Komawa Labarai
SamfuraAI Understanding takaitaccen bayani

Rahoton eTeknix Nvidia ya gabatar da gine-ginen ƙwaƙwalwar ajiya na NVHBM don masu haɓaka AI

eTeknix ya ba da rahoton cewa Nvidia ya gabatar da NVHBM, ƙirar ƙwaƙwalwar ajiya mai girma da aka yi niyya don inganta haɓakar bandwidth da ƙarfin ƙarfi don manyan tsarin AI. Fitar ta ce Amazon yana shirin amfani da shi a cikin abubuwan more rayuwa na Trainium4 na gaba, yayin da Nvidia ke tsammanin Feynman GPUs ɗin sa su ɗauki ƙirar a cikin 2028.

6 min readRead the linked source
Source-provided image accompanying eTeknix reports Nvidia introduced NVHBM memory architecture for AI accelerators
Tushen tusheAn rubuta tushen tushe
Mawallafi
eteknix.com
Tushen hanyar haɗin gwiwa
eteknix.comhttps://www.eteknix.com/nvidia-introduces-its-own-nvhbm-memory-for-artificial-intelligence/
Nau'in tushe
Tushen da aka haɗa - ba a kafa matsayin tushen farko ba.
An kuma ambata

Labari na ƙarshe da aka bita

MaganaFahimtar wannan a cikin daƙiƙa 60

Fara a nan

Mabuɗin sharuddan

Ƙwaƙwalwar ajiya (Agent Memory)
Mahallin da aka adana wani wakilin AI yana amfani da matakai ko zaman don inganta ci gaba.
Alamar alama
Daidaitaccen gwaji ko saitin bayanai da aka yi amfani da shi don aunawa da kwatanta aikin ƙira.
Inference
Lokaci lokacin aiki inda ƙwararren ƙirar ke haifar da tsinkaya ko fitarwa.
Gwada kankaAI Model An Bayyana Tambayoyi

Me ya canza tun bayan bugawa

  1. An fara bugawa
  2. Wannan fasaha ta fasaha ta NVIDIA Sabunta Mai Haɓakawa ta zahiri tana haɓaka NVLink Fusion iri ɗaya da taron NVHBM na al'ada wanda sabuntawar canonical ke wakilta. Yana ƙara dalla-dalla da'awar kamfani game da gine-ginen NVHBM da yuwuwar fa'idodin: har zuwa 30% ƙarin bandwidth ƙwaƙwalwar ajiya fiye da daidaitaccen HBM4e, har zuwa 67% ƙananan PHY da yanki na tallafi, har zuwa 30% ƙarin siliki na babban-mutu don ƙididdigewa ko wasu fasalulluka, har zuwa 15% ƙananan amfani da wutar lantarki na HBM, da da'awar 30% ƙarshen-zuwa-ƙarshen aiki tare da haɓaka aikin NVX.
  3. Wannan shine taron samfurin NVHBM iri ɗaya wanda aka rufe ta sabuntawar Nvidia NVLink Fusion na canonical. eTeknix yana ƙara da'awar aikin da aka ruwaito na fiye da 30% mafi girman bandwidth, 15% ƙananan amfani da wutar lantarki, har zuwa 25% tanadin processor-silicon da har zuwa 80% dawo da fakiti-silicon, tare da rahotannin tsare-tsaren tallafi na gaba wanda ya shafi Amazon Trainium4 da Nvidia's 2028 Feynman architecture Waɗannan cikakkun bayanai ana danganta su ga eTeknix kuma ba a tabbatar da kansu a cikin kayan da aka kawo ba.

Me ya faru

eTeknix ya ba da rahoton cewa Nvidia ya gabatar da NVHBM, babban gine-ginen ƙwaƙwalwar ajiyar bandwidth wanda aka tsara tare da Amazon's Annapurna Labs. Fitar ta ce ƙirar tana motsa mai sarrafa ƙwaƙwalwar ajiya daga babban mai haɓakawa zuwa tushen fakitin ƙwaƙwalwar ajiya kuma an yi niyya don tsarin AI tare da manyan samfura.

eTeknix ya ba da rahoton cewa Nvidia ya gabatar da NVHBM, babban ginin ƙwaƙwalwar ajiyar bandwidth don ayyukan aikin fasaha na fasaha. Labarin ya ce Nvidia ya haɓaka ƙirar tare da Annapurna Labs, sashin guntu na Amazon. Makasudin da aka ruwaito shine nau'in tsarin AI wanda samfurinsa ya ƙunshi tiriliyoyin sigogi, inda motsin bayanai da sauri tsakanin masu sarrafawa da ƙwaƙwalwar ajiya na iya iyakance aiki da ƙara yawan amfani da wutar lantarki. Madogararsa ta gabatar da NVHBM azaman fasahar ababen more rayuwa don masu haɓaka AI maimakon azaman samfuran ƙwaƙwalwar mabukaci gaba ɗaya. Ba ya samar da farashin samfur, kwanan watan aikawa ko cikakken ƙayyadaddun fasaha.

