Imiterere ya FP8 na Hasi-Yuzuye
FP8 ni 8-biti ireremba-nimero yimibare ituma moderi ya AI ibika uburemere no gukoresha imibare ukoresheje kimwe cya kane cyo kwibuka kumibare isanzwe ya 32-bit.
Incamake
It is a key trick for making giant models cheaper and faster to train and serve.
Kwibira cyane
Imiyoboro ya Neural igizwe na miliyari y'imibare. Ubusanzwe iyo mibare yakoresheje bits 32 (FP32) cyangwa 16 bits (FP16 / BF16) buri umwe. FP8 ibagabanya kuri bits 8 gusa, igabanya kwibuka hamwe numuyoboro mugari hafi ya kabiri na 16-bit. Hariho uburyo bubiri busanzwe bwa FP8: E4M3 (bits 4 yerekana, 3 mantissa bits) itanga ibisobanuro byinshi ariko intera ntoya, kandi E5M2 (5 yerekanwe, 2 mantissa) itanga intera yagutse ariko intambwe yoroheje. Gucuruza-ni ubudahemuka: bits nkeya bisobanura kuzenguruka amakosa. Kugirango ugumane ukuri, ibikorerwa bikurikizwa kuri tensor cyangwa kuri buri kintu cyo gupima ibintu bikuraho indangagaciro murwego rwa FP8. Hopper ya NVIDIA na Blackwell GPUs yongeyeho ibyuma bya moteri ya FP8 ya matrix, bituma iba ingirakamaro mumahugurwa no gufata umwanzuro. Imiterere mishya nka MXFP8, MXFP4, na NVFP4 isunika ndetse hepfo hamwe na micro-scaring bloks.
Ubushishozi
Ikibazo cya FP8 ni intera igaragara. Hamwe nintoki gusa zerekana bits, nini cyangwa ntoya ibikorwa birenze cyangwa byuzuye kuri zeru. Gukosora ni igipimo: kugwiza tensor kubintu kugirango indangagaciro zayo zigwe mumadirishya ihagarariwe na FP8, kora FP8 kugwiza-kwegeranya, hanyuma ugabanye inyuma, akenshi ukusanya amafaranga igice mubice bisobanutse neza (FP16 / FP32). E4M3 isanzwe ikoreshwa muburemere no gukora, E5M2 kuri gradients aho intera irenze ibintu neza.
Ingaruka z'Ingamba
Igiciro na bije
Ibyemezo byubwubatsi bitwara imikorere nigiciro cyimikorere kumyaka.
Ibyemezo bisobanutse
Ubuhanga bwa tekinike bufasha amakipe guhitamo umurongo ukwiye, ntabwo ari shyashya gusa.
Kugenzura ubuziranenge
Guhitamo neza bya injeniyeri bigabanya ibintu byizewe mubikorwa.
Kazoza ka FP8 na Imiterere-Ntoya
Icyitonderwa kirimo kwiruka hepfo. Nyuma ya FP8 haje imiterere ya biti 4-bito (MXFP4, NVFP4) ipakira igipimo gito gisangiwe kuri buri gice gito, kandi ibyuma bya Blackwell byihutisha FP4 muburyo butaziguye. Tegereza ivangwa-risobanutse neza aho ibice bitandukanye bikoresha bit-ubugari butandukanye, wongeyeho amahugurwa meza yo kumenya neza bityo 4-bit ihinduka nkibisanzwe. Endgame irimo gukanda imipaka-yerekana urugero kuri bike, bihendutse nta gihombo cyiza gipima.
Gushyira mu bikorwa Isi
Guhugura indimi nini kuri NVIDIA Hopper / Blackwell GPUs ukoresheje FP8 kugirango winjize hafi inshuro ebyiri na BF16
Gukora chatbot umwanzuro muri FP8 kuburyo icyitegererezo gihuye na GPU nkeya kandi gisubiza ibyifuzo byinshi kumasegonda
Koresha E5M2 mugutumanaho gahoro gahoro mugihe cyamahugurwa yatanzwe kugirango ugabanye umurongo wa interineti hagati yumutwe
Kohereza MXFP4 / NVFP4-yuzuye-moderi kugirango ihuze imipaka-yerekana urugero kuri GPU imwe yibuka cyane kugirango ifate umwanzuro uhendutse
Ingaruka & Kurinda
Gutezimbere igipimo kimwe gishobora guhisha intege nke za sisitemu.
Ibikorwa Remezo no kubungabunga akenshi usanga bidahabwa agaciro.
Icyuho cyumutekano no kwitegereza birashobora kwiyongera uko sisitemu igenda igorana.
Igishushanyo mbonera
Sobanura ubukererwe, ubuziranenge, nigiciro cyibiciro mbere yo kubishyira mubikorwa.
Ibipimo byerekana umutwaro ufatika hamwe namakuru yimiterere.
Gukurikirana ibikoresho kubikosa, drift, ningaruka zabakoresha.
Tegura inzira yo gusubiza ibyabaye mbere yo gupima.
Komeza Ubushakashatsi
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
Ubuyobozi bukurikira
Icyitegererezo cyo Gukurikirana
Ibibazo bikunze kubazwa
What is FP8 and Low-Precision Formats?
FP8 ni 8-biti ireremba-nimero yimibare ituma moderi ya AI ibika uburemere no gukoresha imibare ukoresheje kimwe cya kane cyo kwibuka kumibare isanzwe ya 32-bit. Nuburyo bwingenzi bwo gukora moderi nini zihendutse kandi byihuse guhugura no gukora.
Umubare wa FP8 ukoresha bangahe ugereranije numero isanzwe ya FP32?
FP8 ikoresha bits 8 mugihe FP32 ikoresha 32 bits, FP8 rero ifata kimwe cya kane cyububiko nubunini.
Ni iki ibirango bya 'E4M3' na 'E5M2' bisobanura imiterere ya FP8?
E4M3 bisobanura bits 4 zerekana na 3 mantissa; E5M2 bisobanura 5 ibyerekana na 2 mantissa bits. Ibindi byinshi byerekana kwagura intera; byinshi mantissa bits ongeraho neza.
Kuki imiyoboro ya FP8 ikoresha ibintu bipima kuri tensors?
Hamwe nibintu bike byerekana bits, indangagaciro nini zirengerwa nuduto duto munsi ya zeru. Igipimo gikuraho tensors mumadirishya akoreshwa ya FP8 mbere yimibare.
Nubuhe bucuruzi-bubi ningenzi nyamukuru yo gukoresha FP8?
Bits nkeya bisobanura coarser guhagararirwa hamwe nikosa ryinshi. Inyungu ni mike cyane yibuka nubunini, kandi imibare yihuse kubikoresho bishyigikiwe.
Nibihe bisekuruza bya NVIDIA GPU byabanje kongeramo ibikoresho byabigenewe kubibare bya FP8?
GPUs ishingiye kuri Hopper nka H100 yazanye inkunga ya FP8 Tensor Core, hanyuma Blackwell yaje kuyigeza kuri FP4.