ≈ 0.96. A rack-density constraint analysis shows that a 45 kW rack requires f ≥ 0.56–0.67 irrespective of energy price, and that in low-electricity-price plateau regions the optimum liquid fraction is governed by rack power density rather than by energy cost. The study provides quantitative design guidelines for high-altitude AI data centres and identifies the trade-off between free-cooling availability and air-side heat-transfer degradation.

Zenodo (CERN European Organization for Nuclear Research)
Qinghai University (CN)
Life below water
Heat Transfer and Optimization
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.