NVLink ke eng? Joang e Sebetsa, Bandwidth, le Ts'ebetso ea GPU
2026-07-31 494

Litsamaiso tse ncha tsa multi-GPU li itšetlehile ka ho feta ts'ebetso e thata ea GPU. Phetho e potlako ea processor ho fetisetsa data, ho fumana bophelo bohle, le ho sebetsa hammoho e ka ba le phello e kholo ho lebelo le akaretsang. Mqolo ona o hlalosa bandwidth ea NVLink le mehlolo, o bapisa NVLink le PCIe le NVSwitch, o koahela phihlello ea sebaka sa GPU le NVLink-C2C, 'me o bontša hore na NVLink e sebelisoa joang mesebetsing ea AI le HPC. Ho boetse ho thusa ho etsa qeto ha NVLink e ka ntlafatsa ts'ebetso le ha PCIe e ka 'na e se e le ea ho lekana.

Catalog

NVLink ke eng?

NVLink ke theknoloji ea NVIDIA ea ho hokela ka lebelo le phahameng ho fetisetsa data pakeng tsa GPUs le li-processor tse ling mefuteng ea likhomphutha tse potlakileng. Sebaka sa eona se ka sehloohong ke ho fokotsa bottlenecks ba puisano ha li-processor tse 'maloa li hlokahala ho fetisetsa data e ngata ha li sebetsa hammoho.

Mehlolo ea multi-processor, ts'ebetso e ka sitisoa ke hore data e kena ka potlako ho feta lisebelisoa. NVLink e fa tsela e phahameng ea bandwidth, e fokolang latency ea puisano e fokotsang mekoloko le ho thusa li-processor tse hoketsoeng ho fetisetsa data ka mokhoa o motle.

Joang NVLink e Sebetsa?

Figure 2. NVLink GPU-to-GPU Communication Path

NVLink e crea tsela e potlakileng ea puisano pakeng tsa li-GPU tse lumellanang le li-processor tse ling. Ha li-GPU tse peli li hlokahala ho fetisetsa data, phihlello ea peer-to-peer e ka letsetsa data ho tsamaisa ka kotloloho lipakeng tsa libaka tsa bona tsa memory ho fapana le ho romela liphetho tsohle ho CPU memory.

NVLink e tšehetsa puisano e lebisang bobele, kahoo data e ka tsamaisoa ka bobeli ho li-processor tse hokitsoeng. Li-NVLink tse ngata li ka sebetsa hammoho, li fa tsela e pharaletseng ho li-transfer tsa data ho processor ho processor.

NVLink e fa tsela ea 'mele bakeng sa puisano ena, ha software e etsa qeto hore na le lebaka lefe le ts'ebeliso e lokela ho sebetsa. CUDA, li-driver tsa lisebelisoa, NCCL, le kopo li laola liphetho. Ha phihlello ea peer e ts'ehetsoa le ho bonts'oa, GPU e le 'ngoe e ka bala ho, ngola ho, kapa ho kopitsa data ho memory ea GPU e 'ngoe ka NVLink e fumanehang.

Mehlolo le Bandwidth ea NVLink

Bandwidth ea NVLink e fumaneha e ntse e ba e phahame ka mor'a mekhahlelo e meholo ea GPU ea NVIDIA. Genere ea pele ea NVLink e hlahile le Tesla P100 e thehiloeng ho Pascal, e tšehetsa likhokelo tse 'ne tsa NVLink le ho fihla ho 160 GB/s ea total bidirectional bandwidth ka GPU. Volta V100 e eketsa sena ho likhokelo tse tšeletseng le 300 GB/s ka GPU.

Genere ea borai ea NVLink e ile ea fihla le Ampere A100. A100 e tšehetsa likhokelo tse 12 tsa NVLink le ho fihla ho 600 GB/s ea total bidirectional bandwidth ka GPU. Hopper H100 e eketsa palo ea likhokelo ho ea ho 18, e eketsa bandwidth ea NVLink ea borai ho 900 GB/s ka GPU.

Genere ea 5 e fihla le Blackwell, e eketsa total bidirectional bandwidth ho 1.8 TB/s ka GPU. Genere ea 6, e amanang le platform ea Rubin, e eketsa mofuta o makholo ho 36 le ho fihla ho 3.6 TB/s ka GPU.

