What is NVLink?
NVLink is NVIDIA's proprietary interconnect for connecting GPUs directly to each other, or a GPU to a CPU, with much higher bandwidth than a regular PCIe slot provides. In a data-center server with several discrete GPUs, NVLink is what lets those GPUs exchange data fast enough to work together on one model as if they were closer to a single larger device than several separate cards talking over a comparatively slow bus. It's the piece of infrastructure that makes techniques like tensor parallelism practical at data-center scale, since splitting a matrix multiplication across GPUs only pays off if those GPUs can exchange partial results quickly.
Why PCIe isn't enough for GPU-to-GPU traffic
PCIe is a general-purpose bus, shared by all the devices on a system and not designed specifically for GPU-to-GPU communication. When several GPUs are jointly serving one model, splitting either its weights or its computation, they need to constantly exchange partial results, and doing that over PCIe becomes a bottleneck once the GPU count and traffic go up. NVLink is a dedicated, direct link built for exactly this: point-to-point bandwidth between GPUs that's a large multiple of what PCIe offers, which is why data-center GPU pods intended to serve large models as one coherent 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 on the same superchip package, which is what the DGX Spark's GB10 uses to link its Grace CPU and Blackwell GPU so they can address one shared 128GB memory pool, the mechanism behind unified memory as described on our unified memory post. It's easy to conflate the two because they share a name and a lineage, but one links GPUs to GPUs across a multi-GPU system, the other links a CPU to a GPU within one chip package.
Where this leaves a single DGX Spark
A single Spark is one GB10 superchip with one GPU. There's no second discrete GPU on the same board to connect via NVLink, so NVLink in the classic multi-GPU sense isn't part of a single Spark's own architecture, and it's worth being direct about that rather than overstating it. NVLink becomes relevant to GPUwerk customers in two ways: understanding why larger data-center GPU clusters, the kind you'd otherwise be renting by the hour from a hyperscaler, are built the way they are, and when connecting multiple Sparks into a cluster, though that link between separate Spark units is a network connection rather than NVLink. See our two-node cluster guide and scaling to a fleet guide for how multi-Spark setups are actually networked and what that means for throughput compared to an NVLink-connected GPU pod.