Mumakore gumi apfuura, magadzirirwo edhisiki rema "hyperscale" ave achinyanya kutarisisa ma "x86 server fleets" akagadzirirwa kushandiswa pakuverenga zvinhu zvakasiyana-siyana. Nguva iyoyo iri kupera. Nekugadzira AI, ma "foundation models", simulation, uye accelerated analytics zvino zvichishandisa huwandu husina kumboonekwa hwe "compute", ma "hyperscalers" ari kuchinja nekukurumidza kuenda kuma "GPU-first architectures" - uko ma "graphics processing units", "accelerators", uye custom silicon hazvisi zvechipiri zve "add-ons", asi mainjini makuru e "compute".

Kuchinja uku kuri kuchinja dhizaini yedatacenter, economics, supply chains, uye software ecosystems pasi rose. Heino nzira iyo vagadziri vema "hyperscalers" vari kugadzirira ramangwana reGPU-first, uye izvi zvinorevei kune vamwe vese vari muindasitiri.

Kugadzirazve Datacenters dzeHigh-Density GPU Clusters

Kare, ma racks akagadzirwa akapoteredza CPU thermals — kazhinji haapfuure 8–12 kW pa raki.
Ma AI clusters emazuva ano anopfuura 30 kW, 60 kW, uye kunyange 100+ kW pa rack.

Vashandi veHyperscal vari kupindura ne:

Kutonhodza kwemvura sechikanganiso

• Zvidimbu zve plate inotonhora zvakananga ku GPU nodes

• Zvishandiso zvekupisa zvemusuwo wekumashure zvema hybrid fleets

• Kuvandudzwa kwezvivakwa zvemvura zvechivako

• Zvikamu zvekugovera zvinotonhorera (CDUs) mumutsara-wero magadzirirwo

Mapodhi Akakosha Ane Density Yepamusoro

• Mitsetse yeGPU chete ine nzvimbo yakasimba yekupisa

• Nzira dzemhepo dzakaparadzana

• Simba nekutonhodza zvisingaenderane nehurongwa hwemakombiyuta

Kuronga kugona kweThermal-aware

AI clusters ikozvino inotyaira sarudzo yenzvimbo, kwete maCPU.

Kugona kutonhora kunosarudza:

• kuti maGPU mangani anogona kuiswa

• kwaanogona kuva yakaiswa

• kuti mapoka anogona kukura nekukurumidza sei

Kugadzirazve Kutumira Simba reDatacenter

Raki imwe chete yeAI accelerators inogona kukwevera 50+ kW, zvichikonzera kushushikana kukuru pa zvivakwa zvemagetsi.

Vashandi veHyperscal vari kuita izvi ne: