Citation and evidence

NVLink

21 min full readUpdated 21 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI ComputeAI HardwareData CentersNVIDIA

Cite this article

NVLink is NVIDIA's proprietary high-bandwidth, low-latency, cache-coherent point-to-point interconnect for directly connecting GPUs to other GPUs and, in some configurations, to host CPUs. Introduced in 2014 and first shipped in 2016 with the Tesla P100, it reached 1.8 TB/s of aggregate bidirectional bandwidth per GPU in its fifth generation (Blackwell, 2024), roughly fourteen times the bandwidth of PCIe Gen5 x16.[6][16] A sixth generation, NVLink 6, arrives with NVIDIA's Vera Rubin platform in 2026; NVIDIA quoted 3.6 TB/s per GPU at its January 2026 launch, while its current specification page lists a preliminary 3 TB/s.[19][20] Across five generations NVLink has become the dominant scale-up fabric inside modern AI servers, supercomputers, and rack-scale systems such as the DGX GB200 NVL72, where it ties 72 GPUs into a single 130 TB/s memory domain.[6] The companion NVSwitch chip is the crossbar that connects many NVLink endpoints together, and in 2025 NVIDIA extended the technology to third-party silicon through NVLink Fusion.[16]

NVLink was created to overcome the bandwidth and latency limitations of PCI Express in multi-GPU systems.[12] NVLink is the physical interconnect; the related NVSwitch chip is the crossbar that connects many NVLink endpoints together inside a node or a rack, and NVLink-C2C ("chip-to-chip") is a die-to-die variant used to bond NVIDIA's Grace CPU to Hopper or Blackwell GPUs into a single coherent superchip.[5] NVIDIA Collective Communications Library (NCCL) collectives such as all-reduce, all-gather, and reduce-scatter run directly on top of NVLink whenever GPUs in a distributed training job share an NVLink fabric, which is why most large GPU-based LLM training runs depend on it.

NVLink exists to move tensors between GPUs fast enough that large neural networks can be trained and served across many accelerators as if they were one device. Modern frontier models do not fit on a single GPU, so the work is sharded across many GPUs using a mix of data parallelism, tensor parallelism, pipeline parallelism, and expert parallelism. Each of those strategies generates heavy GPU-to-GPU communication (all-reduce, all-gather, reduce-scatter), and NVLink is the fabric that carries it inside a server or rack at hundreds of gigabytes per second per GPU, far beyond what PCIe can sustain. The detailed mechanics of why each parallelism strategy needs NVLink are covered in the sections below.

"NVIDIA NVLink High-Speed Interconnect: Maximizes throughput for Superior Application Performance", official video from NVIDIA, published November 14, 2014. Watch on YouTube.

Background and motivation

By the early 2010s, the dominant host interconnect for accelerators was PCI Express. PCIe Gen3 x16, the standard interface for GPUs through the Maxwell generation, delivered roughly 16 GB/s per direction (about 32 GB/s aggregate). For a single accelerator that was acceptable, but for the kinds of multi-GPU servers that deep learning was beginning to demand, PCIe became the bottleneck. Gradient exchange, parameter synchronization, and activation transfer between GPUs all traveled across the same PCIe root complex and, often, across the host CPU. As models grew, the time spent shuffling tensors across PCIe began to dominate the time spent on compute.

NVIDIA announced NVLink in March 2014 as an answer to this problem.[12] The design goals were straightforward: much higher bandwidth than PCIe, lower latency, support for direct GPU-to-GPU memory access, and a cache-coherent memory model so that multiple GPUs could share data without the software overhead of explicit copies. The first commercial implementation arrived two years later inside the Tesla P100 and the original DGX-1 system.[8]

Five generations of NVLink shipped through the Blackwell era (a sixth, NVLink 6, arrives with Vera Rubin in 2026[19][20]), each tied to a specific NVIDIA GPU computing microarchitecture.[12] The table below summarizes the per-link signaling rate, the data rate per direction per link, the number of links exposed by the flagship data-center GPU of that generation, and the resulting aggregate bidirectional bandwidth per GPU.

