Tensor Cores
Tensor Cores bụ ngwaike pụrụ iche n'ime NVIDIA GPU nke ọgbara ọhụrụ na-arụ ọrụ matriks na-amụba ma na-achịkọta ngwa ngwa.
Nchịkọta
Ha bụ isi ihe kpatara otu GPU nwere ike ịzụ ma na-agba ọsọ nnukwu netwọkụ akwara ozi ngwa ngwa karịa mkpokọta ebumnuche n'ozuzu ga-ekwe.
Ime miri emi
Ewebata ya na ihe owuwu Volta na 2017, Tensor Cores bụ sekit raara onwe ya nye na-agbakọ obere ọnụọgụ matrix gbakwunyere (D = A x B + C) n'otu ọrụ, kama ime nke ọ bụla na-amụba otu n'otu oge na cores CUDA ọkọlọtọ. N'ihi na ihe fọrọ nke nta ka ọ bụrụ akụkụ ọ bụla nke netwọkụ akwara na-ebelata ka ọ bụrụ ụbara matriks, nke a dabara na mgbakọ na mwepụ AI chọrọ n'ezie. Ọgbọ GPU ọ bụla gbasaara ihe ha na-arụ: Volta mere taịlị 4 × 4 FP16, ebe emechara Ampere, Hopper, na Blackwell architectures gbakwunyere usoro dị ala dị ka TF32, BF16, INT8, FP8, na FP4. Obere nkenke pụtara ọnụ ọgụgụ ndị ahazigharịrị n'otu elekere, na-abawanye n'ike n'ike mmepụta maka ọzụzụ na ntinye aka ma na-edobe izi ezi nke anabatara.
Nghọta nka nka
A Tensor Core na-amụba obere matrices abụọ ma na-achịkọta nsonaazụ ya n'otu nzọụkwụ jikọtara ọnụ, na-erigbu eziokwu ahụ bụ na a na-ejigharị otu ụkpụrụ ntinye ahụ n'ọtụtụ ihe nrụpụta. Ọ na-agụkarị ntinye n'ụzọ ziri ezi (FP16, BF16, ma ọ bụ FP8) mana na-akwakọba nchikota na-agba ọsọ na nkenke dị elu (na-abụkarị FP32) iji kpachie njehie okirikiri. Ọbá akwụkwọ sọftụwia dị ka cuBLAS na cuDNN, yana usoro dị ka PyTorch, tile nnukwu matrices n'ime obere ngọngọ ndị a na-akpaghị aka ka ụdị na-enweta ọsọ ọsọ na-enweghị ntinye akwụkwọ ntuziaka.
Mmetụta atụmatụ
Ọnụ ego na mmefu ego
Mkpebi ihe owuwu ụlọ na-akwalite arụmọrụ yana ọnụ ahịa ọrụ ruo ọtụtụ afọ.
Mkpebi doro anya
Nkà mmụta nka na-enyere ndị otu egwuregwu aka ịhọrọ nchịkọta ziri ezi, ọ bụghị naanị nke kachasị ọhụrụ.
Quality akara
Nhọrọ injinia ka mma na-ebelata ihe omume ntụkwasị obi na mmepụta.
Ọdịnihu nke Tensor Cores
Tensor Cores gara n'ihu na-aga n'ihu n'ụzọ ziri ezi na-adị ala: Hopper gbakwunyere FP8 na Blackwell webatara 4-bit FP4 na njikwa ngwaike jikwaa, na-ewebata usoro nke ọ bụla ugboro abụọ maka ntinye-ibu ọrụ. Na-atụ anya nkwado siri ike maka sparity (ịwụpụ ihe efu efu), usoro microscaling nke na-etinye ihe dị n'ọ̀tụ̀tụ̀ na obere ngọngọ nke ọnụ ọgụgụ, yana njikọ miri emi na sistemu ebe nchekwa ka cores na-eri nri. Ka ụdị na-eto eto, igwe matriks, ọ bụghị ọsọ elekere, ka bụ ebe etiti maka ịrụ ọrụ ngwaike AI.
