I-Nvidia yethula i-H100, i-AI GPU yayo eshesha kakhulu okwamanje, kumafu nabathengisi

I-Nvidia’s I-H100 Hopper GPUsezithembisa ukwenza izinguquko i-artificial intelligence (AI) ngesivinini namandla angakaze abonwe, manje atholakala kabanzi kumakhasimende kuzo zonke izinkundla ezahlukahlukene, inkampani imemezele ngoLwesibili engqungqutheleni yayo yaminyaka yonke yonjiniyela.

I-H100 ilandela I-Nvidia A100 Ama-GPU, abesisekelo sesimanjemanje imodeli yolimi olukhulu imizamo yentuthuko. NgokukaNvidia, i-H100 ishesha izikhathi eziyisishagalolunye ngokuqeqeshwa kwe-AI kanye nezikhathi ezingama-30 ngokushesha ukuze kucatshangwe kune-A100.

>>Landela okuqhubekayo kwe-VentureBeat I-Nvidia GTC entwasahlobo ka-2023 ukufakwa <<

I-GPU entsha izuza kokwakhelwe ngaphakathi I-Transformer Injini, ebaluleke kakhulu ekuthuthukisweni kwe i-AI ekhiqizayo amamodeli afana ne-GPT-3. Iphinde ibe nemiyalelo yokuhlela ye-dynamic (DPX), esiza ukusheshisa ukwenziwa kwekhodi.

“Wonke ama-OEM amakhulu anezixazululo zeseva ye-H100 ukusheshisa ukuqeqeshwa kwemodeli yolimi olukhulu, futhi bonke abahlinzeki abakhulu bamafu bebematasa bememezela izimo zabo ze-H100,” kusho u-Ian Buck, i-VP ye-Hyperscale kanye ne-HPC e-Nvidia, ngesikhathi somhlangano nabezindaba. “Sijabule kakhulu ukuthi izinhlelo ze-H100 manje sezikhiqizwa ngokugcwele futhi manje sezitholakala kumakhasimende.”

Uhlu lwe-Hopper lwama-OEMs, amafu

Ama-Hyperscaler nabahlinzeki bamafu bonke benze izimemezelo ekusekeleni i-H100.

UBuck uphawule ukuthi ngesonto eledlule, iMicrosoft imemezele ukubuka kuqala kwayo okuyimfihlo kwezimo ze-H100 Nvidia. Izimo ezintsha zizosiza amandla womabili amamodeli esizukulwane esilandelayo se-OpenAI kanye namamodeli e-Nvidia ukuze anike amandla isigaba esisha sezixazululo ze-AI ezinkulu. Emuva ngoNovemba 2022, iMicrosoft neNvidia bandise ubudlelwano babo ukuze bakhe uhlelo I-AI supercomputer efwini, esikhathini esizayo esizosebenzisa kakhulu i-H100.

UBuck uphinde waphawula ukuthi i-Amazon izomemezela i-Amazon Web Services (AWS) EC2 UltraClusters yezimo ze-p5, ezisekelwe ku-H100. UBuck uthe izimo ze-p5 zingafinyelela ku-20,000 GPUs zisebenzisa ubuchwepheshe be-AWS Elastic Fabric Adapter (EFA).

Ukwengeza, uBuck uthe i-tech giant Meta manje isiqala ukufaka izinhlelo zayo ze-“Grand Teton” H100 ezikhungweni zayo zedatha ukuze kwakhiwe ikhompyutha enkulu elandelayo ye-Meta ye-AI.

I-slide Buck ekhonjiswe phakathi nesithangami nabezindaba yaphawula ukuthi ozakwethu abaningi manje sebezoba bukhoma ne-H100. Phakathi kwabathengisi abasohlwini kubalwa i-Alibaba Cloud, iBaidu Cloud, iCisco, iDell, iFujitsu, iGigabyte, iHewlett Packard Enterprise, iLenovo, iSupermicro neVultr.

Yini eza ngemva kwe-H100? Ukucabanga okwengeziwe

Ama-GPU angasatshalaliswa ukuze aqeqeshe amamodeli amasha kanye nokuqagela.

“Ukuqeqeshwa kuyisinyathelo sokuqala — ukufundisa a inethiwekhi ye-neural ukumodela indlela yokwenza umsebenzi, ukuphendula umbuzo noma ukukhiqiza isithombe,” kusho uBuck. “I-Inference ukuthunyelwa kwalawo mamodeli ekukhiqizeni ezimweni zangempela zokusetshenziswa.”

Ukusiza ukusekela ukuthunyelwa okubanzi kwamakhono okuqonda, i-Nvidia imemezele i-L4 GPU yayo entsha. U-Buck uchaze ukuthi i-L4 iyi-accelerator yendawo yonke yevidiyo ephumelelayo, i-AI kanye nemifanekiso. I-Nvidia isivele inomamukeli wokuqala we-L4: Ifu le-Google. I-Google izobe ihlanganisa i-L4 kunkundla ye-Vertex AI futhi inikeze abasebenzisi bayo ukufinyelela okuqondile ngokusebenzisa izimo zekhompuyutha ezingokoqobo ze-G2.

“Kuyi-slot eyodwa elula, i-Low Profile GPU engangena kunoma iyiphi iseva, iguqule noma iyiphi iseva noma isikhungo sedatha sibe isikhungo sedatha ye-AI,” kusho uBuck. “Le GPU ishesha izikhathi ezingu-120 kuneseva ye-CPU evamile futhi isebenzisa amandla angaphansi ngo-99%.