Riaan I-GTC Ngaphezulu, Ingqungquthela ye-GPU Technology yaseNvidia yasentwasahlobo iphinde yavulwa ngenothi eyisihluthulelo ebuswa ubuchwepheshe be-AI obukhiqizayo.
Lokho akufanele kusimangaze. Naphezu kwenqubekelaphambili eyenziwe ukuthandwa kwe-AMD, Intel, nabanye, i-Nvidia isalokhu ingumhlinzeki omkhulu wama-GPU nama-accelerator asetshenziselwa ukulayisha umsebenzi wokufunda ngomshini.
Kepha ngemuva kwakho konke ubukhazikhazi obunikezwe amamodeli we-AI wakamuva we-Nvidia, izinhlaka zokusheshisa, nehadiwe – okunokuningi esingakhuluma ngakho – kucashe udaba lokusimama. Iphuzu elavela kaningi ngesikhathi senkulumo eyisihluthulelo ye-GTC ye-CEO uJensen Huang – bheka ngezansi – lapho agcizelela khona ukuthi kungasebenzi futhi kubiza kangakanani ukwenza le mithwalo yomsebenzi kumaseva enhloso evamile.
Kukhona ukufaneleka kwengxabano yakhe. Nakuba kungewona wonke umthwalo ofaniswa kalula, lawo angathambekela ekusebenzeni kahle kakhulu. Kunesizathu esenza ukuthi ama-supercomputer asheshiswe yi-GPU alawule Okuhlaza 500.
“I-Cloud computing ikhule ngamaphesenti angu-20 ngonyaka ibe yimboni enkulu ye-$ 1 trillion. Amanye amaseva we-CPU ayizigidi ezingu-30 enza iningi lokucubungula,” kusho u-Huang. “Njengoba ‘uMthetho kaMoore’ uphela, ukusebenza kwe-CPU okwandayo kuza namandla angeziwe, futhi igunya lokunciphisa ukukhishwa kwekhabhoni lingqubuzana nesidingo sokwandisa ama-datacenters.”
Uthi ukusheshisa ukuphela kwendlela yokunciphisa ukusetshenziswa kwamandla. Kodwa akukhona ukuthi uNvidia ufuna nje ukukudayisela ama-GPU amaningi ngesikhathi lapho isidingo siphansi kuzo zonke izingxenye eziningi, inkampani inephothifoliyo ephelele yama-accelerator nesofthiwe ongayisebenzisa nabo efuna ukukuthengisa nawe. I-Nvidia inama-GPU anikezelwe ekuqeqeshweni kwe-AI, amanye aklanyelwe ukuhlehlisa, ukubuka ngamehlo engqondo, nokucutshungulwa kwevidiyo, kanye nekhithi yokuxhumana nokucubungula idatha ukuze kuhlanganiswe konke ndawonye.
“Ama-datacenter kumele asheshise wonke umthwalo womsebenzi ukuze athole amandla kabusha kanye nama-GPU amahhala okukhiqiza imali engenayo kanye nokulayishwa kwe-Nvidia’s Bluefield futhi asheshise uhlelo lokusebenza lwe-datacenter kanye nesofthiwe yengqalasizinda,” kusho u-Huang lapho ekhuluma ne-Bluefield-3 DPUs yenkampani.
Futhi ngemithwalo yomsebenzi engakwazi ukusheshiswa kusetshenziswa ama-GPU noma ama-DPU, i-Nvidia ine-Grace CPU yayo esekelwe engalweni. “Yonke i-144-core Grace Superchip inamandla aphansi kangangokuthi ikwazi ukupholiswa emoyeni,” kusho u-Huang, ephakamisa i-heatsink ye-chip engu-5×8-inch 1U.