Dangane da eTeknix, zaɓin ƙirar ƙira shine sanya mai sarrafa ƙwaƙwalwar ajiya kai tsaye a cikin tushe ya mutu na ƙwaƙwalwar bandwidth mai girma maimakon ajiye shi a cikin babban injin sarrafawa, wanda labarin ke kira XPU. Rahoton ya yarda cewa sanya mai sarrafawa a cikin tushe ba shine keɓaɓɓen Nvidia ba kuma ya ce Samsung, SK Hynix da Micron suna binciko hanyoyin irin wannan. Nvidia ta aiwatar da iƙirarin, duk da haka, an gabatar da shi azaman hanya don rage adadin dabarar mai sarrafawa da ake buƙata akan mai haɓakawa da kunkuntar ƙirar ƙwaƙwalwar ajiya ta zahiri. Majiyar ta ce wannan na iya sauƙaƙa mai shiga tsakani, ɓangaren fakitin da ke haɗa na'ura da ƙwaƙwalwar ajiya.

Fitar ta ba da rahoton alkaluma da yawa da aka dangana ga Nvidia: fiye da 30% ƙarin bandwidth da 15% ƙananan amfani da wutar lantarki idan aka kwatanta da HBM4E. Hakanan ya ce motsi mai sarrafawa na iya 'yantar da kusan kashi 25% na babban yanki na silicon guntu, yayin da rage girman fa'idar zai iya murmurewa kusan 80% na silicon mai amfani a cikin kunshin. Waɗannan alkalumman da'awar kamfani ne kamar yadda eTeknix ya ruwaito; kayan da aka kawo ya ƙunshi wani maƙasudin ma'auni mai zaman kansa, hanyar gwaji, samfurin samfur ko tabbaci na ɓangare na uku. eTeknix ya kara ba da rahoton cewa Nvidia yana son daidaita ƙayyadaddun ƙayyadaddun ƙayyadaddun ƙayyadaddun ƙayyadaddun bayanai a cikin masu yin ƙwaƙwalwar ajiya da yawa don haka abokan haɗin gwiwa a cikin yanayin yanayin halittar NVLink Fusion na iya gwadawa da haɗa masu haɓakawa na al'ada cikin sauƙi.

eTeknix ya ce Amazon yana cikin kamfanoni na farko da ke shirin ɗaukar NVHBM. Labarin ya ba da rahoton cewa Amazon yana da niyyar haɗa ginin ƙwaƙwalwar ajiya tare da haɗin gwiwar NVLink na Nvidia a cikin na'urori masu sarrafa AWS Trainium4 na gaba, ƙirƙirar tsarin gine-ginen matakin rack wanda ke kawo kwakwalwan Amazon da Nvidia GPUs tare. A gefen Nvidia, rahoton ya ce ana sa ran NVHBM zai fara farawa tare da tsarin gine-gine na Feynman GPU, wanda aka tsara don 2028. Waɗannan shirye-shirye ne na gaba kamar yadda aka bayyana ta hanyar fita. Rahoton bai tabbatar da samuwa na yau ba, kayan aikin da aka kammala, tura abokin ciniki ko ingantaccen jadawalin ƙaddamar da kasuwanci.

Bayanan tushe: eteknix.com ↗

Me ya sa yake da mahimmanci

Zane-zanen da aka bayar da rahoton ya yi niyya ga ƙuntatawar ababen more rayuwa na tsakiya a cikin AI: motsi bayanai tsakanin ƙwaƙwalwar ajiya da na'urori masu sarrafawa yadda ya kamata. Idan an sami nasarorin da aka ruwaito a samarwa, NVHBM na iya ba masu ƙirar haɓaka ƙarin bandwidth, ƙananan amfani da wutar lantarki da ƙarin yanki na silicon don albarkatun lissafin.

Ƙwaƙwalwar ajiya ya zama yanki mai iyakancewa na ƙirar haɓaka AI saboda masu sarrafawa suna iya aiki da kyau kawai lokacin da bayanan da suke buƙata ya isa da sauri. Rahoton eTeknix ya tsara NVHBM a matsayin ƙoƙari na magance wannan matsala a matakin kunshin. Haɗin haɗin da ake da'awar na babban bandwidth da ƙananan amfani da wutar lantarki zai fi mahimmanci a cikin manyan horo da tsarin ƙima, inda yawancin masu haɓakawa ke aiki tare da motsin ƙwaƙwalwar ajiya na iya cinye albarkatu masu yawa. Madogararsa ba ta ƙididdige yadda waɗannan haɓakawa da ake da'awar za su fassara zuwa horon ƙirar ƙira cikin sauri, ƙarancin farashin aiki ko mafi kyawun sabis na mai amfani na ƙarshe.