GenerationYearMicroarchitectureFlagship GPUPer-link signalingData rate per direction per linkLinks per GPUAggregate bandwidth per GPU
NVLink 1.02016PascalTesla P10020 GT/s20 GB/s4160 GB/s
NVLink 2.02017VoltaTesla V10025 GT/s25 GB/s6300 GB/s
NVLink 3.02020AmpereA10050 GT/s25 GB/s12600 GB/s
NVLink 4.02022HopperH10050 GT/s25 GB/s18900 GB/s
NVLink 5.02024BlackwellB100 / B200 / GB200100 GT/s50 GB/s181,800 GB/s

Expanded article table

A few conventions are worth flagging because the marketing numbers can be confusing. NVIDIA generally quotes "aggregate bandwidth" as the bidirectional sum across all links, so a 900 GB/s H100 is really 450 GB/s in each direction.[7] Per-link bandwidth is also typically the bidirectional figure: a single NVLink 4.0 link is 25 GB/s in each direction, often written as 50 GB/s bidirectional. Generations 3.0 and 4.0 deliver the same 25 GB/s per direction per link at 50 GT/s, but NVLink 4.0 switches from NRZ to PAM4 signaling, so each differential pair carries twice the data.[3][12] The improvement from A100 to H100 came from raising the link count from 12 to 18 and from cutting the number of differential pairs per link in half, which let NVIDIA pack more links into roughly the same die area.[3] The jump to NVLink 5.0 doubled per-link bandwidth from 50 GB/s to 100 GB/s while holding the link count at 18, which is how Blackwell reaches 1.8 TB/s, exactly twice the H100.[16]

NVLink 1.0 debuted on the GP100 die inside the Tesla P100. Each link carried 20 GB/s in each direction, and each P100 exposed four links, giving 80 GB/s per direction or 160 GB/s aggregate per GPU.[8] That was already roughly five times the bandwidth of PCIe Gen3 x16 and made it possible to build the first DGX-1 with eight P100s wired together in a hybrid cube-mesh topology.[8] NVLink 1.0 was also the version IBM integrated into POWER8+ CPUs, which is how the Summit supercomputer's predecessors first put NVLink directly between a CPU and a GPU instead of going through PCIe.[12]

Volta's V100 stepped the per-link signaling from 20 GT/s to 25 GT/s and increased the link count from four to six. The result was 25 GB/s per direction per link and 300 GB/s aggregate per GPU.[1] NVLink 2.0 also introduced cache coherence with IBM POWER9, which is the configuration Oak Ridge's Summit (two POWER9 CPUs and six V100 GPUs per node) and Lawrence Livermore's Sierra (two POWER9 CPUs and four V100 GPUs per node) supercomputers used, with NVLink 2.0 connecting CPUs and GPUs in a shared coherent address space.[12]

Ampere's A100 doubled the per-pair signaling rate to 50 Gbit/s while halving the number of pairs per link, then doubled the link count from six to twelve. Each NVLink 3.0 link still delivered 25 GB/s per direction, but with twelve of them per A100 the aggregate bandwidth reached 600 GB/s.[12] The A100 generation also introduced second-generation NVSwitch, which NVIDIA used inside the DGX A100 to connect eight A100s in an all-to-all topology where every pair of GPUs gets the full 600 GB/s of bisection bandwidth.[12]

Hopper's H100 kept the 50 GT/s symbol rate but moved to PAM4 signaling with half as many pairs per link, and pushed the link count from twelve to eighteen. Each H100 SXM5 module exposes 18 NVLink 4.0 links for 900 GB/s of aggregate bandwidth, which NVIDIA consistently advertises as roughly seven times the bandwidth of PCIe Gen5 x16 (128 GB/s).[2][7] The H100 NVL variant, which is a PCIe-form-factor product designed for inference of large language models, exposes 600 GB/s of NVLink instead of 900 GB/s.[7] Hopper paired with third-generation NVSwitch, whose total switching capacity rose to 13.6 Tbit/s from 7.2 Tbit/s in the prior generation, and added in-network reduction (the SHARP accelerator) so that all-reduce can finish inside the switch fabric rather than touching every endpoint.[3][10]

Blackwell's B100, B200, and the dual-die GB200 superchip use fifth-generation NVLink. The signaling rate doubles to 100 GT/s, and the per-direction bandwidth per link doubles correspondingly to 50 GB/s. With 18 links per GPU, aggregate bandwidth reaches 1.8 TB/s, twice that of an H100 and roughly fourteen times PCIe Gen5 x16.[6][16] NVIDIA describes a single Blackwell GPU as supporting "up to 18 NVLink 100 gigabyte-per-second (GB/s) connections for a total bandwidth of 1.8 terabytes per second (TB/s)."[18] NVLink 5.0 is the foundation of the GB200 NVL72 rack, which NVIDIA described at launch as "the largest NVIDIA NVLink domain ever offered."[6] Through the NVLink Switch System, fifth-generation NVLink can be extended beyond a single rack to connect up to 576 GPUs in one non-blocking domain with 1 petabyte per second of total aggregate bandwidth, according to NVIDIA's 2025 specifications.[18]

