InfiniBand na RDMA ịkparịta ụka n'Ịntanet
InfiniBand bụ njikọ njikọ ọsọ ọsọ dị ala nke na-ejikọ sava na GPU na ụyọkọ AI, yana RDMA na-ahapụ otu igwe ka ọ gụọ ma ọ bụ dee ebe nchekwa ọzọ na-etinyeghị CPU.
Nchịkọta
Together they are the plumbing that keeps thousands of GPUs fed with data during large-model training.
Ime miri emi
Mgbe ị na-azụ ihe nlereanya gafere puku kwuru puku GPU, netwọkụ na-abụkarị ihe mgbochi, ọ bụghị ibe. InfiniBand bụ ihe agbanwere akwa ebumnuche-wuru maka nke a: ọ na-enye bandwidth ọ bụla na narị gigabits kwa nkeji (NDR na-agba ọsọ na 400 Gb/s) na latency microsecond-scale. Isi aghụghọ ya bụ Nweta ebe nchekwa Direct Direct (RDMA), nke na-ebugharị data ozugbo n'etiti ebe nchekwa ọnụ abụọ, na-agafe kernel sistemụ na mbipụta CPU nke na-ebelata TCP/IP nkịtị. Nke a 'kernel bypass' na-ahapụ okirikiri CPU ma belata latency. InfiniBand na-enyekwa njikwa ọsọ ngwaike maka akwa na-enweghị efu, yana NVIDIA's Quantum switches yana ihe nkwụnye ConnectX na-achịkwa AI supercomputers. RoCE (RDMA n'elu Converged Ethernet) na-eweta uru RDMA yiri ya na netwọk Ethernet.
Nghọta nka nka
RDMA na-arụ ọrụ site na ngwaa na ụzọ kwụ n'ahịrị. Ngwa biputere arịrịọ ọrụ izipu na ịnata kwụ n'ahịrị; ihe nkwụnye netwọkụ (HCA) na-agụ ha ma na-ebufe data ozugbo na mpaghara ebe nchekwa edebanyere aha na nke etinyere na ebe nchekwa dịpụrụ adịpụ. N'ihi na NIC na-ejikwa mbufe na ngwaike na os kernel na-agafe agafe, enwere nnomi data efu na enweghị ngwugwu CPU ọ bụla maka mbufe nnukwu. InfiniBand's njikọ-layer kredit njikwa eruba kredit na-egbochi oke njupụta, na-eme ka akwa ahụ ghara ịla n'iyi na-enweghị oke ifufe mbufe.
Mmetụta atụmatụ
Ọnụ ego na mmefu ego
Mkpebi ihe owuwu ụlọ na-akwalite arụmọrụ yana ọnụ ahịa ọrụ ruo ọtụtụ afọ.
Mkpebi doro anya
Nkà mmụta nka na-enyere ndị otu egwuregwu aka ịhọrọ nchịkọta ziri ezi, ọ bụghị naanị nke kachasị ọhụrụ.
Quality akara
Nhọrọ injinia ka mma na-ebelata ihe omume ntụkwasị obi na mmepụta.
Ọdịnihu nke InfiniBand na netwọk RDMA
Bandwidth na-arịgo: XDR InfiniBand na-achọ 800 Gb/s kwa njikọ, yana maapụ ụzọ 1.6 Tb/s. Asọmpi na-akawanye njọ ka Ultra Ethernet Consortium na-echepụta Ethernet nke dabara InfiniBand maka ibu ọrụ AI, yana dị ka netwọk netwọk (SHARP) na-ebupụ mgbakọ na mwepụ mkpokọta n'ime ndị na-agbanwe onwe ha. Na-atụ anya nchikota GPU-na-netwọk siri ike, njikọ njikọ anya iji belata ike, na akwa gbagoro ruo ụyọkọ narị puku kwuru puku ngwa ngwa ka ụdị oke na-etolite.
Mmejuputa n'ezie n'ụwa
Ijikọ puku kwuru puku GPUs na AI supercomputer ka data gradient na-agagharị n'etiti ọnụ na microseconds n'oge ọzụzụ ekesa.
Ikwe ka otu ihe nkesa gụọ ebe nchekwa onye ọzọ ozugbo (RDMA) iji mee ka sistemụ faịlụ na ọdụ data kesara ngwa ngwa na-enweghị CPU n'elu.
Na-agba ọsọ NCCL na-ebelata arụmọrụ n'elu InfiniBand iji mekọrịta ihe atụ n'ofe ụyọkọ GPU.
Iji RoCE weta nnyefe dị ala ụdị RDMA gaa na netwọk data etiti Ethernet dị
Ihe ize ndụ & okporo ụzọ nche
Ịkwalite otu akara ngosi nwere ike zoo adịghị ike sistemụ sara mbara.
A na-eledakarị ihe akụrụngwa na ụgwọ ọrụ anya.
Ọdịiche nchekwa na nleba anya nwere ike itolite ka sistemu na-adịwanye mgbagwoju anya.
Map mmejuputa
Kọwaa latency, ịdịmma na ebumnuche ọnụ ahịa tupu mmejuputa ya.
Benchmark n'okpuru ibu dị adị na ọnọdụ data.
Nleba anya akụrụngwa maka mperi, ịkpafu na mmetụta onye ọrụ.
Kwadebe ụzọ nzaghachi azụghachi azụ na ihe omume tupu ịchachaa.
Nọgide na-eme nchọpụta
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
Ntuziaka na-esote
Pipeline injinia na ụdị data njiri mara
Ajụjụ a na-ajụkarị
What is InfiniBand and RDMA Networking?
InfiniBand bụ njikọ njikọ ọsọ ọsọ dị ala nke na-ejikọ sava na GPU na ụyọkọ AI, yana RDMA na-ahapụ otu igwe ka ọ gụọ ma ọ bụ dee ebe nchekwa ọzọ na-etinyeghị CPU. Ha ọnụ bụ ọkpọkọ na-eme ka ọtụtụ puku GPU na-enye data nri n'oge ọzụzụ ihe nlereanya buru ibu.
Kedu ihe RDMA na-enye ohere ka otu kọmputa mee?
Nweta ebe nchekwa Direct Direct na-akpali data kwụ ọtọ n'etiti ebe nchekwa ọnụ abụọ, na-agafe kernel OS na mbipụta CPU, nke na-ebelata latency ma na-ahapụ okirikiri CPU.
Kedu ihe kpatara InfiniBand ji enweta latency dị ala karịa netwọk TCP/IP nkịtị?
Ihe nkwụnye InfiniBand na-ebufe data ozugbo gaa na/site na ebe nchekwa etinyere na ngwaike wee gafee kernel OS, na-ewepụ nkwụsịtụ CPU kwa otu na nnomi data.
Ihe dị ka bandwit ọ bụla NDR InfiniBand na-enye?
NDR (ọnụego data na-esote) InfiniBand na-agba ọsọ na 400 Gb/s kwa njikọ, yana XDR lekwasịrị anya 800 Gb/s na ọgbọ na-abịa.
Na RDMA, kedu ihe a na-eji 'panyị kwụ n'ahịrị'?
Ngwa biputere arịrịọ ọrụ izipu na ịnata kwụ n'ahịrị (ụzọ kwụ n'ahịrị); ihe nkwụnye ọwa onye ọbịa na-agụ ha ma na-enyefe ebe nchekwa ozugbo.
Kedu ihe bụ RoCE?
RoCE na-anọchi anya RDMA n'elu Converged Ethernet, na-enye ohere ohere dị ala, kernel-bypass na-ebufe na akwa Ethernet ọkọlọtọ kama ịbụ InfiniBand nke ala.