InoteveraGaidhi rinotevera
Estimating GPU Memory Needed for LLMs
Tekinoroji
Nhungamiro yehunyanzvi
High Bandwidth Memory
Yakakwira Bandwidth Memory (HBM) yakarongedzerwa ndangariro yakaiswa padyo neGPU iyo inopa data nekukurumidza zvakanyanya kupfuura yakajairika RAM.
Teerera kugwaro iriAI-yakarondedzerwa · ~2 min
Pa peji ino2 min verenga
Pfupiso
Ndizvo zvinochengeta AI accelerators kudya, kudzivirira ane simba compute cores kubva pakugara zvisina basa ivo vakamirira maremu emhando uye data.
Kudzika Kwakadzika
HBM inogadzirisa bhodhoro rekutanga: yemazuva ano AI machipi anogona kuita matiririyoni ekushanda pasekondi, asi chete kana data rasvika nekukurumidza zvakakwana. Yakajairwa GDDR ndangariro inobatanidza pamusoro pebhazi rakatetepa, nepo HBM inorongedza akawanda DRAM inofa yakatwasuka uye inoabatanidza nezviuru zvemawaya madiki akatwasuka anonzi kuburikidza-silicon vias (TSVs). Aya masheki anogara pasilicon interposer millimeters kubva kuGPU, ichipa yakanyanyisa nzira yedata, funga zviuru zvebhiti kamwechete panzvimbo yemazana. Mhedzisiro ndeye bandwidth kuyerwa mumaterabytes pasekondi. Zvizvarwa zvafambira mberi kubva paHBM2 kuenda kuHBM2e, HBM3, neHBM3e, chimwe nechimwe chichisimudza zvese zviri zviviri kugona nekumhanya. Kune mahombe emitauro mamodheru, ane huremu hunofanirwa kutenderedzwa nguva dzose, HBM huwandu uye bandwidth inowanzova yakakosha kupfuura komputa komputa.
Kuitwa muNyika yeChokwadi
Kubata makumi kana mazana emagigabytes ehuremu hwemhando huru yemutauro padyo neGPU kuitira kuti vagone kufambiswa panguva yega yega nhanho.
Kugonesa NVIDIA H100 uye H200 datacenter GPUs kuti isvike akawanda terabytes pasekondi yendangariro bandwidth yekudzidziswa.
Kusimbisa AI masangano ekudzidzisa uko maGPU mazhinji ega ega anovimba neHBM kudzivirira kumira pakati pematrix mashandiro.
Inotsigira yakakwirira-resolution yekugadzira mufananidzo uye vhidhiyo modhi iyo inofanirwa kufambisa yakakura activation tensor mukati nekubuda mundangariro nekukurumidza.
Ramangwana reHigh Bandwidth Memory
Memory bandwidth ikozvino inotungamira inomanikidza paAI, saka HBM iri kufambira mberi nekukurumidza. HBM3e iri kutakura mumafambisirwo emagetsi, ine HBM4 iri kumusoro ichivimbisa nzvimbo dzakafara, mastacks akareba, uye huwandu hwakawanda pasuru. Tarisira padyo dhizaini pakati pekurangarira uye pfungwa, pamwe tsika base inofa uye kugadzirisa-pedyo-yeuko, pamwe nemakwikwi anotyisa pakati pevatengesi vakaita seSK hynix, Samsung, uye Micron. Sezvo mamodheru achikura, kuwana data rakawanda padhuze nekombuta, nekukurumidza uye kune yakaderera simba, inoramba iri pakati peAI hardware kufambira mberi.
Kuzvishandisa
Nei zvichikosha
- Mutengo uye bhajeti. Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore.
- Sarudzo dzakajeka. Dzidzo yehunyanzvi inobatsira zvikwata kusarudza murwi wakakodzera, kwete iwo mutsva chete.
- Kudzora kwemhando yepamusoro. Sarudzo dzeinjiniya dziri nani dzinoderedza zviitiko zvekuvimbika mukugadzira.
Chenjera
- Kugadzirisa imwe bhenji kunogona kuvanza yakafara system kushaya simba.
- Infrastructure uye mari yekugadzirisa inowanzotarisirwa pasi.
- Chengetedzo uye kucherechedzwa mapundu anogona kukura sezvo masisitimu anowedzera kuoma.
Matanho anotevera
- Tsanangura latency, mhando, uye mutengo zvinangwa usati waitwa.
- Benchmark pasi pechokwadi mutoro uye data mamiriro.
- Chishandiso chekutarisa zvikanganiso, kudonha, uye mushandisi maitiro.
- Gadzirira nzira dzekudzosera kumashure uye dzezviitiko usati wawedzera.
Ramba Uchiongorora
Tsamba yemahara
Tora iyo yemazuva ese AI pfupiso
Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.
Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.
Zviedze iwe pachako
Take the High Bandwidth Memory quiz
Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.
Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo
Mibvunzo inowanzo bvunzwa
Chii chinonzi High Bandwidth Memory?
Yakakwira Bandwidth Memory (HBM) yakarongedzerwa ndangariro yakaiswa padyo neGPU iyo inopa data nekukurumidza zvakanyanya kupfuura yakajairika RAM. Ndizvo zvinochengeta AI accelerators kudya, kudzivirira ane simba compute cores kubva pakugara zvisina basa ivo vakamirira maremu emhando uye data.
Ndeipi dambudziko rekutanga rinoitwa Yepamusoro Bandwidth Memory kero yeAI accelerators?
AI machipi anogona kuita matiririyoni ekushanda pasekondi chete kana data yasvika nekukurumidza zvakakwana; HBM inopa iyo bandwidth kuitira kuti macores asafe nenzara.
HBM inovakwa sei mumuviri zvakasiyana neyakajairwa GDDR ndangariro?
HBM inorongedza DRAM inofira pamusoro pemumwe uye inoabatanidza nezviuru zvewaya dzakamira (TSVs), ichigadzira nzira yakafara kwazvo yedata.
Chii chinovimbwa nacho HBM kunyanya kuti iwane yakakwira bandwidth?
Panzvimbo pekuvhara nekukasira, HBM inofumura yakafararira interface (1024 bits pastack uye kumusoro) mabyte mazhinji anofambisa kamwechete.
Sei HBM masheki akaiswa pane silicon interposer padyo neGPU?
Waya mapfupi pane yakagovaniswa interposer inodzikisa simba rinodiwa pane bhiti rinotamiswa uye kuderedza latency pakati pendangariro uye compute.
Kune mhando dzemitauro mikuru chaizvo, nei HBM ichikosha zvakanyanya senge compute?
LLM inference inoramba ichiyerera hombe huremu matrices, saka ndangariro huwandu uye bandwidth kazhinji inova iyo inomisa chinhu pane compute throughput.
Ramba uchidzidza
Mazano anoenderana
Mamwe madhairekitori akasarudzirwa nyaya iyi