InfiniBand da RDMA Networking
InfiniBand babban haɗin gwiwa ne mai sauri, ƙarancin latency wanda ke haɗa sabobin da GPUs a cikin gungu na AI, kuma RDMA yana ƙyale injin ɗaya karanta ko rubuta ƙwaƙwalwar ajiyar wani ba tare da haɗa CPU ba.
Dubawa
Together they are the plumbing that keeps thousands of GPUs fed with data during large-model training.
Zurfafa nutsewa
Lokacin da kuke horar da samfuri a cikin dubunnan GPUs, hanyar sadarwar takan zama ƙulli, ba kwakwalwan kwamfuta ba. InfiniBand shine masana'anta da aka canza don wannan: yana ba da bandwidth ta hanyar haɗin gwiwa a cikin ɗaruruwan gigabits a sakan daya (NDR yana gudana a 400 Gb/s) da latency-microsecond. Its key trick is Remote Direct Memory Access (RDMA), which moves data directly between the memory of two nodes, bypassing the operating-system kernel and CPU copies that slow ordinary TCP/IP. Wannan 'kernel bypass' yana 'yantar da kewayon CPU kuma yana rage latti. InfiniBand kuma yana ba da ikon sarrafa kwararar kayan masarufi don masana'anta mara asara, kuma NVIDIA's Quantum switches tare da adaftar ConnectX sun mamaye manyan kwamfutocin AI. RoCE (RDMA akan Haɗin Ethernet) yana kawo fa'idodin RDMA iri ɗaya zuwa cibiyoyin sadarwar Ethernet.
Fahimtar Fasaha
RDMA yana aiki ta hanyar fi'ili da nau'ikan layi. Aikace-aikacen yana aika buƙatun aiki don aikawa da karɓar layukan layi; adaftar hanyar sadarwa (HCA) tana karanta su kuma tana canja wurin bayanai kai tsaye zuwa wuraren da aka riga aka yi rajista, maƙallan ƙwaƙwalwar ajiya akan mai watsa shiri mai nisa. Saboda NIC tana sarrafa canja wuri a cikin kayan masarufi kuma ana ƙetare kernel na OS, babu kwafin bayanai sifili kuma babu kowane fakitin CPU da ke katsewa don canja wurin girma. InfiniBand's link-Layer credit-tuus Control Control-flower yana hana zubar da ruwa, yana mai da masana'anta zama mara asara ba tare da guguwa ba.
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 InfiniBand da RDMA Networking
Bandwidth yana ci gaba da hawa: XDR InfiniBand yana kaiwa 800 Gb/s akan hanyar haɗin gwiwa, tare da taswirar hanya zuwa 1.6 Tb/s. Gasar tana ƙaruwa yayin da Ultra Ethernet Consortium ke ƙirƙira Ethernet wanda ya dace da InfiniBand don ayyukan AI, kuma kamar yadda in-Network computing (SHARP) ke sauke lissafin gama-gari cikin masu sauya kansu. Yi tsammanin haɗawar GPU-zuwa-cibiyar sadarwa mai ƙarfi, haɗin kai na gani don yanke wuta, da yadudduka masu ƙima zuwa gungu na ɗaruruwan dubban masu haɓakawa yayin da ƙirar kan iyaka ke girma.
Aiwatar da Gaskiyar Duniya
Haɗa dubunnan GPUs a cikin babban kwamfuta na AI don haka bayanan gradient suna motsawa tsakanin nodes a cikin microsecond yayin horon da aka rarraba.
Bari wani uwar garken ya karanta ƙwaƙwalwar ajiyar wani kai tsaye (RDMA) don haɓaka tsarin fayilolin da aka rarraba da kuma bayanan bayanai ba tare da CPU ba.
Gudun NCCL duk-rage ayyuka akan InfiniBand don daidaita ma'aunin ƙira a cikin gungu na GPU
Yin amfani da RoCE don kawo canjin yanayin rashin jinkiri na salon RDMA zuwa cibiyoyin cibiyar bayanai na Ethernet data kasance
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 InfiniBand and RDMA Networking 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
Feature Injiniya Bututun da Data Versioning
Tambayoyin da ake yawan yi
What is InfiniBand and RDMA Networking?
InfiniBand babban haɗin gwiwa ne mai sauri, ƙarancin latency wanda ke haɗa sabobin da GPUs a cikin gungu na AI, kuma RDMA yana ƙyale injin ɗaya karanta ko rubuta ƙwaƙwalwar ajiyar wani ba tare da haɗa CPU ba. Tare su ne aikin famfo wanda ke kiyaye dubban GPUs ciyar da bayanai yayin babban horon samfuri.
Menene ainihin RDMA ke ba da damar kwamfuta ɗaya ta yi?
Nesa Ƙwaƙwalwar Ƙwaƙwalwar Ƙwaƙwalwar Ƙwaƙwalwar Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwaƙwa tọn ne na Ƙwaƙwal na Ƙadda ) ya yi, yana ƙetare kernel na OS da kwafin CPU, wanda ke yanke latency kuma ya sake sake zagayowar CPU.
Me yasa InfiniBand ke samun ƙarancin jinkiri fiye da sadarwar TCP/IP na yau da kullun?
Adaftar InfiniBand suna canja wurin bayanai kai tsaye zuwa/daga ƙwaƙwalwar da aka ƙulla a cikin kayan aiki da ketare kwaya ta OS, suna kawar da katsewar kowane fakitin CPU da kwafin bayanai.
Kusan menene bandwidth ta hanyar haɗin gwiwa NDR InfiniBand ke bayarwa?
NDR (Bayan Bayanai na gaba) InfiniBand yana gudana a 400 Gb/s akan kowace hanyar haɗin gwiwa, tare da XDR wanda ke niyya 800 Gb/s a cikin ƙarni na gaba.
A cikin RDMA, menene 'masu layi na layi' ake amfani dasu?
Aikace-aikace suna aika buƙatun aiki don aikawa da karɓar layukan layi (biyu masu layi); adaftar tashar tashar mai masaukin baki yana karanta su kuma yana yin canja wurin ƙwaƙwalwar ajiya kai tsaye.
Menene RoCE?
RoCE yana tsaye ga RDMA akan Haɗin Ethernet, yana ba da izinin ƙarancin latency, canja wurin kernel-bypass akan daidaitaccen masana'anta na Ethernet maimakon InfiniBand na asali.