NVSwitch

NVLink by itself is a point-to-point link. To connect more than a handful of GPUs in a fully connected, high-bandwidth fashion you need a switch, and that switch is NVSwitch. NVSwitch first appeared in 2018 inside the DGX-2, where 16 V100 GPUs were tied together by twelve NVSwitch chips, giving every GPU full 300 GB/s of NVLink bandwidth to every other GPU in the system.[9] That was the first time a 16-GPU server behaved, from the application's perspective, as a single shared-memory accelerator.[9]

NVSwitch generations track NVLink generations but are numbered separately. First-generation NVSwitch shipped with V100 in DGX-2.[9] Second-generation NVSwitch shipped with A100 in DGX A100, and used six switch chips per system rather than twelve.[12] Third-generation NVSwitch shipped with H100 in DGX H100 and HGX H100 baseboards, with 13.6 Tbit/s per chip and the SHARP in-network reduction engine.[3][10] Fourth-generation NVSwitch shipped with Blackwell in 2024, with each switch chip exposing 72 NVLink 5.0 ports.[6]

The most important thing NVSwitch enables is non-blocking all-to-all bandwidth. Without a switch, an N-GPU system has to share its NVLink budget across every pair, which limits both the topology and the bisection bandwidth. With NVSwitch, every GPU can talk to every other GPU at full link bandwidth simultaneously, which is exactly what NCCL collectives like all-reduce and all-gather want.[9]

NVIDIA's DGX line has been the reference design for NVLink-based servers from the beginning, and the DGX generations roughly map onto the NVLink generations.[13]

SystemYearGPUsNVLink generationTopologyNotable detail
DGX-1 (P100)20168 x P1001.0Hybrid cube meshFirst commercial NVLink server
DGX-1 (V100)20178 x V1002.0Hybrid cube meshFirst V100 system, no NVSwitch
DGX-2201816 x V1002.0Full NVSwitch fabricFirst NVSwitch system, 2.4 TB/s bisection
DGX A10020208 x A1003.0Full NVSwitch fabricSix 2nd-gen NVSwitch chips
DGX H10020228 x H1004.0Full NVSwitch fabric3.6 TB/s bisection, SHARP reductions
DGX GB200 NVL72202472 x B2005.0Rack-scale NVSwitch130 TB/s NVLink domain bandwidth

Expanded article table

The DGX-1 with eight P100s used a hybrid cube-mesh because the GPUs only had four NVLinks each, so a fully connected topology was not possible without a switch.[8] The DGX-2 was the first system to behave as a single 16-GPU device, and it produced 2.4 TB/s of bisection bandwidth and 75 GB/s of all-reduce bandwidth.[9][10] The DGX A100 dropped back to eight GPUs but used the much larger A100 NVLink budget to give every pair the full 600 GB/s.[12] The DGX H100 raised that to 3.6 TB/s bisection and 450 GB/s of all-reduce.[10]

The big jump arrived with the GB200 NVL72. Instead of stopping at eight GPUs in a single chassis, NVIDIA used fifth-generation NVLink and fourth-generation NVSwitch to extend the NVLink domain across an entire liquid-cooled rack.[6] Seventy-two Blackwell GPUs and 36 Grace CPUs sit in eighteen compute trays, with nine NVLink Switch trays providing the all-to-all fabric.[6] The result is a single NVLink domain with 130 TB/s of aggregate GPU communication bandwidth, which NVIDIA treats as the basic unit of an "AI factory" rack.[6] Before NVL72, mainstream NVLink servers topped out at eight GPUs (DGX H100), although the 16-GPU DGX-2 and Hopper's NVLink Switch System, which could join up to 256 H100 GPUs across 32 nodes, had already extended NVLink beyond a single eight-GPU board.[9][3] NVL72 made a 72-GPU NVLink domain the standard rack-level building block.

