NVIDIA DGX A100: sethala se thehiloeng ho Ampere se fana ka li-petaflops tse hlano tsa ts'ebetso

Sistimi ea DGX A100, e thehiloeng ho eona Jen-Hsun Huang haufinyane a e ntša ka ontong, e kenyelletsa li-GPU tse robeli tsa A100, li-switch tse tšeletseng tsa NVLink 3.0, balaoli ba marang-rang ba Mellanox ba robong, li-processor tse peli tsa AMD EPYC Roma tse nang le li-cores tse 64, 1 TB ea RAM le 15 TB ea SSD e tšehetsoeng ke NVMe.

NVIDIA DGX A100: sethala se thehiloeng ho Ampere se fana ka li-petaflops tse hlano tsa ts'ebetso

NVIDIA DGX A100 ke moloko oa boraro oa sistimi ea komporo ea k'hamphani, e etselitsoeng haholo ho rarolla mathata a bohlale ba maiketsetso. Hona joale litsamaiso tse joalo li hahiloe holim'a li-processor tsa morao-rao tsa A100 tse tsoang lelapeng la Ampere, tse bakang keketseho e matla ea ts'ebetso ea bona, e fihlileng ho 5 petaflops. Ka lebaka la sena, DGX A100 e khona ho sebetsana le mefuta e mengata ea AI e rarahaneng le palo e kholoanyane ea data.

Bakeng sa sistimi ea DGX A100, NVIDIA e bonts'a feela kakaretso ea memori ea HBM2, e fihlang ho 320 GB. Lipalo tse bonolo tsa lipalo li re lumella ho fumana hore na GPU e 'ngoe le e' ngoe e na le 40 GB ea mohopolo, 'me litšoantšo tsa sehlahisoa se secha li hlakisa hore molumo ona o abuoa har'a mekotla e tšeletseng. Grafiken memory bandwidth e boetse e boleloa - 12,4 TB / s bakeng sa sistimi eohle ea DGX A100 ka kakaretso.

Ha re nahana hore sistimi ea DGX-1, e ipapisitseng le li-Tesla V100s tse robeli, e hlahisitse petaflops e le 'ngoe ka lipalo-palo tse tsoakiloeng, mme ho boleloa hore DGX A100 e sebetsa ho li-petaflops tse hlano, re ka nka hore lipalong tse ikhethileng Ampere GPU e le 'ngoe e lebelo ho feta makhetlo a mahlano. pele ho eona ka meralo ea Volta. Maemong a mang, molemo o fetoha habeli.

NVIDIA DGX A100: sethala se thehiloeng ho Ampere se fana ka li-petaflops tse hlano tsa ts'ebetso

Ka kakaretso, sistimi ea DGX A8 e fana ka ts'ebetso e phahameng ea ts'ebetso ea 100 motsotsoana ts'ebetsong e felletseng (INT1016), ts'ebetsong ea "half-precision floating point" (FP16) - 5 petaflops, ts'ebetsong ea lintlha tse phaphametseng tse habeli (FP64) - 156 teraflops. . Ho feta moo, DGX A32 e fihlella ts'ebetso e phahameng ea 100 petaflops ho komporo ea tensor ea TF2,5. Ha re hopoleng hore teraflops e le 'ngoe ke ts'ebetso ea lintlha tse 1012 ka motsotsoana, petaflops e le' ngoe ke ts'ebetso ea lintlha tse 1015 ka motsotsoana.

Karolo ea bohlokoa ea li-accelerator tsa NVIDIA A100 ke bokhoni ba ho arola lisebelisoa tsa GPU e le 'ngoe ka likarolo tse supileng tsa sebele. Sena se o lumella ho eketsa haholo ho feto-fetoha ha maemo karolong e le 'ngoe ea maru. Mohlala, sistimi e le 'ngoe ea DGX A100 e nang le li-GPU tse robeli tsa 'mele e ka sebetsa joalo ka li-GPU tse 56. Theknoloji ea Multi-Instance GPU (MIG) e u lumella ho khetha likarolo tsa boholo bo fapaneng har'a li-cores tsa khomphutha hape e le karolo ea memori ea cache le memori ea HBM2, 'me li ke ke tsa qothisana lehlokoa le tse ling bakeng sa bandwidth.

