Kuye kwathiwa iinkampani zetekhnoloji zikhangela ii-GPUs okanye zisendleleni yokuzifumana. Ngo-Epreli, i-CEO ye-Tesla u-Elon Musk uthenge ii-GPU ezili-10,000 kwaye wathi inkampani iya kuqhubeka nokuthenga isixa esikhulu see-GPU kwi-NVIDIA. Kwicala leshishini, abasebenzi be-IT batyhala ngamandla ukuqinisekisa ukuba ii-GPU zihlala zisetyenziselwa ukwandisa imbuyekezo kutyalo-mali. Nangona kunjalo, ezinye iinkampani zinokufumanisa ukuba ngelixa inani le-GPUs lisanda, ukungasebenzi kwe-GPU kuba nzima kakhulu.
Ukuba imbali isifundise nantoni na malunga ne-high-performance computing (HPC), kukuba ukugcinwa kunye nokunxibelelana akufanele kubingelelwe ngeendleko zokugxila kakhulu kwi-computing. Ukuba ugcino alukwazi ukuhambisa idatha kwiiyunithi zekhompyuter, nokuba unezona GPUs zininzi emhlabeni, awuyi kufezekisa ukusebenza kakuhle.
NgokukaMike Matchett, umhlalutyi kwiDatha enkulu yeHlabathi eNcinci, iimodeli ezincinci zinokuqhutywa kwimemori (RAM), zivumela ukugxila ngakumbi ekubalweni. Nangona kunjalo, iimodeli ezinkulu ezifana ne-ChatGPT ezineebhiliyoni zeenodi azikwazi ukugcinwa kwimemori ngenxa yeendleko eziphezulu.
“Awukwazi ukulingana neebhiliyoni zeenodi kwinkumbulo, ke ukugcinwa kubaluleke ngakumbi,” utshilo uMatchet. Ngelishwa, ukugcinwa kwedatha kudla ngokungahoywa ngexesha lenkqubo yokucwangcisa.
Ngokubanzi, kungakhathaliseki ukuba yintoni na imeko yokusetyenziswa, kukho iingongoma ezine eziqhelekileyo kwinkqubo yoqeqesho lomzekelo:
1. Uqeqesho olungumzekelo
2. Isicelo seNtengiso
3. Ukugcinwa kwedatha
4. IComputing ekhawulezileyo
Xa udala kunye nokuthunyelwa kweemodeli, ezininzi iimfuno zibeka phambili ubungqina obukhawulezayo be-concept (POC) okanye iimeko zokuvavanya ukuqalisa uqeqesho lwemodeli, kunye neemfuno zokugcina idatha zinganikwa ingqwalasela ephezulu.
Nangona kunjalo, umceli mngeni ulele kwinto yokuba uqeqesho okanye ukuhanjiswa kwenkcazo kungathatha iinyanga okanye iminyaka. Iinkampani ezininzi zinyusa ngokukhawuleza ubungakanani bazo bemodeli ngeli xesha, kwaye iziseko zophuhliso kufuneka zande ukuze zihambelane neemodeli ezikhulayo kunye neeseti zedatha.
Uphando oluvela kuGoogle kwizigidi zemisebenzi yoqeqesho lwe-ML lubonisa ukuba umyinge we-30% yexesha loqeqesho lichithwa kumbhobho wedatha yegalelo. Ngelixa uphando oludlulileyo lugxile ekwandiseni ii-GPUs ukukhawulezisa uqeqesho, imingeni emininzi isahleli ekuphuculeni iindawo ezahlukeneyo zombhobho wedatha. Xa unamandla amakhulu okubala, eyona ntsini yenyani iba yindlela onokukhawuleza ngayo ukondla idatha kwizibalo ukuze ufumane iziphumo.
Ngokukodwa, imiceli mngeni ekugcinweni kwedatha kunye nolawulo ifuna ukucwangciswa kokukhula kwedatha, ikuvumela ukuba uqhubeke ukhupha ixabiso ledatha njengoko uqhubela phambili, ngakumbi xa ungena kwiimeko zokusetyenziswa ezikumgangatho ophezulu ezifana nokufunda okunzulu kunye neenethiwekhi ze-neural, ezibeka iimfuno eziphezulu ugcino ngokomthamo, ukusebenza, kunye nokwaleka.
Ukuthi ngqo:
Ukubaleka
Ukufunda ngoomatshini kufuna ukuphatha inani elikhulu ledatha, kwaye njengoko umthamo wedatha usanda, ukuchaneka kweemodeli kukwaphucula. Oku kuthetha ukuba amashishini kufuneka aqokelele kwaye agcine idatha eninzi yonke imihla. Xa ugcino lungakwazi ukulinganisa, imithwalo enzima yedatha idala imiqobo, inciphisa ukusebenza kwaye ibangele ixesha elibizileyo le-GPU.
Ukuba bhetyebhetye
Inkxaso eguquguqukayo kwiiprotokholi ezininzi (kubandakanywa i-NFS, i-SMB, i-HTTP, i-FTP, i-HDFS, kunye ne-S3) iyimfuneko ukuhlangabezana neemfuno zeenkqubo ezahlukeneyo, kunokuba zikhawulelwe kuhlobo olulodwa lokusingqongileyo.
