Up tókànItọsọna atẹle
Awọn ọna kika Serialization awoṣe
Imọ-ẹrọ
Imọ Itọsọna
FP8 ati Kekere-konge Awọn ọna kika
FP8 jẹ ọna kika nọmba oju omi lilefoofo 8-bit ti o jẹ ki awọn awoṣe AI tọju awọn iwuwo ati ṣiṣe iṣiro nipa lilo idamẹrin ti iranti awọn nọmba 32-bit boṣewa.
Gbọ si itọsọna yiiAI-ṣe afihan · ~2 min
Lori iwe yi2 min ka
Akopọ
O jẹ ẹtan bọtini fun ṣiṣe awọn awoṣe omiran din owo ati yiyara lati ṣe ikẹkọ ati ṣiṣẹ.
Jin Dive
Awọn nẹtiwọọki nkankikan jẹ ti awọn ọkẹ àìmọye awọn nọmba. Ni aṣa awọn nọmba wọnyẹn lo 32 bit (FP32) tabi 16 die-die (FP16/BF16) ọkọọkan. FP8 dinku wọn si awọn iwọn 8 nikan, gige iranti ati bandiwidi ni aijọju ni idaji dipo 16-bit. Awọn ipalemo FP8 ti o wọpọ meji lo wa: E4M3 (awọn iwọn olutayo mẹrin, awọn iwọn mantissa 3) yoo fun ni konge diẹ sii ṣugbọn iwọn ti o kere ju, ati E5M2 (olupilẹṣẹ 5, 2 mantissa) n funni ni iwọn ti o gbooro ṣugbọn awọn igbesẹ ti o kere ju. Iṣowo-pipa jẹ iṣootọ: awọn iwọn diẹ tumọ si awọn aṣiṣe iyipo. Lati duro deede, awọn ilana lo fun-tensor tabi awọn ifosiwewe idinaki ti o ṣe iwọn awọn iye sinu iwọn lilo FP8. NVIDIA's Hopper ati Blackwell GPUs ṣafikun awọn ẹrọ matrix FP8 hardware, ti o jẹ ki o wulo fun ikẹkọ mejeeji ati itọkasi. Awọn ọna kika tuntun bii MXFP8, MXFP4, ati NVFP4 Titari paapaa ni isalẹ pẹlu awọn bulọọki iwọn-miki ti o pin.
Real-World imuse
A team benchmarks an FP8 training recipe on hardware with supported matrix operations and compares it with BF16.
An inference team checks weight, activation, and cache memory before selecting FP8 serving.
A training implementation uses an FP8 encoding with a wider range for tensors whose values require it.
A deployment team evaluates supported four-bit formats against an accuracy and memory budget.
Ọjọ iwaju ti FP8 ati Awọn ọna kika Ipeye-Kekere
Lower-precision formats should be evaluated as complete recipes rather than universal speed or quality guarantees. Block scaling, mixed-precision layers, and supported kernels can help balance numerical error with memory and compute costs. Four-bit formats such as MXFP4 and NVFP4 are additional choices on supporting systems; their range, scaling, and hardware requirements differ. A team should benchmark its own workload and retain higher precision where measured errors justify it.
Fifi o sinu iwa
Kini idi ti o ṣe pataki
- Iye owo ati isuna. Awọn ipinnu faaji ṣe awakọ iṣẹ ati idiyele iṣẹ fun awọn ọdun.
- Awọn ipinnu diẹ sii. Ẹkọ imọ-ẹrọ ṣe iranlọwọ fun awọn ẹgbẹ lati yan akopọ to tọ, kii ṣe ọkan tuntun nikan.
- Iṣakoso didara. Awọn yiyan imọ-ẹrọ to dara julọ dinku awọn iṣẹlẹ igbẹkẹle ni iṣelọpọ.
Ṣọra fun
- Ṣiṣepe ala-ilẹ kan le tọju awọn ailagbara eto ti o gbooro.
- Awọn ohun elo amayederun ati awọn idiyele itọju nigbagbogbo ni aibikita.
- Aabo ati awọn ela akiyesi le dagba bi awọn eto ṣe di eka sii.