NVLink-C2C is a die-to-die variant of NVLink built specifically to bond a Grace CPU to a Hopper or Blackwell GPU on the same package or board. It first appeared in the Grace Hopper Superchip (GH200) and now anchors the GB200 superchip as well.[4] NVLink-C2C delivers 900 GB/s of bidirectional bandwidth between the CPU and GPU, which NVIDIA describes as roughly seven times the bandwidth of x16 PCIe Gen5.[4] It is also dramatically more energy-efficient: 1.3 picojoules per bit transferred, which NVIDIA cites as more than five times better than PCIe Gen5.[11]

The more important property is coherence. NVLink-C2C is memory-coherent, meaning the GPU and the CPU share a single, hardware-managed address space.[4] In NVIDIA's words, "Memory coherency enables you to transfer only the data you need, and not migrate entire pages to and from the GPU," with CPU and GPU threads concurrently and transparently accessing both CPU- and GPU-resident memory.[4] CPU threads and GPU threads can both access either CPU-resident or GPU-resident memory transparently, and atomic operations cross the boundary natively.[4] For model parallelism and large-model inference this matters because the GPU can spill embedding tables, KV caches, and other oversized data structures into the CPU's much larger LPDDR5X memory pool without paying PCIe transfer costs every time it needs them.[4]

NVLink Fusion is NVIDIA's program, announced May 18, 2025 at Computex, that licenses NVLink interconnect technology to third parties so they can integrate NVLink directly into their own custom silicon and connect it to NVIDIA GPUs and CPUs.[16] At launch the first custom-silicon partners were MediaTek, Marvell, Alchip Technologies, Astera Labs, Synopsys, and Cadence, while Fujitsu and Qualcomm Technologies were named as CPU partners whose processors can be paired with NVIDIA GPUs over NVLink.[16] The goal is to let customers build semi-custom AI infrastructure in which custom CPUs or accelerators participate in NVLink-based clusters alongside NVIDIA GPUs, rather than being limited to PCIe.[16] NVIDIA CEO Jensen Huang framed the launch in sweeping terms: "A tectonic shift is underway: for the first time in decades, data centers must be fundamentally rearchitected, AI is being fused into every computing platform."[16] NVLink Fusion uses the same fifth-generation NVLink that delivers 1.8 TB/s per GPU, and NVIDIA positions it as a way for cloud providers to scale out AI factories toward millions of GPUs.[16]

Cache coherence and the memory model

NVLink is more than a fast pipe. From the second generation onward, links can carry coherent memory traffic, which means that two GPUs (or a CPU and a GPU) connected by NVLink behave like nodes in a coherent NUMA system rather than two separate devices that happen to be wired together.[12] A GPU can issue a load to a memory address that physically lives in another GPU's HBM, and the cache fabric will resolve it without explicit DMA. CUDA programs see this as Unified Virtual Addressing and, with newer GPUs, as full Unified Memory across the NVLink domain.

This matters for two reasons. First, it lets the CUDA runtime move data lazily, only fetching cache lines when they are actually touched, which is much more efficient than pre-staging entire tensors. Second, it lets compilers and frameworks treat the NVLink domain as one big memory pool. Tensor parallelism, in particular, depends on this: when a single matrix multiply is sharded across eight GPUs, each GPU has to read partial results from every other GPU during the all-reduce that follows, and the cost of that all-reduce is dominated by NVLink bandwidth.[14]

PCIe is still the standard host interconnect, and every NVIDIA data-center GPU also exposes a PCIe interface for talking to the host CPU and to NICs. But for GPU-to-GPU communication inside a server, NVLink wins by an enormous margin. The headline contrast: PCIe Gen5 x16 tops out around 128 GB/s aggregate, while NVLink 5.0 delivers 1,800 GB/s per GPU, roughly fourteen times more.[16]

InterconnectPer-direction bandwidthAggregate bandwidthNotes
PCIe Gen3 x16~16 GB/s~32 GB/sP100-era host link
PCIe Gen4 x16~32 GB/s~64 GB/sA100-era host link
PCIe Gen5 x16~64 GB/s~128 GB/sH100 / Blackwell host link
NVLink 4.0 (H100)450 GB/s900 GB/sAbout 7x PCIe Gen5 x16
NVLink 5.0 (B200)900 GB/s1,800 GB/sAbout 14x PCIe Gen5 x16
NVLink-C2C (GH200)450 GB/s900 GB/sCPU-to-GPU, coherent

Expanded article table

The gap matters most for collectives. A ring all-reduce of a 100 GB gradient buffer moves roughly twice that volume through each GPU, which is about three seconds of pure transfer time at PCIe Gen5 x16's ~64 GB/s per direction but only about a fifth of a second at NVLink 5.0's 900 GB/s per direction. When collectives like this run every training step, that gap compounds.

The headline reason NVLink exists today is large-model training. Modern frontier language models are too big to fit on a single GPU, so training (and increasingly inference) is sharded across many GPUs using a mix of data parallelism, tensor parallelism, pipeline parallelism, and expert parallelism.[14] The communication patterns these strategies generate are very different in their bandwidth requirements.