NVLink Generation	NVIDIA Architecture	Mohlala GPU kapa Platform	Mofuta o phahameng Link ka GPU	Mofuta o phahameng Bidirectional Bandwidth ka GPU
NVLink 1	Pascal	P100	4	160 GB/s
NVLink 2	Volta	V100	6	300 GB/s
NVLink 3	Ampere	A100	12	600 GB/s
NVLink 4	Hopper	H100	18	900 GB/s
NVLink 5	Blackwell	Blackwell GPUs	18	1.8 TB/s
NVLink 6	Rubin	Rubin platform	36	3.6 TB/s

Per-Link khahlanong le Total NVLink Bandwidth

Bandwidth ka link le bandwidth e akaretsang ka GPU e hlalosa boleng bo fapaneng. GPU e ka ba le litlhakore tse ngata tsa NVLink, kahoo bandwidth e akaretsang ka GPU e kopanya bandwidth e fumanehang ho tloha ho linx tse ngata tse tšehetsang.

Mohlala, H100 e na le 18 NVLink 4 litlhakore. Link e 'ngoe le e 'ngoe e fa 50 GB/s ea bidirectional bandwidth, e fanang ka boholo bo phahameng ba 900 GB/s ka GPU. Blackwell e sebelisa 18 NVLink 5 litlhakore ka 100 GB/s bidirectional bandwidth ka link, e hlahisa ho fihlela ho 1.8 TB/s ka GPU.

Linomoro tsena li hlahisa bandwidth e phahameng ea interface. Bandwidth e 'nile e sebelisoa pakeng tsa GPUs tse peli e ka ba tlase ho latela palo ea linx tse sebetsa lipakeng tsa tsona, sebopeho sa sistimi, le konfigurable ya hardware. Ha o bala lintlha tsa NVLink, kamehla sheba hore na boleng bo ama link e le 'ngoe, GPU e le 'ngoe, kapa mokhoa oo o felletseng oa NVLink.

NVLink vs PCIe vs NVSwitch

Setšoantšo 3. NVLink vs PCIe vs NVSwitch Tsamaiso

NVLink, PCIe, le NVSwitch li tsamaisa data ka hare ho mekgwa ea GPU, empa li sebetsa ka mesebetsi e fapaneng. PCIe ke interface e amanang le ts'ebeliso e fapaneng e kopanyang GPUs le lisebelisoa tse ling le sistimi e amohelehang. NVLink e fa puisano e potlakileng lipakeng tsa li-processor tse tšehetsang NVIDIA, ha NVSwitch e sebetsa le NVLink ho hokela lihlopha tse kholo tsa GPUs ka har'a sebopeho se arolelanoang.

Feature	NVLink	PCIe	NVSwitch
Seo e leng sona	Phutheho e potlakileng ea processor	Sistimi e amanang le ts'ebeliso e fapaneng	Switching fabric e sebelisitsoeng le NVLink
Sepheo se ka sehloohong	Puisano e potlakileng lipakeng tsa GPUs le processors tse tšehetsang	Hokela GPUs, SSDs, NICs, le lisebelisoa tse ling ho sistimi	Hokela GPUs tse ngata tse nang le NVLink ka har'a sebopeho se le seng
Puisano pakeng tsa GPU le GPU	E fa puisano e bohale pakeng tsa batho bao ba etselitsoeng hammoho	E tšehetsa liphetohelo pakeng tsa GPU le GPU, empa ts'ebetso e itšetlemeha holim'a sebopeho sa PCIe	E ruta lipalo tsa NVLink lipakeng tsa GPUs tse ngata
Data path	Lintlha tse tobileng tsa processor kapa likhokahano ka NVSwitch	Sephethephethe se tsamaisoa ka har'a PCIe root complexes le switches	E fa litsela tse switched lipakeng tsa GPUs tse amanang le NVLink
Latency	E entsoe bakeng sa puisano e tlase ea processor	E loketse bakeng sa I/O e amanang le ts'ebeliso, empa litsela tsa GPU ho GPU li ka kenyelletsa mekgwa e mengata ea sistimi	E kenyelletsa switching ha e ntse e boloka GPUs li amanang ka litsela tse phahameng tsa NVLink
Topology	Likhokelo tse tobileng ho tloha ho e 'ngoe ho ea ho e 'ngoe kapa likhokelo ho NVSwitch	CPU root complexes le PCIe switches	Switched, e amanang haholo le GPU fabric
Compatibility	E hloka lisebelisoa tse tšehetsang tse amanang le NVIDIA	Ts'ehetso e pharalletseng ea indasteri	E hloka lisebelisoa tse tšehetsang tse amanang le NVIDIA NVLink platforms
E loketse ka ho fetisisa bakeng sa	Mechini e amanang le puisano heavi multi-GPU	Mechini e le 'ngoe ea GPU kapa mesebetsi e nang le sephetho se lekanyelitsoeng sa GPU ho GPU	Mechini e meholo moo GPUs tse ngata li hlokang puisano e ngata
Melemo e meholo	E lekanyelitsoe ho lisebelisoa le mekgwa e amanang le NVIDIA	E ka ba bottleneck ho mesebetsi e amanang le inter-GPU e boima	E eketsa cost ea hardware le bohlokwa ba sistimi