Mmejuputa n'ezie n'ụwa
Ọzụzụ nnukwu ụdị asụsụ dị ka mgbanwe mgbanwe ụdị GPT, ebe ọtụtụ ijeri matrix multiplications kwa nzọụkwụ na-agba ọsọ na Tensor Cores na BF16 ma ọ bụ FP8.
Na-agba ọsọ ntinye ozugbo maka chatbots na ndị na-emepụta ihe onyonyo, na-eji ọnụọgụ INT8 ma ọ bụ FP8 iji jeere ndị ọrụ karịa kwa GPU ozi.
Na-eme ngwa ngwa NVIDIA DLSS na egwuregwu vidiyo, ebe netwọkụ akwara na-ebuli okpokolo agba dị ala site na iji Tensor Cores nke ọ bụla.
Na-eme ka kọmpụta sayensị ọsọ ọsọ dị ka mpịakọta protein (AlfaFold) na ụdị ihu igwe emezigharịrị dị ka matriks-arọ akwara arụ ọrụ.
Ihe ize ndụ & okporo ụzọ nche
Ịkwalite otu akara ngosi nwere ike zoo adịghị ike sistemụ sara mbara.
A na-eledakarị ihe akụrụngwa na ụgwọ ọrụ anya.
Ọdịiche nchekwa na nleba anya nwere ike itolite ka sistemu na-adịwanye mgbagwoju anya.
Map mmejuputa
Kọwaa latency, ịdịmma na ebumnuche ọnụ ahịa tupu mmejuputa ya.
Benchmark n'okpuru ibu dị adị na ọnọdụ data.
Nleba anya akụrụngwa maka mperi, ịkpafu na mmetụta onye ọrụ.
Kwadebe ụzọ nzaghachi azụghachi azụ na ihe omume tupu ịchachaa.
Nọgide na-eme nchọpụta
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
Ntuziaka na-esote
Tensor Parallelism maka Model buru ibu
Ajụjụ a na-ajụkarị
Kedu ihe bụ Tensor Cores?
Tensor Cores bụ ngwaike pụrụ iche n'ime NVIDIA GPU nke ọgbara ọhụrụ na-arụ ọrụ matriks na-amụba ma na-achịkọta ngwa ngwa. Ha bụ isi ihe kpatara otu GPU nwere ike ịzụ ma na-agba ọsọ nnukwu netwọkụ akwara ozi ngwa ngwa karịa mkpokọta ebumnuche n'ozuzu ga-ekwe.
Kedu ọrụ mgbakọ na mwepụ bụ Tensor Cores ahaziri ka ọ mee ngwa ngwa?
Tensor Cores na-arụ multiplication matrix fused gbakwunyere mkpokọta n'otu nzọụkwụ, nke bụ kpọmkwem ọrụ na-achịkwa ọkwa netwọkụ akwara.
Kedu NVIDIA GPU architecture bu ụzọ webatara Tensor Cores?
Tensor Cores ejiri arụ ọrụ Volta rụpụtara na 2017, malite na arụmọrụ matrix FP16 4 × 4.
Kedu ihe kpatara Tensor Cores ji ejikarị ezi nkenke dị ka FP16 ma ọ bụ FP8?
Iji obere ibe n'ibe n'otu ọnụ ọgụgụ pụtara ọtụtụ ụkpụrụ na-agafe na ngwaike okirikiri nke ọ bụla, na-abawanye ọsọ ọsọ na-eji naanị obere, na-enwekarị nnabata, enweghị nkenke.
Iji chekwaa izi ezi, kedu nkenke nkenke Tensor Cores na-ejikarị maka nchikota na-agba ọsọ (mkpokọ)?
Ntinye nwere ike ịbụ nkenke dị ala, mana onye nchịkọta na-agba ọsọ na nkenke dị elu dị ka FP32 iji kpachie nrụpụta nke njehie okirikiri.
Kedu ka usoro AI kwa ụbọchị dị ka PyTorch si eji Tensor Cores?
Ọbá akwụkwọ ndị dị n'okpuru na-agbaji nnukwu arụmọrụ matriks n'ime taịlị Tensor-Core na-akpaghị aka, yabụ na usoro na-enweta ọsọ ọsọ na-enweghị ntinye akwụkwọ ntuziaka.