Blog/What is NVLink?
Definition

What is NVLink?

By Samuel Seidel · Published September 9, 2026

NVLink is NVIDIA's proprietary interconnect that links GPUs directly to each other, or a GPU to a CPU, at much higher bandwidth than a PCIe slot. In a data-center server with several discrete GPUs, NVLink lets them exchange data fast enough to work on one model almost as a single larger device, instead of as separate cards talking over a slower bus. It is what makes tensor parallelism practical at data-center scale: splitting a matrix multiplication across GPUs only pays off if they can swap partial results quickly.

Why PCIe isn't enough for GPU-to-GPU traffic

PCIe is a general-purpose bus shared by every device in the system, not a link designed for GPU-to-GPU traffic. GPUs that jointly serve one model, by splitting its weights or its computation, exchange partial results constantly, and over PCIe that becomes the bottleneck as GPU count and traffic grow.

NVLink is a direct link built for this job, with point-to-point bandwidth many times what PCIe offers. That is why data-center GPU pods meant to serve large models as one unit are built around NVLink rather than PCIe alone.

NVLink versus NVLink-C2C

NVLink and NVLink-C2C (chip-to-chip) are related technologies with different jobs. NVLink connects separate discrete GPUs to each other. NVLink-C2C connects a CPU and a GPU inside one superchip package. The DGX Spark's GB10 uses NVLink-C2C to link its Grace CPU and Blackwell GPU so both address one shared 128 GB memory pool, the mechanism behind unified memory. The shared name invites confusion, but one links GPU to GPU across a multi-GPU system, the other links CPU to GPU within one package.

Where this leaves a single DGX Spark

A single Spark is one GB10 superchip with one GPU. There is no second GPU on the board to connect, so NVLink in the classic multi-GPU sense is not part of a single Spark's architecture.

For someone renting Sparks on GPUwerk, NVLink matters in two ways. It explains why large data-center GPU clusters, the kind you would otherwise rent by the hour from a hyperscaler, are built the way they are. And it sets expectations for a linked pair of Sparks: the link between two separate Spark units is a network connection, not NVLink. The two-node cluster guide and the scaling to a fleet guide show how multi-Spark setups are networked and what that means for throughput compared with an NVLink-connected GPU pod.

Related pages

First top-up: pay $20, get $30 in credit

One dedicated GB10 superchip, no shared tenancy.

One GPU and one 128 GB memory pool, linked over NVLink-C2C, from $0.79/hour.

Create account