Ama-Tensor Cores
I-Tensor Cores amayunithi ehadiwe akhethekile ngaphakathi kwe-NVIDIA GPU yesimanje enza imisebenzi ye-matrix yokuphindaphinda nokuqongelela ngokushesha okukhulu.
Uhlolojikelele
They are the main reason a single GPU can train and run large neural networks orders of magnitude faster than general-purpose compute would allow.
I-Deep Dive
Eyethulwe ngesakhiwo se-Volta ngo-2017, i-Tensor Cores iyisekhethi ezinikezele ehlanganisa ukuphindaphinda okuncane kwe-matrix kanye nokwengezwa (D = A x B + C) ekusebenzeni okukodwa, kunokwenza ukuphindaphinda ngakunye ngesikhathi kuma-cores we-CUDA ajwayelekile. Ngoba cishe yonke ingqimba yenethiwekhi ye-neural yehlisela ekuphindaphindeni kwe-matrix, lokhu kufana nezibalo ezidingwa yi-AI. Isizukulwane ngasinye se-GPU sandise esikuphathayo: I-Volta yenza amathayela angu-4x4 FP16, kwathi kamuva izakhiwo ze-Ampere, Hopper, ne-Blackwell zengeza amafomethi anemba eliphansi njenge-TF32, BF16, INT8, FP8, ne-FP4. Ukunemba okuphansi kusho izinombolo ezengeziwe ezicutshungulwa iwashi ngalinye, okukhuphula ngokumangalisayo umphumela wokuqeqeshwa nokuchazwa kuyilapho kugcinwa ukunemba kwamukelekile.
I-Technical Insight
I-Tensor Core iphindaphinda omatikuletsheni ababili abancane futhi iqongelela umphumela ngesinyathelo esisodwa esihlanganisiwe, isebenzisa iqiniso lokuthi amanani afanayo okokufaka aphinde asetshenziswe ezintweni eziningi zokukhiphayo. Ngokuvamile ifunda okokufaka ngokunemba okuncishisiwe (FP16, BF16, noma FP8) kodwa iqongelela isamba esisebenzayo ngokunemba okuphezulu (ngokuvamile i-FP32) ukuze ikhawulele iphutha lokusondeza. Imitapo yolwazi yesofthiwe efana ne-cuBLAS ne-cuDNN, nezinhlaka ezifana ne-PyTorch, amathayela omatikuletsheni abakhulu kulawa mabhulokhi amancane ngokuzenzakalelayo ukuze amamodeli athole ukusheshisa ngaphandle kokufaka amakhodi okwenziwa ngesandla.
I-Strategic Impact
Izindleko kanye nesabelomali
Izinqumo zezakhiwo ziqhuba ukusebenza kanye nezindleko zokusebenza iminyaka.
Izinqumo ezicacile
Imfundo yobuchwepheshe isiza amaqembu ukuthi akhethe isitaki esifanele, hhayi nje esisha.
Ukulawulwa kwekhwalithi
Izinketho ezingcono zobunjiniyela zinciphisa izehlakalo ezinokwethenjelwa ekukhiqizeni.
Ikusasa le-Tensor Cores
I-Tensor Cores iqhubeka nokunemba okuhlala kuphansi: U-Hopper wengeze i-FP8 futhi uBlackwell wethula i-4-bit FP4 ene-Hardware ephethwe isikali, cishe okuphindwe kabili okuphumayo esinyathelweni ngasinye ukuze kucatshangelwe umthwalo osindayo. Lindela ukusekelwa okuqinile kwe-sparsity (ukweqa izisindo eziyiziro), amafomethi e-microscaling anamathisela izici zesikali kumabhulokhi amancane ezinombolo, nokuhlanganiswa okujulile nezinhlelo zememori ukuze ama-cores ahlale ondlekile. Njengoba amamodeli ekhula, injini ye-matrix, hhayi isivinini sewashi eluhlaza, ihlala iyindawo yempi emaphakathi yokusebenza kwezingxenyekazi zekhompiyutha ze-AI.
Ukuqaliswa Komhlaba Wangempela
Ukuqeqesha amamodeli amakhulu olimi afana neziguquli zesitayela se-GPT, lapho izigidigidi zokuphindaphinda kwe-matrix isinyathelo ngasinye zisebenza ku-Tensor Cores ku-BF16 noma ku-FP8.
