Usoro DGX A100, dabere na Jen-Hsun Huang na nso nso a , gụnyere asatọ A100 GPUs, isii NVLink 3.0 switches, itoolu Mellanox netwọk njikwa, abụọ AMD EPYC Rome-generation processors na 64 cores, 1 TB nke RAM na 15 TB nke SSD na nkwado NVMe.
NVIDIA DGX A100 bụ ọgbọ nke atọ nke sistemu kọmpụta ụlọ ọrụ, emebere ya maka idozi nsogbu ọgụgụ isi. Ugbu a, a na-ewu usoro ndị dị otú ahụ na ndị nrụpụta eserese A100 kachasị ọhụrụ sitere na ezinụlọ Ampere, nke na-akpata mmụba dị ukwuu na arụmọrụ ha, nke ruru 5 petaflops. N'ihi nke a, DGX A100 nwere ike ijikwa ụdị AI dị mgbagwoju anya yana nnukwu data buru ibu.
Maka sistemụ DGX A100, NVIDIA na-egosi naanị mkpokọta nchekwa HBM2, nke ruru 320 GB. Mgbakọ mgbakọ na mwepụ dị mfe na-enye anyị ohere ikpebi na GPU ọ bụla nwere 40 GB nke ebe nchekwa, na ihe oyiyi nke ngwaahịa ọhụrụ ahụ na-eme ka o doo anya na a na-ekesa olu a n'etiti nchịkọta isii. Ekwurukwa bandwidth ebe nchekwa eserese - 12,4 TB / s maka sistemụ DGX A100 niile.
N'iburu n'uche na usoro DGX-1, dabere na asatọ Tesla V100s, mepụtara otu petaflops na ngụkọ nke ọma, na DGX A100 na-ekwu na ọ na-arụ na petaflops ise, anyị nwere ike iche na na ngụkọ akọwapụtara otu Ampere GPU dị ngwa ngwa karịa ugboro ise. nke bu ya ụzọ na Volta architecture. N'ọnọdụ ụfọdụ, uru na-aghọ okpukpu iri abụọ.
Na mkpokọta, usoro DGX A8 na-enye arụmọrụ kachasị elu nke arụmọrụ 100 kwa nke abụọ na arụmọrụ integer (INT1016), na arụmọrụ ọkara nke ziri ezi na-ese n'elu mmiri (FP16) - 5 petaflops, na arụmọrụ abụọ na-ese n'elu mmiri (FP64) - 156 teraflops. . Na mgbakwunye, DGX A32 na-arụ ọrụ kacha elu nke 100 petaflops na TF2,5 tensor computing. Ka anyị cheta na otu teraflops bụ ọrụ 1012 na-ese n'elu mmiri kwa nkeji, otu petaflops bụ ọrụ 1015 floating point kwa nkeji.
Akụkụ dị mkpa nke NVIDIA A100 accelerators bụ ike kewaa akụrụngwa nke otu GPU n'ime akụkụ asaa mebere. Nke a na-enye gị ohere ịbawanye mgbanwe nhazi n'otu akụkụ igwe ojii. Dịka ọmụmaatụ, otu sistemụ DGX A100 nwere GPU anụ ahụ asatọ nwere ike ịrụ ọrụ dị ka GPU mebere 56. Teknụzụ Multi-Instance GPU (MIG) na-enye gị ohere ịhọrọ akụkụ nke nha dị iche iche ma n'etiti kọmpụta kọmpụta yana dịka akụkụ nke ebe nchekwa cache na ebe nchekwa HBM2, na ha agaghị asọmpi ibe ha maka bandwit.
Ọ dị mma ịmara na atụnyere sistemụ DGX gara aga, anatomi nke DGX A100 enweela mgbanwe ụfọdụ. Ọnụ ọgụgụ nke ọkpọkọ okpomoku na radiators nke modulu SXM3, nke etinyere A100 ndị na-eme ihe nkiri na ebe nchekwa HBM2, abawanyela nke ukwuu ma e jiri ya tụnyere modul Tesla V100 nke ọgbọ Volta, n'agbanyeghị na njedebe ha na-ezo n'anya onye nkịtị. site na mkpuchi elu. Oke bara uru maka imewe a bụ 400 W nke ike ọkụ. E gosikwara nke a site na njirimara gọọmentị nke A100 na ụdị SXM3, nke e bipụtara taa.
