NVLink

NVLink is NVIDIA's proprietary high-speed interconnect for GPU-to-GPU communication within a single node. Unlike InfiniBand which connects GPUs across nodes, NVLink enables GPUs within the same server to share memory and communicate at bandwidths far exceeding PCIe. The fifth generation NVLink (NVLink 5, shipping with Blackwell) delivers 1.8 TB/s of bidirectional bandwidth, connecting all 8 GPUs in a DGX B200 into a unified memory domain via NVSwitch.

NVLink operates through NVSwitch chips that create a fully connected mesh between GPUs within a node. In the Blackwell architecture, NVLink 5 with NVSwitch 4 enables all 8 GPUs to access a combined 1.4 TB of HBM3e memory as a single addressable pool. This is critical for training large language models, where model parameters must be distributed across GPUs but accessed with minimal latency. The NVLink domain defines the practical unit of compute for many workloads — if your model fits within one NVLink domain, inter-node communication overhead drops dramatically.

NVLink configuration determines the practical compute unit. In due diligence, we evaluate whether target companies understand the distinction between NVLink-connected GPUs and PCIe-connected GPUs — a distinction that directly affects workload performance and pricing justification.