NVSwitch
NVIDIA's switch chip designed to complement NVLink, used to build all-to-all GPU interconnect topologies. It enables GPUs within and across racks to communicate at full bandwidth and serves as a foundational interconnect component in large-scale AI clusters such as DGX SuperPOD.
Timeline (last 90 days)
The 8-GPU configuration of the HGX H200 interconnects 8 H200 GPUs with high bandwidth via NVLink and NVSwitch, with a total VRAM of up to 1.1 TB (HBM3e).
NVSwitch is the companion switch chip used to build fully connected topologies and is the foundational interconnect architecture for large-scale clusters such as DGX SuperPOD.
NVIDIA achieves high-bandwidth communication between GPUs and between servers through technologies such as InfiniBand, NVLink, and NVSwitch
All Facts (5)
8卡DGX B200系统的GPU间总带宽达到14.4TB/s
85%UnverifiedGPU interconnect bandwidth within data centers is typically on the order of 100–400 Gbps (via InfiniBand HDR/NDR or NVSwitch), while wide-area network bandwidth between data centers may be only 1–10 Gbps.
70%UnverifiedThe 8-GPU configuration of the HGX H200 interconnects 8 H200 GPUs with high bandwidth via NVLink and NVSwitch, with a total VRAM of up to 1.1 TB (HBM3e).
50%UnverifiedNVSwitch is the companion switch chip used to build fully connected topologies and is the foundational interconnect architecture for large-scale clusters such as DGX SuperPOD.
50%UnverifiedNVIDIA achieves high-bandwidth communication between GPUs and between servers through technologies such as InfiniBand, NVLink, and NVSwitch
50%