Ka mantsoe a bonolo, PCIe e hokela lisebelisoa ho sistimi, NVLink e fa puisano e potlakileng lipakeng tsa processor le processor, le NVSwitch e atolosa NVLink ho mekgwa e meholo ea GPU. NVSwitch ha e na ho theha NVLink hobane e itšetlemeha ho likhokahano tsa NVLink ho tsamaisa sephethephethe sa GPU.

PCIe e ka ba le bokhoni ha puisano pakeng tsa GPU le GPU e lekanyelitsoe. NVLink e ba molemo ha GPUs li fetisang data khafetsa, ha NVSwitch e amanang haholo ha GPUs tse ngata tse amanang le NVLink li hloka puisano e amohelehang ka har'a sistimi e le 'ngoe.

Na NVLink e kopanya Memory ea GPU?

NVLink ha e kopanye ka boitsebiso VRAM ea GPUs tse 'maloa ho pool e kholo ea memori. GPUs tse peli tse nang le 80 GB ea memori ngoe le ngoe li ntse li na le libaka tse fapaneng tse 80 GB tsa memori eseng ho fetoha GPU e le 'ngoe e nang le 160 GB ea VRAM ea lehae.

Ha lisebelisoa le mekhahlelo li tšehetsa, CUDA peer-to-peer mekhahlelo ea mehopolo e fa GPU e 'ngoe monyetla oa ho fumana mehopolo ho GPU e' ngoe. GPU e ka bala, e ngola, kapa e kopitsa data ho peer memory ntle le hore e qale ho fetisetsa data ka CPU memory. NVLink e fana ka tsela e phahameng ea bandwidth bakeng sa transfer ena ha li-GPUs li amanang ka eona.

Remote GPU memory, empa, e lula e amanngoa le GPU e 'ngoe. Software e ntse e etsa qeto hore na data e bolokoa kae, GPU e fumana data ena, le hore na data e lokela ho kopitsoa kapa ho fumanoa ka lisebelisoa tse ling. CUDA Unified Memory e ka nolofalletsa ho fumana mehopolo ho tloha ho lisebelisoa, empa ha e qhealthse karohano ea 'mele pakeng tsa GPU memory.

Ka lebaka lena, ho beheng mehopolo ho bohlokoa hanyane ka mekhahlelo e mengata ea GPU. Ho fumana mehopolo ea GPU e holehang kamehla ho ka ama ts'ebetso le fa NVLink e fana ka khokahano e potlakileng.

Mokhoa oo NVLink-C2C e extending NVLink ho Chip-to-Chip Connections

Setšoantšo sa 4. NVLink-C2C Khokahano Pakeng tsa CPU le GPU

NVLink-C2C e fa NVIDIA NVLink ho tsoa ho khokahanong ea processor level ho likhokahano tse khutšoane tsa chip-to-chip ka har'a package kapa superchip. E ka hokahanya CPUs, GPUs, le mefuta e meng e amanang ea processor dies, e fa monyetla oa ho fetisetsa data ka khokahano e amanang haholo.

Lipplatform tsa NVLink-C2C tse tšehetsang le tsona li fumaneha ho processors tse amanang li ka boloka maikutlo a tšoanang ka mehopolo. Mohlala, Grace Hopper e sebelisa NVLink-C2C pakeng tsa Grace CPU le Hopper GPU, e lumellang li-processor ka bobeli ho fumana mehopolo ho tsamaisoa ha ho amanang ha mehopolo ea bona e lula e arohane.

