Ụdị FP8 na nke dị ala
FP8 bụ usoro ọnụọgụ 8-bit na-ese n'elu mmiri nke na-ahapụ ụdị AI na-echekwa nha ma na-agba ọsọ mgbakọ na mwepụ site na iji nkeji anọ nke ebe nchekwa nke nọmba 32-bit ọkọlọtọ.
Nchịkọta
It is a key trick for making giant models cheaper and faster to train and serve.
Ime miri emi
A na-eme netwọkụ akwara ozi nke ọtụtụ ijeri ọnụọgụ. Na omenala, ọnụọgụ ndị ahụ na-eji 32 bit (FP32) ma ọ bụ 16 bit (FP16/BF16) nke ọ bụla. FP8 na-ebelata ha ka ọ bụrụ naanị 8 ibe n'ibe, na-ebelata ebe nchekwa na bandwit na ọkara na 16-bit. Enwere nhazi FP8 abụọ a na-ahụkarị: E4M3 (4 exponent bits, 3 mantissa ibe n'ibe) na-enye nkenke karịa mana ọ dị ntakịrị, yana E5M2 (5 exponent, 2 mantissa) na-enye ohere sara mbara mana nzọụkwụ siri ike. Azụmahịa ahụ bụ ikwesị ntụkwasị obi: ntakịrị ntakịrị pụtara mperi mperi. Iji nọgide na-abụ nke ziri ezi, usoro nhazi na-etinye ihe na-eme ka ọ bụrụ tensor ma ọ bụ ngọngọ nke na-eweghachi ụkpụrụ n'ime oke FP8 nwere ike iji. NVIDIA's Hopper na Blackwell GPUs gbakwunyere igwe FP8 ngwaike matrix, na-eme ka ọ baa uru maka ọzụzụ na ntinye aka. Ụdị ọhụrụ dị ka MXFP8, MXFP4, na NVFP4 na-ebuli ọbụna ala site n'iji ngọngọ micro-scaling.
Nghọta nka nka
Ihe ịma aka FP8 bụ oke ike. N'iji naanị njuaka nke ngwa ngwa, nnukwu ma ọ bụ obere ọrụ na-ejupụta ma ọ bụ na-eruba na efu. Ndozi ahụ na-arị elu: mụbaa tensor site na ihe ka ụkpụrụ ya wee daa na windo FP8 a na-anọchi anya ya, mee ka FP8 mụbaa, wee kewaa azụ azụ, na-agbakọkarị nchikota akụkụ na nkenke dị elu (FP16/FP32). A na-ejikarị E4M3 eme ihe maka ịdị arọ na ịrụ ọrụ, E5M2 maka gradients ebe oke dị mkpa karịa nkenke.
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 FP8 na Ụdị nkenke dị ala
Nkenkenke na-agbada ala. Mgbe FP8 bịarutere 4-bit micro-scaling formats (MXFP4, NVFP4) nke na-ebu obere nha nha n'otu obere ngọngọ, yana ngwa Blackwell na-agbatị FP4 ozugbo. Na-atụ anya ntụzịaka agwakọta-nkenke ebe ọkwa dị iche iche na-eji obosara-bit dị iche iche, gbakwunyere ọzụzụ mara mma nke ọma ka 4-bit bụrụ nke ndabara maka ntinye. Egwuregwu ngwụcha ahụ na-amanye ụdị n'oke oke n'obere ibe dị ọnụ ala na-enweghị mfu dị mma.
Mmejuputa n'ezie n'ụwa
Ọzụzụ ụdị asụsụ buru ibu na NVIDIA Hopper/Blackwell GPUs site na iji FP8 ka ọ bụrụ ihe nrụpụta okpukpu abụọ karịa BF16
Ije ozi ntinye aka chatbot na FP8 ka ihe nlereanya dabara na GPU ole na ole wee zaa arịrịọ ndị ọzọ kwa sekọnd.
Iji E5M2 maka nkwurịta okwu gradient n'oge ọzụzụ ekesa iji belata bandwidth netwọk n'etiti ọnụ
Na-ebuga ụdị MXFP4/NVFP4 ka ọ dabara n'ụdị ọnụ ala n'otu GPU ebe nchekwa dị elu maka ntinye dị ọnụ ala.
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 FP8 and Low-Precision Formats 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
Ụdị Serialization Model
Ajụjụ a na-ajụkarị
What is FP8 and Low-Precision Formats?
FP8 bụ usoro ọnụọgụ 8-bit na-ese n'elu mmiri nke na-ahapụ ụdị AI na-echekwa nha ma na-agba ọsọ mgbakọ na mwepụ site na iji nkeji anọ nke ebe nchekwa nke nọmba 32-bit ọkọlọtọ. Ọ bụ aghụghọ dị mkpa maka ime nnukwu ụdị dị ọnụ ala na ngwa ngwa iji zụọ na ije ozi.
Kedu ọnụọgụ FP8 na-eji tụnyere nọmba FP32 ọkọlọtọ?
FP8 na-eji 8 ibe n'ibe ebe FP32 na-eji 32 bits, ya mere FP8 na-ewe otu ụzọ n'ụzọ anọ nke nchekwa na bandwit.
Kedu ihe akara 'E4M3' na 'E5M2' na-akọwa maka usoro FP8?
E4M3 pụtara ibe n'ibe 4 na mantissa ibe n'ibe 3; E5M2 pụtara exponent 5 na 2 mantissa ibe n'ibe. Iberibe ndị ọzọ na-emepe emepe na-agbasa oke; ọzọ mantissa ibe n'ibe tinye nkenke.
Kedu ihe kpatara pipeline FP8 ji etinye ihe mkpali na tenors?
N'ịbụ ndị nwere obere okwu ole na ole, ọnụ ahịa buru ibu na-ejupụta na obere ndị na-eruba na efu. Ịmekọ ihe na-eme ka tenors dị na windo FP8 nwere ike iji tupu mgbakọ na mwepụ.
Kedu azụmaahịa bụ isi ọdịda nke iji FP8?
Obere ibe n'ibe pụtara nnochite anya dị oke mkpa yana mperi okirikiri karịa. Uru ya dị obere ebe nchekwa na bandwit, yana mgbakọ na mwepụ ngwa ngwa na ngwaike akwadoro.
Kedu ọgbọ NVIDIA GPU gbakwunyere nkwado ngwaike raara onwe ya nye maka mgbakọ na mwepụ FP8?
GPU ndị dabere na Hopper dị ka H100 webatara nkwado FP8 Tensor Core, na Blackwell mechara gbasaa nke a na FP4.