Iyo DGX A100 system, yakavakirwa pane iyo Jen-Hsun Huang nguva pfupi yadarika , inosanganisira masere eA100 GPUs, matanhatu NVLink 3.0 switch, mapfumbamwe Mellanox network controllers, maviri AMD EPYC Rome-chizvarwa processors ane 64 cores, 1 TB ye RAM uye 15 TB yeSSD ine NVMe rutsigiro.
NVIDIA DGX A100 chizvarwa chechitatu chekambani yemakomputa masisitimu, akagadzirirwa zvakanyanya kugadzirisa matambudziko ehungwaru hwekugadzira. Iye zvino masisitimu akadaro akavakirwa paazvino A100 graphics processors kubva kumhuri yeAmpere, izvo zvinokonzeresa kuwedzera kwakanyanya mukuita kwavo, iyo yasvika mashanu petaflops. Nekuda kweizvi, iyo DGX A5 inokwanisa kubata yakanyanya kuomarara maAI modhi uye yakakura mavhoriyamu edata.
Kune iyo DGX A100 system, NVIDIA inongoratidza huwandu hwese hwehBM2 memory, inosvika mazana matatu nemakumi maviri GB. Masvomhu akareruka anotibvumira kuona kuti GPU yega yega ine 320 GB yendangariro, uye mifananidzo yechigadzirwa chitsva inoburitsa pachena kuti vhoriyamu iyi yakagoverwa pakati pezvimedu zvitanhatu. Iyo graphics memory bandwidth inotaurwa zvakare - 40 TB / s yeiyo yese DGX A12,4 system yakazara.
Tichifunga kuti iyo DGX-1 system, yakavakirwa pamasere eTesla V100s, yakagadzira imwe petaflops mune yakasanganiswa-chaiyo macalculation, uye iyo DGX A100 inonzi inoita pamapetaflops mashanu, tinogona kufunga kuti mukuverenga chaiko Ampere GPU imwe inomhanya zvakapetwa kashanu kupfuura. yakatangira neVolta architecture. Mune zvimwe zviitiko, mukana wacho unova wakapetwa makumi maviri.
Pakazara, iyo DGX A8 system inopa kuita kwepamusoro-soro kwekushanda kwe100 pasekondi mune yakazara mashandiro (INT1016), muhafu-chaiyo inoyangarara nzvimbo (FP16) - 5 petaflops, mune kaviri-chaiyo inoyangarara nzvimbo (FP64) - 156 teraflops. . Pamusoro pezvo, iyo DGX A32 inowana kuita kwepamusoro kwe100 petaflops muTF2,5 tensor komputa. Ngatiyeukei kuti teraflops imwe ndeye 1012 inoyangarara nzvimbo pasekondi, imwe petaflops ndeye 1015 inoyangarara nzvimbo pasekondi.
Chinhu chakakosha cheNVIDIA A100 accelerator kugona kupatsanura zviwanikwa zveGPU imwe muzvikamu zvinomwe. Izvi zvinokutendera iwe kuti uwedzere zvakanyanya kugadzirisa kuchinjika mune imwechete gore segment. Semuenzaniso, imwe DGX A100 system ine masere emuviri maGPU anogona kuita se56 chaiwo maGPU. Multi-Instance GPU (MIG) tekinoroji inokutendera iwe kuti usarudze zvikamu zvehukuru hwakasiyana pakati pemakomputa cores uye sechikamu checache memory uye HBM2 ndangariro, uye ivo havazokwikwidzana kune bandwidth.
Zvakakosha kucherechedza kuti kana tichienzanisa neyakapfuura DGX masisitimu, iyo anatomy yeDGX A100 yakachinja. Huwandu hwemapombi ekupisa mumaradiator eSXM3 modules, iyo A100 graphics processors ine HBM2 memory inoiswa, yakawedzera zvakanyanya kana ichienzaniswa neTesla V100 modules yechizvarwa cheVolta, kunyangwe migumo yavo yakavanzika kubva pakuona kweavhareji munhu. nemavharo epamusoro. Muganho unoshanda weiyo dhizaini ndeye 400 W yesimba rinopisa. Izvi zvinosimbiswawo nehunhu hwepamutemo hweA100 muSXM3 vhezheni, yakaburitswa nhasi.
Padivi peA100 GPUs pabhodhi reamai kune matanhatu-yechitatu-chizvarwa NVLink interface switch, iyo pamwe chete inopa nzira mbiri dzekuchinjana data nekumhanya kwe4,8 TB/s. NVIDIA zvakare yakatarisira zvakanyanya kutonhora kwavo, vachitonga neakazara-profile radiators ane kupisa pombi. Imwe neimwe GPU yakagoverwa 12 chiteshi cheNVLink interface; maGPU akavakidzana anogona kuchinjanisa data nekumhanya kwe600 GB/s.
Iyo DGX A100 system zvakare inogara mapfumbamwe Mellanox ConnectX-6 HDR network controllers, inokwanisa kufambisa ruzivo nekumhanya kunosvika 200 Gbit/s. Pakazara, iyo DGX A100 inopa nzira mbiri dzekufambisa data nekumhanya kwe3,6 TB/s. Iyo sisitimu zvakare inoshandisa tekinoroji yeMellanox yakanangana nekuyera kwakaringana kwemakomputa masisitimu ane chivakwa chakadaro. PCI Express 4.0 rutsigiro papuratifomu inotemerwa ne AMD EPYC Roma chizvarwa processors; semhedzisiro, iyi interface haishandiswe kwete chete neA100 graphics accelerators, asiwo neakasimba-state madhiraivha ane NVMe protocol.
Pamusoro peiyo DGX A100, NVIDIA yatanga kupa vadyidzani vayo neHGX A100 mabhodhi, anove mamwe ezvikamu zvemaseva masisitimu anozogadzirwa nevamwe vagadziri vega. Rimwe HGX A100 bhodhi rinogona kugarisa mana kana masere NVIDIA A100 GPUs. Pamusoro pezvo, kune zvayaida, NVIDIA yakatounganidza DGX SuperPOD - sumbu remazana zana nemakumi mana DGX A140 masisitimu, ichipa mashandiro e100 petaflops ane mwero mwero. Iyo kambani yakavimbisa kupa rubatsiro rwemaitiro kune vanobatana vanoda kuvaka akafanana makomputa masumbu zvichienderana neDGX A700. Nenzira, zvakatora NVIDIA isingapfuure mwedzi kuvaka iyo DGX SuperPOD panzvimbo yemwedzi yakati wandei kana kunyange makore akajairwa kumabasa akadaro.
Sekureva kwaNVIDIA, kuburitswa kweDGX A100 kwatotanga nemutengo wemadhora zviuru zana nemakumi mapfumbamwe neshanu pakopi, vanodyidzana nekambani iyi vari kutotora masisitimu aya mumasumbu avo emakore, ecosystem yatovhara nyika makumi maviri nenhanhatu, kusanganisira Vietnam neUAE. Uye zvakare, magiraidhi mhinduro ane Ampere architecture anonyatso kufungidzira kuve chikamu chePerlmutter supercomputer system, yakagadzirwa naCray yeUS Dhipatimendi reMagetsi. Ichange iine NVIDIA Ampere graphics processors padivi pe AMD EPYC Milan chizvarwa chepakati processors ine Zen 199 architecture. Supercomputer nodes yakavakirwa paNVIDIA Ampere ichasvika kune mutengi muhafu yechipiri yegore, kunyangwe makopi ekutanga akatosvika murabhoritari yakasarudzika. dhipatimendi reAmerica.
Source: 3dnews.ru