Ukubambezeleka
I/O latency ibalulekile ekwakheni nasekusebenziseni iimodeli njengoko idatha ifundwa kwaye iphinda ifundwe izihlandlo ezininzi. Ukunciphisa i-I / O latency kunokunciphisa ixesha loqeqesho lweemodeli ngeentsuku okanye iinyanga. Uphuhliso lwemodeli olukhawulezayo luguqulela ngokuthe ngqo kwiinzuzo ezinkulu zeshishini.
Umsebenzi
Ukusebenza kweenkqubo zokugcina kubalulekile kuqeqesho olusebenzayo lwemodeli. Iinkqubo zoqeqesho zibandakanya inani elikhulu ledatha, ngokuqhelekileyo kwiiterabytes ngeyure.
Ufikelelo olunxuseneyo
Ukufezekisa iziphumo eziphezulu, iimodeli zoqeqesho zahlula imisebenzi ibe yimisebenzi emininzi ehambelanayo. Oku kuhlala kuthetha ukuba ii-algorithms zokufunda koomatshini zifikelela kwiifayile ezifanayo kwiinkqubo ezininzi (ezinokwenzeka kwiiseva zomzimba ezininzi) ngaxeshanye. Inkqubo yokugcina kufuneka iphathe iimfuno ezifanayo ngaphandle kokunciphisa ukusebenza.
Ngobuchule bayo obubalaseleyo kwi-latency ephantsi, i-output ephezulu, kunye ne-I/O enkulu ehambelanayo, iDell PowerScale yeyona ndawo ifanelekileyo yokugcina ihambelana nekhompyuter ye-GPU. I-PowerScale inciphisa ngokufanelekileyo ixesha elifunekayo kwiimodeli zokuhlalutya eziqeqesha kunye nokuvavanya i-multi-terabyte datasets. Kwi-PowerScale yonke i-flash yokugcina, i-bandwidth inyuka ngamaxesha e-18, isusa i-I / O i-bottlenecks, kwaye inokongezwa kumaqela akhoyo e-Isilon ukukhawuleza nokuvula ixabiso leedatha ezingacwangciswanga.
Ngaphezu koko, i-PowerScale's multi-protocol access accessabilities ibonelela ngokuguquguquka okungenamkhawulo wokuqhuba umsebenzi, ukuvumela idatha ukuba igcinwe kusetyenziswa iprotocol enye kwaye ifikeleleke kusetyenziswa enye. Ngokukodwa, iimpawu ezinamandla, ukuguquguquka, ukulinganisa, kunye nokusebenza komgangatho weshishini weqonga lePowerScale linceda ukujongana nale mingeni ilandelayo:
- Ukukhawulezisa ukutsha ukuya kuthi ga kumaxesha angama-2.7, ukunciphisa umjikelo woqeqesho lwemodeli.
- Ukuphelisa i-bottlenecks ye-I / O kunye nokubonelela ngoqeqesho olukhawulezayo lwemodeli kunye nokuqinisekiswa, ukuchaneka kwemodeli ephuculweyo, ukuphuculwa kwemveliso yesayensi yedatha, kunye nembuyekezo eyandisiweyo kutyalo-mali lwekhompuyutha ngokusetyenziswa kweempawu zebakala loshishino, ukusebenza okuphezulu, ukuhambelana, kunye nokuncipha. Ukuphucula ukuchaneka kwemodeli kunye nedatha enzulu, enesisombululo esiphezulu ngokufikelela kwi-119 PB yendawo yokugcina esebenzayo kwiqela elinye.
- Ukufezekisa ukusasazwa kwinqanaba ngokuqala amancinci kunye nokuzimela ukulinganisa ikhompyutha kunye nokugcinwa, ukuhambisa ukhuseleko lwedatha olomeleleyo kunye neendlela zokhuseleko.
- Ukuphucula imveliso yesayensi yedatha ngokuhlalutya kwendawo kunye nezisombululo eziqinisekisiweyo zangaphambili, ngokukhawuleza, ukuthunyelwa komngcipheko ophantsi.
-Ukusebenzisa uyilo oluqinisekisiweyo olusekwe kweyona teknoloji ibalaseleyo, kubandakanya i-NVIDIA GPU yokukhawulezisa kunye nereferensi yezakhiwo ezineenkqubo ze-NVIDIA DGX. Ukusebenza okuphezulu kunye nokuhambelana kwe-PowerScale kuyahlangabezana neemfuno zokusebenza kogcino kwinqanaba ngalinye lokufunda koomatshini, ukusuka ekufumaneni idatha kunye nolungiselelo ukuya kwimodeli yoqeqesho kunye nokuthelekelela. Kunye nenkqubo yokusebenza ye-OneFS, zonke ii-nodes zinokusebenza ngokungenamthungo ngaphakathi kweqela elifanayo eliqhutywa yi-OneFS, elineempawu zenqanaba leshishini ezifana nolawulo lwentsebenzo, ulawulo lwedatha, ukhuseleko, kunye nokukhuselwa kwedatha, okwenza ukuba kugqitywe ngokukhawuleza koqeqesho lwemodeli nokuqinisekiswa kwamashishini.
Ixesha lokuposa: Jul-03-2023