Next awọn igbesẹ
- Ṣetumo lairi, didara, ati awọn ibi-afẹde idiyele ṣaaju imuse.
- Aṣepari labẹ ẹru ojulowo ati awọn ipo data.
- Abojuto ohun elo fun awọn aṣiṣe, fiseete, ati ipa olumulo.
- Mura ipadasẹhin pada ati awọn ipa ọna esi iṣẹlẹ ṣaaju iwọn.
Tesiwaju Ṣiṣawari
Ìwé ìròyìn ọfẹ
Gba ìkànìyàn AI ojoojúmọ́
Ìtàn AI mẹta tó jẹ́rìí ní gbogbo owurọ́ ọjọ́ iṣẹ́, kọ́wé ní Gẹ̀ẹ́sì tó rọrùn. Ọfẹ́ fún gbogbo àkókò, kò sí ìpolówó.
Ìmeèlì kan ní gbogbo ọjọ́ iṣẹ́. Yọkúrò ní ìkànìyàn ní ìkànìyàn kan. A kò ta tàbí pín àdírẹ́sì rẹ.
Ṣe idanwo fun ara rẹ
Take the FP8 and Low-Precision Formats quiz
Ìdáhùn lẹ́sẹ̀kẹsẹ̀ lórí gbogbo ìdáhùn, àti iwe-ẹ̀rí tó lè pin pẹ̀lú ID tó lè jẹ́rìí lẹ́yìn tí o bá kọ́ ẹ̀kọ́.
Ṣe àtìlẹ́yìn ẹ̀kọ́ AI ọfẹ́. AI Understanding jẹ́ àjọ àìní èrè 501(c)(3) — kò sí ìpolówó, kò sí àkúnya owó, ní gbogbo àkókò. Ṣe ẹbun kan
Awọn ibeere ti a beere nigbagbogbo
Kini FP8 ati Awọn ọna kika Ipeye-Kekere?
FP8 jẹ ọna kika nọmba oju omi lilefoofo 8-bit ti o jẹ ki awọn awoṣe AI tọju awọn iwuwo ati ṣiṣe iṣiro nipa lilo idamẹrin ti iranti awọn nọmba 32-bit boṣewa. O jẹ ẹtan bọtini fun ṣiṣe awọn awoṣe omiran din owo ati yiyara lati ṣe ikẹkọ ati ṣiṣẹ.
Awọn die-die melo ni nọmba FP8 nlo ni akawe si nọmba FP32 boṣewa kan?
FP8 nlo 8 die-die nigba ti FP32 nlo 32 die-die, nitorina FP8 gba idamẹrin ti ibi ipamọ ati bandiwidi.
Kini awọn aami 'E4M3' ati 'E5M2' ṣe apejuwe nipa ọna kika FP8 kan?
E4M3 tumo si 4 olutayo die-die ati 3 mantissa die-die; E5M2 tumo si 5 olutayo ati 2 mantissa die-die. Awọn ege olupilẹṣẹ diẹ sii gbooro iwọn; diẹ mantissa die-die fi konge.
Kini idi ti awọn opo gigun ti FP8 lo awọn ifosiwewe igbelosoke si awọn tenors?
Pẹlu awọn ege olupilẹṣẹ diẹ, awọn iye nla nṣàn ati awọn ti o kere ju ti nṣàn si odo. Iwontunwọnsi ṣe atunṣe awọn tenors sinu ferese ohun elo FP8 ṣaaju ki iṣiro naa.
Iṣowo-pipa wo ni isalẹ akọkọ ti lilo FP8?
Awọn die-die ti o kere julọ tumọ si aṣoju irẹwẹsi ati aṣiṣe iyipo diẹ sii. Anfaani naa kere pupọ si iranti ati bandiwidi, ati iyara iyara lori ohun elo atilẹyin.
Eyi ti iran NVIDIA GPU akọkọ ṣafikun atilẹyin ohun elo igbẹhin fun matrix matrix FP8?
Awọn GPU ti o da lori Hopper bii H100 ṣe afihan atilẹyin FP8 Tensor Core, ati Blackwell nigbamii fa eyi si FP4.
Tesiwaju kikọ
Jẹmọ awọn itọsọna
Awọn itọsọna diẹ sii ti a yan fun koko yii