NVIDIA DGX A100: sethala se thehiloeng ho Ampere se fana ka li-petaflops tse hlano tsa ts'ebetso

Ke habohlokoa ho hlokomela hore ha ho bapisoa le mekhoa e fetileng ea DGX, anatomy ea DGX A100 e bile le liphetoho tse itseng. Palo ea liphaephe tsa mocheso ho li-radiator tsa li-module tsa SXM3, tseo ho tsona li-processor tsa A100 tse nang le mohopolo oa HBM2 li kentsoeng, li eketsehile haholo ha li bapisoa le li-module tsa Tesla V100 tsa moloko oa Volta, le hoja lipheletso tsa tsona li patiloe ponong ea motho ea tloaelehileng. ka likoahelo tse ka holimo. Moeli o sebetsang oa moralo ona ke 400 W ea matla a mocheso. Sena se boetse se tiisitsoe ke litšobotsi tsa molao tsa A100 phetolelong ea SXM3, e hatisitsoeng kajeno.

Haufi le li-GPU tsa A100 tse leboteng la bo-mme ho na le li-switches tse tšeletseng tsa moloko oa boraro oa NVLink, tseo hammoho li fanang ka phapanyetsano ea data ea litsela tse peli ka lebelo la 4,8 TB/s. NVIDIA e boetse e hlokometse ka botebo ho pholiso ea bona, ho ahlola ka li-radiator tse felletseng tse nang le liphaephe tsa mocheso. GPU e 'ngoe le e' ngoe e abetsoe likanale tse 12 tsa sebopeho sa NVLink; li-GPU tsa boahelani li ka fapanyetsana data ka lebelo la 600 GB/s.

Sistimi ea DGX A100 e boetse e na le balaoli ba marang-rang ba Mellanox ConnectX-6 HDR ba robong, ba khonang ho fetisa tlhahisoleseling ka lebelo le fihlang ho 200 Gbit/s. Ka kakaretso, DGX A100 e fana ka phetiso ea lintlha tse peli ka lebelo la 3,6 TB / s. Sistimi e boetse e sebelisa mahlale a Mellanox a nang le sepheo sa ho holisa lits'ebetso tsa komporo ka meaho e joalo. Ts'ehetso ea PCI Express 4.0 sethaleng e khethoa ke li-processor tsa AMD EPYC Roma; ka lebaka leo, sebopeho sena ha se sebelisoe feela ke li-accelerator tsa litšoantšo tsa A100, empa hape le li-drive tse tiileng tse nang le protocol ea NVMe.

NVIDIA DGX A100: sethala se thehiloeng ho Ampere se fana ka li-petaflops tse hlano tsa ts'ebetso

Ho phaella ho DGX A100, NVIDIA e se e qalile ho fana ka balekane ba eona ka liboto tsa HGX A100, e leng e 'ngoe ea likarolo tsa lisebelisoa tsa seva tseo bahlahisi ba bang ba tla li hlahisa ka bobona. Boto e le 'ngoe ea HGX A100 e ka amohela li-GPU tse 'ne kapa tse robeli tsa NVIDIA A100. Ntle le moo, bakeng sa litlhoko tsa eona, NVIDIA e se e bokelletse DGX SuperPOD - sehlopha sa litsamaiso tse 140 tsa DGX A100, tse fanang ka ts'ebetso ho li-petaflops tse 700 tse nang le litekanyo tse itekanetseng ka kakaretso. Khamphani e tšepisitse ho fana ka thuso ea mekhoa ho balekane ba lakatsang ho theha lihlopha tse tšoanang tsa komporo ho latela DGX A100. Ka tsela, ho ile ha nka NVIDIA nako e fetang khoeli ho haha ​​​​DGX SuperPOD sebakeng sa likhoeli tse 'maloa kapa lilemo tse tloaelehileng bakeng sa mesebetsi e joalo.

NVIDIA DGX A100: sethala se thehiloeng ho Ampere se fana ka li-petaflops tse hlano tsa ts'ebetso

Ho ea ka NVIDIA, ho tsamaisoa ha DGX A100 ho se ho qalile ka theko ea $ 199 ka kopi, balekane ba k'hamphani ba se ba ntse ba tšoara litsamaiso tsena ka lihlopha tsa bona tsa maru, ecosystem e se e akaretsa linaha tse 000, ho kenyeletsoa Vietnam le UAE. Ntle le moo, tharollo ea lits'oants'o tse nang le boqapi ba Ampere e tla ba karolo ea sistimi e kholo ea Perlmutter, e entsoeng ke Cray bakeng sa Lefapha la Matla la US. E tla ba le li-processor tsa graphics tsa NVIDIA Ampere hammoho le li-processor tse bohareng tsa AMD EPYC Milan tse nang le meralo ea Zen 26. Li-node tsa Supercomputer tse thehiloeng ho NVIDIA Ampere li tla fihla ho moreki halofong ea bobeli ea selemo, le hoja likopi tsa pele li se li fihlile laboratoring e khethehileng. lefapha la Amerika.



Source: 3dnews.ru

Eketsa ka tlhaloso