Ajiye silikon da aka ruwaito zai iya shafar yadda aka ƙirƙira abubuwan haɓakawa na gaba. Idan har zuwa 25% na yankin mai sarrafawa ba a ƙara buƙata don mai sarrafa ƙwaƙwalwar ajiya, masu zanen kaya na iya yuwuwar keɓance wannan sararin don ƙarin albarkatun kwamfuta, kodayake eTeknix baya ba da rahoton takamaiman guntu da ke yin haka. Da'awarta cewa kusan kashi 80% na fakitin siliki mai amfani za'a iya dawo dasu haka kuma yana bayyana ɗakin ɗakin ƙira, ba ingantaccen sakamako ba. Ƙimar aiki za ta dogara ne akan adadin yankin da aka dawo da shi a zahiri za a iya amfani da shi, ko kunshin ya kasance mai araha kuma ko ana iya samar da ƙwaƙwalwar ajiya a ma'aunin da ake buƙata.

Tsarin daidaitawa yana da mahimmanci saboda ƙwaƙwalwar haɓaka tana da amfani kawai lokacin da na'urori masu sarrafawa, marufi, firmware da tsarin masana'antu ke aiki tare. eTeknix ya ba da rahoton cewa Nvidia yana son masu yin ƙwaƙwalwar ajiya da yawa don tallafawa ƙayyadaddun NVHBM na kowa, wanda zai iya rage aikin haɗin gwiwa don kamfanoni na gina haɓakar al'ada ta hanyar NVLink Fusion. Haɗin Amazon da aka ƙaddamar zai ƙara mahimmanci fiye da na Nvidia GPUs na kansa idan Trainium4 ƙarshe ya haɗu da hanyar sadarwar Nvidia ko abubuwan haɓakawa. Har yanzu, rahoton da aka kawo ba ya rubuta yarjejeniya tare da masu samar da ƙwaƙwalwar ajiya mai suna, kammala gwajin hulɗar aiki ko alƙawura daga abokan cinikin girgije fiye da ɗaukan Amazon na gaba wanda ya bayyana.

Saboda haka labarin ya shafi fiye da sabon lakabin ƙwaƙwalwar ajiya. Ya bayyana Nvidia yana ƙoƙarin tsara wani muhimmin Layer na tarin kayan aikin AI yayin haɗa wannan Layer zuwa yanayin yanayin haɗin kai da kuma taswirar processor na Amazon. Wannan na iya yin tasiri kan yadda aka gina tsarin AI na al'ada da nawa iko Nvidia ke riƙe akan abubuwan da ke kewaye. Tasirin jama'a ya kasance mai yiwuwa: labarin yana ba da wata shaida cewa NVHBM har yanzu ta canza ƙarfin girgije da ake samu, farashin mai haɓakawa, amfani da makamashi ko samun dama ga ayyukan AI.

Interactive Mechanism

Ingantacciyar hanyar sadarwa: Yadda A zahiri yake Aiki

Bincika fasahar da ke bayan wannan ci gaban ta hanyar mu'amala.

Document Size:128K tokens
Needle Placement Depth (Location in document):50% into text
Attention Context Buffer Map:
Target Fact (50%)
Equivalent Pages~320Standard book pages
Retrieval Accuracy99.9%Needle recall score
RAM / KV Cache5.1 GBMemory overhead
Prompt CachingActive~80% discount on reuse
Core takeaway: Million-token context windows allow querying whole codebases or legal archives in one prompt. However, KV cache memory scales with context length, making prompt caching crucial for real-time production.
Duba ra'ayi na hulɗa+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

Abin kallo na gaba

Babban abubuwan da ba a sani ba sune wadatar kasuwanci, abokan masana'antu, gwajin aiki mai zaman kansa da lokacin turawa. eTeknix ya ce Amazon yana shirin haɗa Trainium4 na gaba kuma Nvidia yana tsammanin fitowar Feynman ta 2028, amma rahoton da aka kawo bai tabbatar da cewa ko dai samfurin na jigilar kaya ba ko kuma an tabbatar da fa'idodin da ake da'awar.

Batun farko da za a kallo shine ko Nvidia ya buga cikakken bayanin NVHBM kuma ko masu binciken kayan masarufi masu zaman kansu ko abokan ciniki sun gwada adadin bandwidth da aka ruwaito. Rahoton da aka kawo yana ba da kwatancen HBM4E amma baya kwatanta nauyin aiki, ƙarfin ƙwaƙwalwar ajiya, yanayin aiki ko hanyoyin aunawa. Idan ba tare da waɗannan cikakkun bayanai ba, ba za a iya amfani da kaso ba don tsinkayar ribar da ake samu a zahiri don wani samfuri ko cibiyar bayanai. Gwajin mai zaman kansa kuma zai fayyace ko haɓakar sun shafi gabaɗaya ko kuma kawai ga tsararrun da aka zaɓa kawai.