Phapang e kholo ke hore na theknoloji e 'ngoe le e' ngoe e sebelisoa kae. Standard NVLink e tloaelehileng e amohela li-GPUs tse arohaneng kapa e hokahanya li-GPU ho NVSwitch ho maemo a borai kapa tsamaiso. NVLink-C2C e beha mohopolo o tšoanang oa ho hokahana ka lebelo le leholo haufi le processors ka boeona, e etsa hore e be e loketseng bakeng sa CPU ho GPU, CPU ho CPU, le mekhahlelo e meng e amanang ka chip-to-chip.

NVLink bakeng sa AI le Multi-GPU Workloads

NVLink e bohlokoa haholo ha li-GPU tse 'maloa li sebetsa ka mesebetsi e le 'ngoe 'me li batla ho fetohela data khafetsa. Melecheng ena, ts'ebetso ha e ntse e ipapisitse le speed ea GPU empa hape le hore na data e tsamaea ka potlako pakeng tsa accelerators.

Nakong ea AI training, li-GPU li ka fetohela gradients, parameters, activations, le liphetho tse mahareng. NVIDIA NCCL e laola litlhahiso tse ngata tsa transfer ka mekgwa e joalo ka AllReduce, AllGather, ReduceScatter, Broadcast, le puisano e amanang.

Tensor parallelism e hlahisa puisano e ngata hobane likarolo tse ling tsa calculations e sebetsa ho li-GPU tse fapaneng 'me liphetho tse khutšoanyane li tlameha ho fetoloa pele ts'ebetso e ka tsoela pele. Model parallelism le eona e fetola data pakeng tsa li-GPU ha likarolo tse fapaneng tsa mohlala li sebetsa.

Multi-GPU LLM inference e na le lits'ebetso tse tšoanang ha mohlala o moholo o khetholloa ho li-GPU tse 'maloa. Ha GPU e le 'ngoe e tlameha ho emela data ho tsoa ho e 'ngoe, mekhahlelo ea puisano e ka fokotsa throughput e akaretsang.

HPC workloads tse kang simulation ea saense, computational fluid dynamics, le modelling ea molecular le tsona li fetola data ea mekhahlelo, liphetho tse khutšoanyane, le tlhahisoleseding ea synchronization pakeng tsa li-GPU. Liphetho tse potlakileng tse pakeng tsa GPU ho GPU li ka fokotsa nako ea ho emela le thuso ho etsa hore mosebetsi o behe ka mokhoa o khutšoanyane.

Bakeng sa mecha e bobebe ea puisano, NVLink e ka fokotsa nako e sebelisoang ke li-GPU ho fetola data le thuso mecha ea GPU e mengata e atleha ka tsela e betere.

Nako e 'nang le NVLink

NVLink e loketse ha li-GPU tse 'maloa li sebetsa ka mosebetsi o le mong 'me li batla ho fetola data khafetsa. E ka thusa ha puisano pakeng tsa GPU le GPU e fokotsa mosebetsi 'me tsamaiso e tšehetsa NVLink.

NVLink e khetheha ha:

• Li-GPU tse 'maloa li sebetsa hammoho ho workload e le 'ngoe

• Mecha e meholo ea data e tsamaea pakeng tsa li-GPU

• Puisano ea GPU e fokotsa ts'ebetso

• Li-GPU le platform li tšehetsa NVLink

• Software e ka sebelisa puisano e amanang ho li-GPU tse 'maloa

PCIe e ka ba le leseli ha:

• Tsamaiso e sebelisa GPU e le 'ngoe feela

• Li-GPU tse ngata li sebetsana le mesebetsi e arohaneng

• Sepe se amanang pakeng tsa GPU se tlase

• PCIe e se e entse ts'ebetso e hlokahalang

• Theko e tlase le tšehetso e pharaletseng ea lisebelisoa li bohlokoa haholo

Theko le moralo oa tsamaiso li boetse li bohlokoa. Mechahalo e nang le NVLink e ka hloka li-GPU tse tšehetsang, moralo o nepahetseng oa tsamaiso, ho fetola ho eketsehileng, le matla le ho pholosa haholo le mekhahlelo e meholo.

Ka kakaretso, NVLink e molemo ka ho fetisisa ha puisano pakeng tsa li-GPU e le moeli o hlakileng oa ts'ebetso. Ha li-GPU li ngata ha li fetohe data kapa PCIe e se e fa bandwidth e lekaletseng, NVLink e ka eketsa theko le qhaphoho ntle le site tse hlakileng tsa ts'ebetso.