Tensor parallelism is the most NVLink-hungry. It splits a single matrix multiply across N GPUs and requires an all-reduce after every transformer layer's attention and MLP blocks.[14] The size of the activations being reduced is large (gigabytes per layer for big models) and the operation has to complete before the next layer can start. In practice this means tensor-parallel groups must live inside a single NVLink domain.[14] For H100-based clusters that means tensor parallelism is bounded at eight (one DGX H100 node); for GB200 NVL72 it can go up to seventy-two without leaving the NVLink fabric.[6] That is the single biggest reason hyperscalers care about NVL72: it lets them use much larger tensor-parallel degrees, which in turn lets them serve trillion-parameter models with low latency. NVIDIA claims the 72-GPU NVLink domain delivers up to 30x faster real-time trillion-parameter LLM inference than the prior generation.[6]

Pipeline parallelism is gentler on the network because the only thing that crosses GPUs is the activation tensor between pipeline stages, but it still benefits from NVLink because pipeline bubbles get smaller as inter-stage latency drops. Data parallelism uses all-reduce on gradients once per step, and FSDP uses all-gather on parameters before each forward pass; both are friendlier to InfiniBand-class inter-node networks but still benefit from NVLink for the intra-node portion of the collective.

NCCL, NVIDIA's collective communication library, is what actually drives the wires. NCCL automatically discovers the topology, picks the best available transport (NVLink, NVSwitch, PCIe, InfiniBand, or RoCE), and chooses an algorithm (ring, tree, or SHARP) for each collective. When NCCL detects an NVLink fabric it prefers it over everything else, and on NVL72 it will try to route as much of the all-reduce as possible through the in-switch SHARP engine.[10]

The effect on training is large enough to be visible in published numbers. The Megatron-LM paper from NVIDIA, Stanford, and Microsoft Research reports weak scaling of GPT models from 1 billion to 1 trillion parameters on 32 to 3,072 A100 GPUs at 44 to 52 percent of peak FLOPS, and finds tensor parallelism works best within a single NVLink-connected DGX A100 server because of its expensive all-reduce communication.[14] Many frontier labs do not publish their parallelism configurations in detail, and not every frontier model runs on NVLink (Google trains Gemini on its own TPUs), but on GPU clusters the NVLink-connected server or rack is the usual unit of tensor parallelism.

NVLink is proprietary, and that has been both its greatest strength and the main complaint against it. Because NVIDIA controls the spec, it can iterate quickly: each generation has roughly doubled aggregate bandwidth, and the addition of NVSwitch and SHARP has happened on NVIDIA's own schedule rather than a standards-body schedule. The cost is that there is no second source. If you want NVLink, you have to buy NVIDIA GPUs, NVIDIA switches, and (increasingly) NVIDIA networking. The H100 PCIe NVL variant, which exposes only 600 GB/s instead of 900 GB/s, is a reminder that even NVIDIA's own product segmentation can hide NVLink behind extra cost.[7] The 2025 NVLink Fusion program softens this slightly by licensing NVLink to partner silicon, but the fabric, switches, and software ecosystem remain NVIDIA's.[16]

Competitors have responded with their own scale-up fabrics. AMD's Infinity Fabric connects its Instinct MI300X and MI325X GPUs at up to 896 GB/s aggregate per GPU on Infinity Fabric 4. Intel's Xe Link does the same job for its Ponte Vecchio (Data Center GPU Max) accelerators, while Intel's Gaudi chips scale up over integrated RoCE Ethernet ports. None of these is interchangeable with NVLink at the silicon level, and as of 2025 none matched the NVL72 rack-scale topology; AMD's 72-GPU Helios rack, built on the MI400 series, is its answer for 2026.[21]

The more interesting development is UALink (Ultra Accelerator Link), an open standard whose consortium was incorporated in 2024 with founding members including AMD, Intel, Broadcom, Cisco, Google, HPE, Meta, and Microsoft.[15] The UALink 200G 1.0 specification was ratified on April 8, 2025, defining an open scale-up fabric that can connect up to 1,024 accelerators with direct read, write, and atomic semantics at 200 Gbit/s per lane.[17] Its explicit goal is to be an open alternative to NVLink and NVSwitch.[15] Whether UALink ships in volume, and whether it can match NVIDIA's bandwidth and software ecosystem, is one of the open questions in AI infrastructure for the next several years.

