UK What NVLink does
NVLink is a point-to-point interconnect that NVIDIA designed to move data between GPUs, and on some systems between a GPU and a CPU, without going through PCI Express. It first shipped on the Tesla P100 in 2016 and has been redesigned with every data centre GPU generation since.
The problem it solves is simple. A large AI model is often split across several GPUs, and those GPUs have to swap partial results many times a second. Over PCIe that traffic shares the same lanes as storage and networking and passes through the host’s root complex. NVLink gives each GPU its own wide, direct path to its neighbours, with lower latency and several times the bandwidth.
With enough NVLink bandwidth, software can treat the memory of linked GPUs almost as one pool, so a model that is too big for one card can still run at full speed across two, four or eight.
NVLink bandwidth by generation
Each generation has roughly doubled the bandwidth available to a GPU, either by adding links or by making each link faster. The figures are NVIDIA’s totals per GPU, counting both directions.
| Generation | Architecture | Introduced | Links per GPU | Bandwidth per GPU |
|---|---|---|---|---|
| NVLink 1 | Pascal (P100) | 2016 | 4 | 160 GB/s |
| NVLink 2 | Volta (V100) | 2017 | 6 | 300 GB/s |
| NVLink 3 | Ampere (A100) | 2020 | 12 | 600 GB/s |
| NVLink 4 | Hopper (H100, H200) | 2022 | 18 | 900 GB/s |
| NVLink 5 | Blackwell (B200, GB200) | 2024 | 18 | 1.8 TB/s |
| NVLink 6 | Rubin | Announced for 2026 | Not yet published | 3.6 TB/s (announced) |
| For comparison: PCIe 5.0 x16 | Any | 2019 (spec) | 16 lanes | Around 128 GB/s |
Even a single PCIe 5.0 x16 slot is several times slower than a Hopper GPU’s NVLink connection. Our guide to PCIe bandwidth by generation shows how the slot figures are worked out.
NVLink bridge vs NVSwitch
NVLink reaches GPUs in two physical ways, and the difference decides how many GPUs can talk to each other at full speed.
-
//
NVLink bridge. A short rigid connector across the top edges of two (or, on some cards, four) PCIe GPUs in neighbouring slots. It links only the cards it spans, and the bridge has to match the slot spacing of the server.
-
//
NVSwitch. A switch chip on the GPU baseboard of HGX and DGX servers. Every GPU on the board connects to the switches, so all eight GPUs can reach each other at full NVLink bandwidth at the same time.
-
//
NVLink Switch (rack scale). The same idea extended across a whole rack. In a GB200 NVL72 system, 72 Blackwell GPUs share one NVLink domain.
-
//
NVLink-C2C. A chip-to-chip version that joins the Grace CPU to a Hopper or Blackwell GPU in NVIDIA’s superchips, giving the GPU fast access to CPU memory.
Which GPUs support NVLink
NVLink support is a deliberate product decision, and it has moved away from desktop cards. Check the datasheet of the exact card, because two GPUs from the same generation can differ.
-
//
Data centre PCIe cards with bridges. The H100 PCIe and H100 NVL pair through NVLink bridges at 600 GB/s. The H200 NVL GPU takes two-way or four-way bridges at 900 GB/s, and the H100 NVL accelerator pairs with three bridges.
-
//
SXM GPUs on HGX boards. H100, H200 and B200 modules on HGX baseboards use NVSwitch rather than bridges.
-
//
Professional Ampere cards. The RTX A6000 and RTX A5000 pair through a two-way bridge at around 112.5 GB/s.
-
//
No NVLink. The L4, L40, L40S, RTX 6000 Ada and RTX PRO Blackwell cards, and every GeForce card since the RTX 30 series, have no NVLink connector. Multi-GPU work on these cards runs over PCIe.
NVLink vs PCIe: when you need it
NVLink matters when GPUs have to share work on one problem. It matters much less when each GPU runs its own job.
-
//
Training large models. Every training step ends with GPUs exchanging gradients, so link bandwidth sets how well a job scales past one card.
-
//
Inference on a model bigger than one GPU. Splitting a large language model across GPUs (tensor parallelism) sends data between them for every token generated.
-
//
Inference that fits on one GPU. Running separate copies of a smaller model on each card, or separate camera streams per GPU, needs little GPU-to-GPU traffic. PCIe is usually enough.
-
//
HPC and simulation. Codes that share large data sets between GPUs, such as CFD or molecular dynamics, gain from the extra bandwidth.
Quick answers on NVLink
The questions below come up most when people plan a multi-GPU system. For server builds, a four-GPU chassis such as the SKY-622G4 MGX GPU server supports NVLink bridges between its PCIe cards, and our data centre GPUs list shows which cards take a bridge.
NVLink: common questions
What does NVLink do?
It links GPUs directly. GPUs exchange data and read each other’s memory over NVLink instead of PCIe, so several GPUs can work on one model with far less waiting.
Is NVLink faster than PCIe?
Yes, by a wide margin. A Hopper GPU has 900 GB/s of NVLink bandwidth, against around 128 GB/s for a PCIe 5.0 x16 slot. Latency between GPUs is also lower.
Does the RTX 4090 or RTX 5090 support NVLink?
No. NVIDIA removed NVLink from GeForce cards after the RTX 30 series. Multi-GPU set-ups with these cards communicate over PCIe.
Does the RTX 6000 Ada or RTX PRO 6000 Blackwell support NVLink?
No. Neither card has an NVLink connector. Where GPU-to-GPU bandwidth matters, data centre cards such as the H200 NVL, or SXM GPUs on an HGX board, are the usual choice.
Is NVLink the same as SLI?
No. SLI was a graphics feature for splitting frame rendering between cards. NVLink is a general-purpose data link. Late GeForce cards used an NVLink bridge to carry SLI, which is why the two are often confused.
We fit data centre GPUs and their NVLink bridges in the GPU servers we build as standard, with bridge spacing, power and airflow checked for the card count before the system ships.