NVIDIA NVLink is a high-bandwidth GPU interconnect. NVSwitch connects multiple NVLink-equipped GPUs within a supported system design. The link can reduce communication time for some multi-GPU workloads, but neither the presence of NVLink nor a datasheet bandwidth guarantees an application speedup. GPU model, system topology, data size and software determine what is actually achieved.
Bandwidth and domain size
NVIDIA's NVLink product specifications give these maximum bidirectional bandwidths per GPU:
| Generation | Platform | Maximum per GPU |
|---|---|---|
| NVLink 3 | Ampere A100 | 600 GB/s |
| NVLink 4 | Hopper H100 | 900 GB/s |
| NVLink 5 | Blackwell | 1,800 GB/s |
| NVLink 6 | Rubin | 3,600 GB/s, preliminary |
The generation alone does not say how many GPUs are in one domain. For example, HGX B200 is an eight-GPU baseboard with fifth-generation NVLink. GB200 NVL72 connects 72 GPUs in a rack-scale NVLink domain and lists 130 TB/s aggregate NVLink bandwidth. That aggregate sums link capacities across the rack (approximately 72 × 1.8 TB/s); it is not a per-GPU rate or a measured collective throughput.
NVIDIA specifies 3.6 TB/s bidirectional per GPU and 260 TB/s aggregate for Vera Rubin NVL72. Its own specifications label the Rubin figures preliminary. Claims about larger future domains should be tied to a particular announced system and its status, not assigned to every NVLink generation.
Inspect a Linux system
The nvidia-smi documentation provides topology and NVLink queries:
nvidia-smi topo -m
nvidia-smi topo -p2p n
nvidia-smi nvlink -stopo -m shows the reported GPU/NIC topology and path labels. topo -p2p n checks NVLink peer-to-peer capability between GPU pairs. nvlink -s reports link state and, for active links, link bandwidth. The output depends on driver, GPU and system architecture; do not infer that all GPU pairs have the same path from a product family name alone.
For achieved GPU-to-GPU communication, use a workload-level measurement. NVIDIA's nccl-tests checks collective correctness and performance; the CUDA p2pBandwidthLatencyTest sample checks peer transfers. Record GPU count, placement, message size and the units used. A peer transfer test does not by itself predict end-to-end model training or inference speed.
When NVLink helps
Tensor parallelism and expert parallelism can exchange data frequently enough to benefit from a high-bandwidth local domain. A data-parallel job may communicate at different intervals, and a workload with little GPU-to-GPU traffic may see little benefit. Communication libraries may choose different algorithms for different message sizes and topologies.
Do not change GPU power limits, driver branches or kernel settings merely to “enable NVLink.” First confirm the hardware supports it, inspect topology and measure the actual workload. For the difference between rack-local and inter-rack bandwidth, see Scale-up vs Scale-out.