One practical limitation worth noting: NVIDIA removed NVLink from its consumer Ada Lovelace GeForce cards in 2022, and Blackwell consumer cards (RTX 50 series) likewise do not expose NVLink.[12] NVLink is now strictly a data-center technology. Hobbyists who want to pool VRAM across multiple GPUs at home cannot do it the way they could with the GeForce RTX 3090, the RTX A6000, and earlier Titan and Quadro cards.

Continue reading

More from See also

Topics:AI ComputeAI HardwareData CentersNVIDIA

References

  1. ^Nvidia, "NVIDIA DGX-1 With Tesla V100 System Architecture" white paper. images.nvidia.com/...m-architecture-whitepaper.pdf
  2. ^Nvidia, "NVIDIA H100 Tensor Core GPU Architecture" white paper, March 2022. web.archive.org/...gtc22-whitepaper-hopper.pdf
  3. ^1 ^2 ^3 ^4 ^5Nvidia Developer Blog, "NVIDIA Hopper Architecture In-Depth." developer.nvidia.com/...pper-architecture-in-depth
  4. ^1 ^2 ^3 ^4 ^5 ^6Nvidia Developer Blog, "NVIDIA Grace Hopper Superchip Architecture In-Depth." developer.nvidia.com/...chip-architecture-in-depth
  5. ^Nvidia, "NVLink-C2C: Chip Interconnect Technology." nvidia.com/...nvlink-c2c
  6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Nvidia, "GB200 NVL72." nvidia.com/...gb200-nvl72
  7. ^1 ^2 ^3 ^4Nvidia, "H100 GPU." nvidia.com/...h100
  8. ^1 ^2 ^3 ^4Nvidia Developer Blog, "NVIDIA DGX-1: The Fastest Deep Learning System." developer.nvidia.com/...stest-deep-learning-system
  9. ^1 ^2 ^3 ^4 ^5 ^6Alben et al., "NVSwitch and DGX-2," Hot Chips 30, 2018. old.hotchips.org/...HotChips2018_DGX2NVS_Final.pdf
  10. ^1 ^2 ^3 ^4 ^5Ishii et al., "The NVLink-Network Switch," Hot Chips 34, 2022. hc34.hotchips.org/...ch%20HotChips%202022%20r5.pdf
  11. ^Sury et al., "NVLink-C2C: A Coherent Off Package Chip-to-Chip Interconnect with 40Gbps/pin Single-ended Signaling," ISSCC 2023. ieeexplore.ieee.org/...10067395
  12. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Wikipedia, "NVLink." en.wikipedia.org/...NVLink
  13. ^Wikipedia, "Nvidia DGX." en.wikipedia.org/...Nvidia_DGX
  14. ^1 ^2 ^3 ^4 ^5Narayanan et al., "Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM," SC 2021. people.eecs.berkeley.edu/...sc_megatron_lm.pdf
  15. ^1 ^2UALink Consortium and reporting in The Next Platform, "UALink Fires First GPU Interconnect Salvo at Nvidia NVSwitch," April 2025. nextplatform.com/...nnect-salvo-at-nvidia-nvswitch
  16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11NVIDIA Newsroom, "NVIDIA Unveils NVLink Fusion for Industry to Build Semi-Custom AI Infrastructure With NVIDIA Partner Ecosystem," May 18, 2025. nvidianews.nvidia.com/...ructure-partner-ecosystem
  17. ^UALink Consortium, "UALink 200G 1.0 Specification Overview," blog post, May 15, 2025. ualinkconsortium.org/...specification-overview-802
  18. ^1 ^2Nvidia, "NVLink and NVLink Switch" (archived March 2025). web.archive.org/...nvlink
  19. ^1 ^2Nvidia, "NVLink and NVLink Switch." nvidia.com/...nvlink
  20. ^1 ^2Nvidia Developer Blog, "Inside the NVIDIA Rubin Platform: Six New Chips, One AI Supercomputer," January 5, 2026. developer.nvidia.com/...chips-one-ai-supercomputer
  21. ^TechPowerUp, "AMD Previews 432 GB HBM4 Instinct MI400 GPUs and Helios Rack-Scale AI Solution," June 13, 2025. techpowerup.com/...d-helios-rack-scale-ai-solution

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

8 revisions · v8 · 4,146 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check of the full article against its cited sources (120 claims, 23 sources fetched); 5 material + 20 minor fixed

Cite this page: AI Wiki. "NVLink." aiwiki.ai, updated 21 Jun 2026, fact-checked 2 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/nvlink

Edit article