Ngokwerekhodi, iGrace Superchip, ehlanganisa i-1TB yememori ye-LPDDR, isenamandla omklamo oshisayo (TDP) alinganiselwe angama-500W, ngakho-ke akuyona ingxenye yamandla ephansi kakhulu ngaphandle lapho. Ukuze uqhathanise, i-AMD’s 96-core/192-thread Epyc 4 ine-TDP elungisekayo engu-360W-400W. Ukusebenza emisebenzini ehlukene, hhayi i-TDP, ngakho-ke kuzonquma ukuncintisana kokusebenza kahle phakathi kwalokhu kokubili ngokwezinga elikhulu kune-TDP iyodwa.
Ngemva kokwakha i-Hardware, kuyacaca ukuthi isinyathelo esilandelayo sikaNvidia ukulandela futhi akhe imitapo yolwazi yesofthiwe ukusheshisa umsebenzi kuyo yonke imboni ethembele kuma-CPU namuhla – empeleni kwakha izimakethe ezintsha zehadiwe yayo. Imitapo yolwazi emisha ye-computational lithography yenkampani yile ubufakazi kulokho nje, u-Huang ethi i-TSMC ingashintsha ama-CPU node ayo angu-40,000 ngamaseva ambalwa afinyelela ku-500 DGX H100 ngenkathi inqamula ukusetshenziswa kwamandla ukusuka ku-35MW kuya ku-5MW kule nqubo.
Inkundla igcwele kakhulu kunakuqala
Uma konke lokhu kuzwakala kujwayelekile, lokho kungenxa yokuthi i-CEO ye-AMD uLisa SU ishaye kakhulu amaphuzu afanayo ngenkathi ekhuluma eNgqungqutheleni Yamazwe Ngamazwe Yezifunda Zezifunda Eziqinile ngasekupheleni kukaFebhuwari.
Ngesikhathi senkulumo yakhe eyisihluthulelo u-Su uxwayise ngokuthi ngaphandle uma kuthathwa izinyathelo ezinqala phakathi neshumi leminyaka ukuthuthukisa ukusebenza kahle kwezakhiwo zekhompiyutha, amakhompiyutha anamandla kakhulu emhlabeni ngeke nje alingise ukusabela kwenuzi, kuzodingeka asebenze kuwo.
Impendulo ye-AMD kulokhu ibandakanya izinhlobonhlobo zezakhiwo zekhompiyutha kanye nezimiso zokuklama, okuhlanganisa ezimbalwa, njengama-chiplets, eyaphayona ngazo. Ukusebenzisa lobu buchwepheshe i-AMD isebenzela ukushumeka ama-accelerator azo zonke izinhlobo kumapulatifomu ayo. Isibonelo, i-MI300 APU ezayo yenkampani ihlanganisa ama-Zen 4 CPU cores nama-CDNA3 GPU ayo kanye nomthwalo wesikebhe wenkumbulo yomkhawulokudonsa ophezulu.
Oqhudelanayo we-x86 uphinde asebenzele ukwakha ama-FPGA, ama-accelerator e-AI, nama-DPU, atholwe kwa-Xilinx nasePensando, abe kuma-chip awo ukuze kusheshiswe imithwalo eminingi yomsebenzi futhi kuthuthukiswe ukusebenza kahle kwekhompyutha.
Futhi hhayi ukushiywa ngaphandle, ingxenye ekhulayo ye-Intel’s 4th Gen Xeon Ukufa sekuvele kusetshenziswe ama-accelerator azinikele ekufundeni komshini, i-cryptography, ukucindezela, ukusakaza idatha, ukuhlaziya, nokuphepha.
I-Intel nayo isebenza ku-APU yayo – yize ithanda igama elithi XPU – elibizwa nge-Falcon Shores, yize lelo pulatifomu ibambezelekile kuze kube okungenani ngo-2025 kulokho osekuphenduke umkhuba womenzi we-chip ogulayo.
Ngakho-ke i-Nvidia ikude nokuphela kwe-chipmaker ekhuthaza ukusetshenziswa okwengeziwe kwama-accelerator. ®