Batu na biyu shi ne wadata da aiki tare. An bayyana shirin Nvidia, kamar yadda eTeknix ya ruwaito, shine don daidaita NVHBM a cikin masu yin ƙwaƙwalwar ajiya da yawa, amma labarin bai ambaci sunayen masu ba da gudummawa ba ko gano lokutan masana'antu. Samun tallafi na gaba zai dogara ne akan ko waɗancan kamfanonin suna samar da sassa masu jituwa a cikin girma kuma ko iyawar marufi na iya ci gaba da tafiya tare da buƙatar haɓaka AI. Hakanan zai zama mahimmanci don ganin ko abokan haɗin gwiwar masu haɓakawa na al'ada za su iya amfani da ƙira ba tare da dogaro da takamaiman musaya na Nvidia ko hanyoyin takaddun shaida ba.

Batu na uku shine lokacin da aka ruwaito taswirar hanyoyin samfur. eTeknix ya ce Amazon yana shirin haɗa NVHBM da NVLink a cikin masu sarrafawa na Trainium4 na gaba kuma Nvidia yana tsammanin fasahar ta bayyana tare da Feynman GPUs a cikin 2028. Rahoton bai faɗi lokacin da Trainium4 zai ƙaddamar da shi ba, ko samfuran kamfanonin biyu za su kasance tare, ko kuma 202 na tsammanin zai kasance tare. Don haka ya kamata a kula da waɗancan kwanakin azaman tsare-tsaren sa ido maimakon samuwa na yanzu.

A ƙarshe, masu karatu ya kamata su kalli shaidar ainihin tasirin turawa: ƙananan amfani da wutar lantarki a matakin tsarin, mafi girman kayan aiki don ƙayyadaddun ayyukan AI, canje-canje a cikin haɓakar haɓakawa ko rage hadaddun kunshin. Babu ɗayan waɗannan sakamakon da aka kafa ta wannan labarin. Abubuwan da ba a sani ba masu ma'ana sun haɗa da farashi, yawan amfanin ƙasa, yanayin zafi, ƙarfin ƙwaƙwalwar ajiya, dacewa tare da tsarin da ake da su da kuma gwargwadon yadda sauran ƙirar ƙira-ƙira ta mutuƙar masu kaya ke kaiwa kasuwa. Har sai an amsa waɗannan tambayoyin, NVHBM sakamako ne na jagorar kayan aikin da aka ruwaito, amma fa'idar sa har yanzu ba ta tabbata ba.

Jagorori masu alaƙa & tambayoyin tambayoyi

AI Model ya bayyanaAI horoMasu canjiMakomar AIGwada abin da kuka sani - gwada gwajin AI kyautaNemo kalmar AI a cikin ƙamus ɗin muBi samfurin AI na sakin tracker

Sabuntawa da gyare-gyare

Ana sabunta wannan labarin na canonical a wurin lokacin da abubuwan haɓakawa suka canza ta zahiri. URL ɗin sa da ainihin ranar bugawa ba sa canzawa.

  • Wannan shine taron samfurin NVHBM iri ɗaya wanda aka rufe ta sabuntawar Nvidia NVLink Fusion na canonical. eTeknix yana ƙara da'awar aikin da aka ruwaito na fiye da 30% mafi girman bandwidth, 15% ƙananan amfani da wutar lantarki, har zuwa 25% tanadin processor-silicon da har zuwa 80% dawo da fakiti-silicon, tare da rahotannin tsare-tsaren tallafi na gaba wanda ya shafi Amazon Trainium4 da Nvidia's 2028 Feynman architecture Waɗannan cikakkun bayanai ana danganta su ga eTeknix kuma ba a tabbatar da kansu a cikin kayan da aka kawo ba.
  • Wannan fasaha ta fasaha ta NVIDIA Sabunta Mai Haɓakawa ta zahiri tana haɓaka NVLink Fusion iri ɗaya da taron NVHBM na al'ada wanda sabuntawar canonical ke wakilta. Yana ƙara dalla-dalla da'awar kamfani game da gine-ginen NVHBM da yuwuwar fa'idodin: har zuwa 30% ƙarin bandwidth ƙwaƙwalwar ajiya fiye da daidaitaccen HBM4e, har zuwa 67% ƙananan PHY da yanki na tallafi, har zuwa 30% ƙarin siliki na babban-mutu don ƙididdigewa ko wasu fasalulluka, har zuwa 15% ƙananan amfani da wutar lantarki na HBM, da da'awar 30% ƙarshen-zuwa-ƙarshen aiki tare da haɓaka aikin NVX.
Duba log ɗin gyaran jama'a
An sami wannan yana da amfani?