Na-esote A100 GPUs na motherboard bụ ọgbọ atọ nke atọ NVLink interface switches, nke jikọtara ọnụ na-enye mgbanwe data ụzọ abụọ na ọsọ nke 4,8 TB / s. NVIDIA lezikwara anya nke ọma maka ịdị jụụ ha, na-ekpe ikpe site na radiators zuru oke na ọkpọkọ okpomọkụ. A na-ekenye GPU ọ bụla ọwa 12 nke interface NVLink; GPU ndị agbata obi nwere ike ịgbanwe data na ọsọ 600 GB/s.
Sistemu DGX A100 nwekwara ụlọ ndị na-ahụ maka netwọkụ Mellanox ConnectX-6 HDR itoolu, nwere ike ibunye ozi na ọsọ ọsọ ruru 200 Gbit/s. Na mkpokọta, DGX A100 na-enye nnyefe data ụzọ abụọ na ọsọ nke 3,6 TB / s. Sistemu ahụ na-ejikwa teknụzụ Mellanox nke nwe ya nke ezubere iji mee ka sistemụ kọmpụta rụọ ọrụ nke ọma na ụdị ụlọ a. Nkwado PCI Express 4.0 na ọkwa ikpo okwu ka AMD EPYC Rome na-arụ ọrụ ọgbọ; n'ihi ya, ọ bụghị naanị A100 ndị na-eme ihe ngosi eserese na-eji interface a, kamakwa site na draịva steeti siri ike na usoro NVMe.
Na mgbakwunye na DGX A100, NVIDIA amalitela ịnye ndị mmekọ ya na bọọdụ HGX A100, nke bụ otu n'ime akụkụ nke sistemu nkesa nke ndị nrụpụta ndị ọzọ ga-emepụta n'onwe ha. Otu bọọdụ HGX A100 nwere ike ịnabata NVIDIA A100 GPU anọ ma ọ bụ asatọ. Na mgbakwunye, maka mkpa nke ya, NVIDIA achịkọtalarị DGX SuperPOD - ụyọkọ nke sistemu 140 DGX A100, na-enye arụmọrụ na 700 petaflops nwere oke nha anya dị mma. Ụlọ ọrụ ahụ kwere nkwa inye ndị mmekọ aka na-achọ ịmepụta ụyọkọ kọmpụta yiri nke ahụ dabere na DGX A100. Site n'ụzọ, ọ were NVIDIA ihe karịrị otu ọnwa iji wuo DGX SuperPOD kama ọtụtụ ọnwa ma ọ bụ ọbụna ọtụtụ afọ maka ọrụ ndị dị otú ahụ.
Dị ka NVIDIA si kwuo, nnyefe nke DGX A100 amalitelarị na ọnụahịa nke $ 199 kwa otu, ndị mmekọ ụlọ ọrụ ahụ na-akwadorịrị usoro ndị a na ụyọkọ igwe ojii ha, gburugburu ebe obibi ekpuchilarị mba 000, gụnyere Vietnam na UAE. Na mgbakwunye, ihe ngwọta eserese nwere ihe owuwu Ampere ga-abụ nke a ga-ebu amụma na ọ bụ akụkụ nke sistemụ kọmpụta Perlmutter, nke Cray maka Ngalaba Energy US mepụtara. Ọ ga-agụnye NVIDIA Ampere graphics processors n'akụkụ AMD EPYC Milan ọgbọ Central processors na Zen 26 architecture. Supercomputer nodes dabere na NVIDIA Ampere ga-eru ndị ahịa na ọkara nke abụọ nke afọ, n'agbanyeghị na mbụ mbipụta abanyela na specialized laabu nke ngalaba America.
isi: 3dnews.ru