FP8 da Ƙarfin Madaidaicin Tsarukan
FP8 tsari ne mai lamba 8-bit mai iyo wanda ke ba da damar samfuran AI su adana ma'auni da gudanar da lissafi ta amfani da kwata na ƙwaƙwalwar daidaitattun lambobi 32-bit.
Dubawa
It is a key trick for making giant models cheaper and faster to train and serve.
Zurfafa nutsewa
Ana yin hanyoyin sadarwar jijiyoyi da biliyoyin lambobi. A al'ada, waɗannan lambobin suna amfani da 32 rago (FP32) ko 16 ragowa (FP16/BF16) kowanne. FP8 yana rage su zuwa rago 8 kawai, yana yanke ƙwaƙwalwar ajiya da bandwidth kusan a cikin rabin da 16-bit. Akwai shimfidu na FP8 guda biyu na gama gari: E4M3 (4 mai faɗakarwa ragowa, 3 mantissa ragowa) yana ba da ƙarin daidaici amma ƙarami, kuma E5M2 (mai magana 5, 2 mantissa) yana ba da faffadan kewayo amma matakai masu ƙarfi. Cinikin ciniki shine aminci: ƴan kaɗan na nufin kurakurai. Don zama daidai, tsarin tsarin yana amfani da kowane-tensor ko kowane toshe abubuwan sikeli waɗanda ke mayar da ƙima zuwa kewayon FP8 mai amfani. NVIDIA's Hopper da Blackwell GPUs sun kara injunan matrix na FP8 na kayan aiki, yana mai da amfani ga duka horo da tunani. Sabbin tsare-tsare kamar MXFP8, MXFP4, da NVFP4 suna turawa har ma da ƙasa tare da ɓangarorin sikelin sikeli.
Fahimtar Fasaha
Kalubalen FP8 shine kewayo mai ƙarfi. Tare da ɗimbin ɓangarorin maɓalli kawai, manyan ko kananun kunnawa sun cika ko ƙarƙashin ƙasa zuwa sifili. Gyaran yana haɓakawa: ninka tensor da ma'ana don haka ƙimarsa ta sauka a cikin taga mai wakiltar FP8, yi FP8 ninka-taruwa, sannan raba baya, yawanci yana tara juzu'i cikin madaidaici (FP16/FP32). E4M3 yawanci ana amfani dashi don ma'auni da kunnawa, E5M2 don gradients inda kewayon ke da mahimmanci fiye da daidaito.
Dabarun Tasiri
Kudin da kasafin kuɗi
Hukunce-hukuncen gine-gine suna haifar da aiki da tsadar aiki na shekaru.
Shawarwari masu haske
Ilimin fasaha yana taimaka wa ƙungiyoyi su zaɓi tari mai kyau, ba kawai sabon abu ba.
Kula da inganci
Zaɓuɓɓukan injiniya mafi kyau suna rage abin dogaro a cikin samarwa.
Makomar FP8 da Tsarukan Madaidaici
Madaidaici yana tsere zuwa ƙasa. Bayan FP8 ya zo 4-bit micro-scaling Formats (MXFP4, NVFP4) waɗanda ke ɗaukar ƙaramin sikelin da aka raba a kowane ƙaramin shinge, kuma kayan aikin Blackwell yanzu yana haɓaka FP4 kai tsaye. Yi tsammanin gauraye-daidaitaccen girke-girke inda yadudduka daban-daban ke amfani da nau'i-nau'i-nau'i daban-daban, tare da mafi kyawun horarwar ƙididdigewa don haka 4-bit ya zama tsoho don ƙaddamarwa. Ƙarshen wasan yana matsar da ƙira-ƙira na kan iyaka akan ƙananan kwakwalwan kwamfuta masu rahusa ba tare da asara mai inganci ba.
Aiwatar da Gaskiyar Duniya
Horar da manyan nau'ikan harshe akan NVIDIA Hopper/Blackwell GPUs ta amfani da FP8 zuwa kusan nau'ikan kayan aiki biyu da BF16
Bayar da ra'ayi na chatbot a cikin FP8 don haka samfurin ya dace da ƴan GPUs kuma yana amsa ƙarin buƙatun daƙiƙa guda.
Amfani da E5M2 don sadarwar gradient yayin horarwar da aka rarraba don yanke bandwidth cibiyar sadarwa tsakanin nodes
Ƙirƙirar ƙirar ƙididdigewa na MXFP4/NVFP4 don dacewa da ƙirar iyaka akan babban GPU mai ƙwaƙwalwar ajiya guda ɗaya don ƙididdigewa mai rahusa
Hatsari & Tsare-tsare
Haɓaka ma'auni ɗaya na iya ɓoye manyan raunin tsarin.
Sau da yawa ana raina kayan more rayuwa da kuma kuɗin kulawa.
Tsaro da gibin lura na iya girma yayin da tsarin ke ƙara haɓaka.
Taswirar Hanya
Ƙayyade latency, inganci, da maƙasudin farashi kafin aiwatarwa.
Alamar ma'auni a ƙarƙashin ainihin kaya da yanayin bayanai.
Kula da kayan aiki don kurakurai, ɗigo, da tasirin mai amfani.
Shirya bijirowa da hanyoyin mayar da martani kafin sikeli.
Ci gaba da Bincike
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
Jagora na gaba
Samfuran Serialization Formats
Tambayoyin da ake yawan yi
What is FP8 and Low-Precision Formats?
FP8 tsari ne mai lamba 8-bit mai iyo wanda ke ba da damar samfuran AI su adana ma'auni da gudanar da lissafi ta amfani da kwata na ƙwaƙwalwar daidaitattun lambobi 32-bit. Dabarar maɓalli ce don yin ƙattai masu arha da sauri don horarwa da hidima.
Rago nawa ne lambar FP8 ke amfani da ita idan aka kwatanta da daidaitaccen lambar FP32?
FP8 yana amfani da 8 ragowa yayin da FP32 ke amfani da 32 ragowa, don haka FP8 yana ɗaukar kashi ɗaya cikin huɗu na ajiya da bandwidth.
Menene alamun 'E4M3' da 'E5M2' suka bayyana game da tsarin FP8?
E4M3 yana nufin 4 ragowa mai magana da 3 mantissa ragowa; E5M2 yana nufin 5 mai magana da 2 mantissa ragowa. Ƙarin ɓangarorin maɓalli suna faɗaɗa kewayon; ƙarin mantissa ragowa suna ƙara daidaito.
Me yasa bututun FP8 ke amfani da abubuwan ƙima ga tenors?
Tare da ƴan ƴan ɗimbin fasiƙanci, manyan dabi'u suna mamayewa kuma ƙanana suna ƙarƙashin ƙasa zuwa sifili. Ƙididdigar ƙididdiga ta sake daidaitawa a cikin taga mai amfani na FP8 kafin lissafi.
Wanne-kashe-kashe shine babban koma bayan amfani da FP8?
Ƙananan rago yana nufin wakilci mai ƙarfi da ƙarin kuskuren zagaye. Amfanin yana da ƙarancin ƙwaƙwalwar ajiya da bandwidth, da lissafi cikin sauri akan kayan aikin tallafi.
Wane ƙarni na NVIDIA GPU ne ya fara ƙara tallafin kayan aikin kwazo don lissafin matrix na FP8?
GPUs na tushen Hopper kamar H100 sun gabatar da tallafin FP8 Tensor Core, kuma Blackwell daga baya ya tsawaita wannan zuwa FP4.