NVLink & NVSwitch
NVIDIA's scale-up fabric: what makes 72 GPUs one computer.
You're looking at a technology record: a building block of AI hardware — what it does, who makes it, and which chips and systems use it.
NVLink is the coherent, memory-speed link between NVIDIA GPUs — 900 GB/s per GPU in generation 4 (Hopper), 1.8 TB/s in generation 5 (Blackwell) — and NVSwitch extends it into an all-to-all domain: 8 GPUs in an HGX board, 72 in a GB200 NVL72 rack. This scale-up domain is NVIDIA's deepest moat; competitors answered with the open UALink standard. NVLink Fusion (2025) opens the fabric to third-party CPUs and accelerators on NVIDIA's terms.
Specifications & key facts
Each figure is labeled by basis: vendor spec, vendor claim, reported, estimate, or official. "Not publicly disclosed" means exactly that — we never invent values.
Overview
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Connects
Maintained by
Competes with
Related
Sources
Record reviewed as of 2026-01-15 — see Methodology.