Tensor
Tensor Cores ay jumtukaay lañu yuñ jagleel GPU NVIDIA yu bees yi, ñuy def matrix yu bari ak dajale ci anam wu gaaw lool.
Résumé
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.
Plongeur bu xóot
Dugaale ak architecture Volta ci 2017, Tensor Cores ay sircuit yuñ jagleel ñuy xayma ab matrix bu ndaw buy yokk (D = A x B + C) ci benn jëfandikoo, moo gën ñuy def bu nekk benn-benn ci cores CUDA standard. Ndax daanaka bépp couche ci reso neuronal dafay wàññeeku ci matrix yu bari, loolu méngoo ak li IA soxla ci math. GPU bu nekk dafay yokk limuy jëfandikoo: Volta defar na 4x4 FP16, ci ganaw ga Ampere, Hopper, ak Blackwell yokk nañu ay formaa yu gëna ndaw yu melni TF32, BF16, INT8, FP8, ak FP4. Lu gëna ndaw ci njub mooy limu gëna bari ci montor bu nekk, loolu dafay yokk bu baax produit bi ci tàggat yaram ak ci inference boole ci tëye njubte gi.
Gis-gis xarala
Tensor Core dafay yokk ñaari matrix yu ndaw ba noppi dajale resultaa bi ci benn jéego buñ boole, di jàppee ni benn valeur bi ñuy dugal dañu koy jëfandikoowaat ci élément yu bari yuy génn. Dafay faral di lire ay dugal ci anam wu wàññeeku (FP16, BF16, wala FP8) waaye dafay dajale xaalis biy dawal ci anam wu gëna dëggu (dafay faral di nekk FP32) ngir wàññi njuumti yi ci wërsëg. Biblioteek losisel yu melni cuBLAS ak cuDNN, ak kaadar yu melni PyTorch, dañuy boole matris yu mag yi ci blok yu ndaw yii ci saasi suko defee model yi mëna gaaw te duñu soxla kodage manuel.
njeextalu pexe
Njëgg ak budget
Dogal yi architecture di jël dañuy indi njariñ ak njëgu liggéey bi ay at ci ginaaw.
dogal yu gëna leer
Njàngalem xarala yi dafay jàppale ekip yi ñu tànn li gën, te baña yam ci li gëna bees daal.
Xool kalite
Tanneef yu gëna baax ci wàllu ingeñër dina wàññi jafe-jafe yi ci wàllu wóor ci liggéey bi.
Ëlëgu Tensor Cores
Tensor Cores wéyam di dem ci gëna néew njub: Hopper yokk FP8 ak Blackwell dugal 4-bit FP4 ak scaling hardware-managed, lu tollu ci ñaari yoon produit bi jéego bu nekk ngir liggéey bu diis. Xaarandi ndimmbal bu gëna dëgër ngir sparsity (teggi poids nul), formaa microscaling yuy takk facteur scale ci bloc yu ndaw yu lim, ak boole bu gëna xóot ak sistemu mémoire ngir core yi mëna des. Lu model yi di gëna màgg, motëru matrix bi, du gaawaayu montor bu ñor bi, mooy des ci xeex bi gëna mag ci performance hardware IA.
Doxal ci àdduna dëgg
Taggat xeetu làkk yu mag yu melni transformatër yu nuroo ak GPT, fu ay miliyaar ciy bari matrix jéego bu nekk di dox ci kaw Tensor Cores ci BF16 wala FP8.
Doxal ay gis-gis ci jamono dëgg ngir chatbots ak defarkati nataal, jëfandikoo INT8 wala FP8 ngir mëna jàppale jëfandikukat yu bari ci GPU bu nekk.
Gaawaay NVIDIA DLSS ci jeu video, fu reso neuronal di yokk kadre yu am dayo bu néew di jëfandikoo Tensor Cores ci kadre bu nekk.
Gaawaale ordinatër yu gëstukat yi lu ci melni pliage protein (AlphaFold) ak modeli meteo yuñ defaraat ñu nekk ay liggéey yu diis ci matrix.
Risk yi ak balustrade yi
Optimize benn benchmark mën na nëbb ñakk kattan yu gëna yaatu ci sistem bi.
Njëg li ñuy fay ci infrastructure yi ak ci toppatoo dañuy faral di suufeel.
Bu sistem yi di gëna xawa jafee xam, jafe-jafe yi am ci wàllu kaaraange ak seetlu mën nañu gëna bari.
Roadmap ngir samp gi
Mandargal latency, kalite, ak njëg yi laata ngay jëfandikoo.
Benchmark ci biir sargal ak done yu dëggu.
Jumtukaay bi di saytu njuumte yi, derive bi ak njeextalu jëfandikukat bi.
Waajal rollback ak yooni tontu ci jafe-jafe yi laata ngay eskale.
Weyal di banneexu
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
Gis bi ci topp
paralelism tensor ngir model yu mag
Laaj yi ñuy faral di laaj
What is Tensor Cores?
Tensor Cores ay jumtukaay lañu yuñ jagleel GPU NVIDIA yu bees yi, ñuy def matrix yu bari ak dajale ci anam wu gaaw lool. Mooy sabab bi tax benn GPU mëna tàggat ak doxal reso neuronal yu mag yu magnitude yu gëna gaaw ci ordinatër yu bari yi.
Ban core math lañu defar Tensor Cores ngir gaaw?
Tensor Cores yi dañuy def matrix fused buy yokk boole ci accumulation ci benn jéego, muy liggéey bi gëna am solo ci reso neuronal yi.
Ban arsitektir GPU NVIDIA moo njëkka dugal Tensor Cores?
Tensor Cores mingi njëkka feeñ ci arsitektir Volta ci 2017, daal di tàmbalee ci FP16 4x4 jëfandikoo matrix.
Lan moo waral Tensor Cores di faral di jëfandikoo njubte gu néew lu melni FP16 wala FP8?
Jëfandikoo bit yu néew ci nimero bu nekk dafay tekki ni valeur yu bari dañuy jaar ci hardware bi ci cycle bu nekk, di yokk bu baax gaawaay bi ak lu tuuti, luñuy faral di muñ, ñàkka am njub.
Ngir baña gaawa jeex, ban njub la Tensor Cores yi di faral di jëfandikoo ngir am xaalis buy daw (accumulation)?
Duggal yi mën nañu baña am njub, waaye accumulateur bi dafay faral di dox ci njub bu gëna rëy lu melni FP32 ngir wàññi njuumti yi ci rounding.
Ak naka la kaadar IA yu melni PyTorch di jëfandikoo Tensor Cores?
Bibliothèque yi ci suuf dañuy xaaj matrix yu mag yi ci ay carreaux yu tollu ni Tensor-Core ci saasi, suko defee kaadar yi am gaawaay bi te duñu soxla kode manuel.