Isebenzisa ukulinganisa kwesikhathi sangempela kuma-chatbots namajeneretha ezithombe, kusetshenziswa ukulinganisa kwe-INT8 noma i-FP8 ukuze kunikezwe abasebenzisi abengeziwe nge-GPU ngayinye.
Ukusheshisa i-NVIDIA DLSS kumageyimu evidiyo, lapho inethiwekhi ye-neural inyusa izinga lozimele abanokulungiswa okuphansi isebenzisa i-Tensor Cores ifreyimu ngayinye.
Ukusheshisa ikhompuyutha yesayensi efana ne-protein-folding (AlphaFold) namamodeli wesimo sezulu ahlelwe kabusha njengemithwalo ye-matrix-heavy neural workload.
Izingozi & Guardrails
Ukuthuthukisa ibhentshimakhi eyodwa kungafihla ubuthakathaka obubanzi besistimu.
Izindleko zengqalasizinda nezokulungisa zivame ukubukelwa phansi.
Izikhala zokuphepha nokubonakala zingakhula njengoba izinhlelo ziba nzima kakhulu.
Ukuqalisa Umhlahlandlela
Chaza ukubambezeleka, ikhwalithi, nezindleko ezihlosiwe ngaphambi kokuqaliswa.
Ibhentshimakhi ngaphansi komthwalo wangempela nezimo zedatha.
Ukuqapha amathuluzi amaphutha, ukukhukhuleka, nomthelela wabasebenzisi.
Lungiselela izindlela zokuhlehlisa nezigameko ngaphambi kokukala.
Qhubeka Uhlole
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Tensor Cores quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Umhlahlandlela olandelayo
I-Tensor Parallelism yamamodeli amakhulu
Imibuzo evame ukubuzwa
What is Tensor Cores?
I-Tensor Cores amayunithi ehadiwe akhethekile ngaphakathi kwe-NVIDIA GPU yesimanje enza imisebenzi ye-matrix yokuphindaphinda nokuqongelela ngokushesha okukhulu. Kuyisizathu esiyinhloko sokuthi i-GPU eyodwa ingaqeqesha futhi isebenzise ama-oda amakhulu wenethiwekhi ye-neural yobukhulu ngokushesha kunokuba ikhompuyutha yenjongo evamile ingavumela.
Yikuphi ukusebenza kwezibalo okuyinhloko i-Tensor Cores eklanyelwe ngokuqondile ukusheshisa?
I-Tensor Cores yenza ukuphindaphinda kwe-matrix ehlanganisiwe kanye nokunqwabelana esinyathelweni esisodwa, okuwumsebenzi ngqo olawula izendlalelo zenethiwekhi ye-neural.
Yiziphi izakhiwo ze-NVIDIA GPU ezethula okokuqala i-Tensor Cores?
I-Tensor Cores ikhishwe nge-architecture ye-Volta ngo-2017, iqala ngokusebenza kwe-FP16 4x4 matrix.
Kungani ama-Tensor Cores evame ukusebenzisa ukunemba okuncishisiwe njenge-FP16 noma i-FP8?
Ukusebenzisa amabhithi ambalwa ngenombolo kusho amanani amaningi ageleza ku-hardware umjikelezo ngamunye, akhuphuka kakhulu isivinini ngokulahleka okuncane kuphela, okubekezelelekayo, kokunemba.
Ukuze kulondolozwe ukunemba, yikuphi ukunemba okuvame ukusetshenziswa yi-Tensor Cores ngesamba esisebenzayo (iqoqo)?
Okokufaka kungase kube ukunemba okuphansi, kodwa i-accumulator ivamise ukusebenza ngokunemba okuphezulu njenge-FP32 ukuze kukhawulwe ukunqwabelana kwamaphutha okusondeza.
Ngabe izinhlaka ze-AI zansuku zonke ezifana ne-PyTorch zizisebenzisa kanjani i-Tensor Cores?
Imitapo yolwazi engaphansi iphula imisebenzi emikhulu ye-matrix ibe amathayela osayizi we-Tensor-Core ngokuzenzakalela, ngakho izinhlaka zithola ukusheshisa ngaphandle kokufaka ikhodi mathupha.