NVIDIA NVLink Switch 5
900 GB/s
Fields
id
interconnect/nvswitch-5Kind
nvswitch
Version
NVLink Switch 5
Bandwidth (GB/s, per link, unidirectional)
900
Bandwidth basis
PER GPU, PER DIRECTION. NVIDIA's spec table gives 'NVLink Switch GPU-to-GPU Bandwidth 1,800 GB/s' - BIDIRECTIONAL AGGREGATE per GPU pair; recorded per-direction value is half, 900 GB/s. Total aggregate published for the 72-GPU NVL72 configuration is 130 TB/s. A switch fabric and a direct-attached NVLink 5 link deliver the same per-GPU rate; only the domain size changes.
Link count
72
Topology
Switched non-blocking fabric supporting both 8-GPU and 72-GPU NVLink domains (GB200/GB300 NVL72).
Scale-up
yes
Scale-out
no
Switching
Nine-switch-class rack-scale domain with SHARP in-network reduction and multicast. NVIDIA states NVLink scale-out lets NVL72 support 9x the GPU count of a single 8-GPU system while keeping communication in balance with compute.
notes
Blackwell generation. The step change is domain size, not link speed: 8 -> 72 GPUs in one non-blocking domain. Inference consequence: a 70B-class model at 4-bit fits comfortably in one 8-GPU domain, so TP degree beyond 8 buys little and instead adds all-reduce traffic that scales with TP degree - the NVL72 domain is aimed at MoE expert parallelism (many low-bandwidth, latency-tolerant all-to-all transfers) and trillion-parameter weights rather than at more TP. NVIDIA's table is marked preliminary.
way
each
Notes
Blackwell generation. The step change is domain size, not link speed: 8 -> 72 GPUs in one non-blocking domain. Inference consequence: a 70B-class model at 4-bit fits comfortably in one 8-GPU domain, so TP degree beyond 8 buys little and instead adds all-reduce traffic that scales with TP degree - the NVL72 domain is aimed at MoE expert parallelism (many low-bandwidth, latency-tolerant all-to-all transfers) and trillion-parameter weights rather than at more TP. NVIDIA's table is marked preliminary.
Cross-references
This record references no other records — it is a leaf. It is cited by 13.