HejuruUbuyobozi bukurikira
Core ML kuri On-Device Apple Kohereza
Tekiniki
Ubuyobozi bwa tekiniki
Tensor Cores ni ibikoresho byabikoresho byabugenewe imbere muri NVIDIA GPU igezweho ikora matrix kugwiza-no gukusanya ibikorwa byihuse.
Umva iki gitaboAI-byavuzwe · ~2 min
Ninimpamvu nyamukuru GPU imwe ishobora guhugura no gukoresha imiyoboro minini yimitsi itondekanya ubunini bwihuse kuruta rusange-intego ya compte yakwemerera.
Yinjijwe hamwe nubwubatsi bwa Volta muri 2017, Tensor Cores ni imirongo yabugenewe ibara matrike yo kugwiza mato hiyongereyeho (D = A x B + C) mugikorwa kimwe, aho gukora buri kugwiza icyarimwe kumurongo wa CUDA isanzwe. Kuberako hafi buri cyiciro cyurusobe rw'imitsi rugabanuka kugwiza matrix, ibi bihuye n'imibare AI ikeneye. Buri gisekuru cya GPU cyaguye ibyo bakora: Volta yakoze amatafari ya 4x4 FP16, mugihe nyuma yububiko bwa Ampere, Hopper, na Blackwell bwongeyeho imiterere yo hasi cyane nka TF32, BF16, INT8, FP8, na FP4. Ubusobanuro bwo hasi busobanura imibare myinshi yatunganijwe kumasaha, izamura cyane ibicuruzwa byamahugurwa hamwe numusozo mugihe bikomeza kwemerwa.
Guhugura imiterere nini yindimi nka GPT yuburyo bwa transformateur, aho miliyari zo kugwiza matrix kuri buri ntambwe ikora kuri Tensor Cores muri BF16 cyangwa FP8.
Gukoresha igihe nyacyo cyo kuganira kubiganiro na generator zamashusho, ukoresheje ingano ya INT8 cyangwa FP8 kugirango ukoreshe abakoresha benshi kuri GPU.
Kwihutisha NVIDIA DLSS mumikino ya videwo, aho urusobe rw'imitsi rushyira hejuru imiterere-yo hasi ikoresheje Tensor Cores buri kadamu.
Kwihutisha kubara siyanse nka protein-igwiza (AlphaFold) hamwe nikitegererezo cyikirere cyavuguruwe nka matrix-iremereye imitsi yimirimo.
Tensor Cores ikomeza kugenda yerekeza kumurongo wo hasi: Hopper yongeyeho FP8 hanyuma Blackwell yinjiza 4-bit FP4 hamwe nogupima ibyuma byifashishwa, byikuba kabiri byinjira buri ntambwe kubikorwa byakazi biremereye. Tegereza ubufasha bukomeye kubusa (gusimbuka zeru zeru), imiterere ya microscaling ihuza ibipimo byibipimo bito bito byimibare, hamwe no guhuza byimbitse hamwe na sisitemu yo kwibuka kugirango ibice bikomeze kugaburirwa. Mugihe icyitegererezo gikura, moteri ya matrix, ntabwo yihuta yisaha, ikomeza kuba ikibuga cyibanze hagati yimikorere yibikoresho bya AI.
Ibaruwa y'amakuru y'ubuntu
Inkuru eshatu zizewe za AI buri gitondo cy'iminsi y'akazi, zanditse mu rurimi rworoshye. Ubuntu iteka ryose, nta kwamamaza.
Imeri imwe buri munsi y'akazi. Kwiyimura mu buryo bumwe. Ntitwigera tugurisha cyangwa ngo dusangire aderesi yawe.
Isuzume
Ibisubizo ako kanya kuri buri kibazo, n'icyemezo gishobora gusangirwa gifite ID yagenzurwa umaze gutsinda isomo.
Fasha uburezi bwa AI bw'ubuntu. AI Understanding ni umuryango utegamiye kuri leta wa 501(c)(3) — nta kwamamaza, nta kwishyura, iteka ryose. Tanga impano
Tensor Cores ni ibikoresho byabikoresho byabugenewe imbere muri NVIDIA GPU igezweho ikora matrix kugwiza-no gukusanya ibikorwa byihuse. Ninimpamvu nyamukuru GPU imwe ishobora guhugura no gukoresha imiyoboro minini yimitsi itondekanya ubunini bwihuse kuruta rusange-intego ya compte yakwemerera.
Tensor Cores ikora matrix yagabanijwe hiyongereyeho kwirundanya muntambwe imwe, nigikorwa rwose cyiganjemo imiyoboro y'urusobe.
Tensor Cores yatangijwe nubwubatsi bwa Volta muri 2017, duhereye kubikorwa bya matrix ya FP16 4x4.
Gukoresha bits nkeya kuri numero bisobanura indangagaciro nyinshi zinyura mubyuma buri cyiciro, byongera cyane umuvuduko hamwe na bike gusa, mubisanzwe byihanganirwa, gutakaza neza.
Iyinjiza irashobora kuba idasobanutse neza, ariko ikusanyirizo mubisanzwe ikora muburyo bunoze nka FP32 kugirango igabanye iyubakwa ryamakosa.
Amasomero yibanze asibanganya ibikorwa bya matrix muri Tensor-Core-nini ya tile mu buryo bwikora, bityo ibikorerwa bikabona umuvuduko udafite code yintoki.
Komeza wige
Abandi bayobozi batoranijwe kuriyi ngingo
HejuruUbuyobozi bukurikira
Core ML kuri On-Device Apple Kohereza
Tekiniki