NVIDIA NVSwitch#
NVIDIA® NVSwitch™ is a high-bandwidth, low-latency fabric that connects multiple GPUs in one system. For vGPU™, peer-to-peer traffic between GPUs on the fabric uses NVLink™ instead of relying only on PCIe. The switch behaves as a crossbar. GPUs can reach each other at NVLink rates. This matters for multi-GPU workloads that exchange data in memory on a single node.
Support is limited to specific hardware platforms, vGPU types, hypervisor releases, and guest operating systems.
For information about using the NVSwitch, refer to the NVIDIA Fabric Manager documentation.
Note
On NVIDIA HGX™ B200 and B300 with VMware vSphere, NVSwitch support requires the Fabric Manager CRX (Container Runtime for ESXi) component. The Fabric Manager CRX component is not included in NVIDIA AI Enterprise Infra 9.0. For NVIDIA AI Enterprise deployments on B200 and B300 HGX with VMware vSphere, use NVIDIA AI Enterprise Infra 8.2.
Platform Support for NVIDIA NVSwitch#
NVIDIA HGX B300 8-GPU baseboard
NVIDIA HGX B200 8-GPU baseboard
NVIDIA HGX H200 8-GPU baseboard
NVIDIA HGX H100 8-GPU baseboard
NVIDIA HGX H800 8-GPU baseboard
NVIDIA HGX A100 8-GPU baseboard
NVIDIA NVSwitch Limitations#
Only time-sliced vGPUs are supported. MIG-backed vGPUs are not supported. Unified memory (UVM) is supported only with 1:1 (full-sized) time-sliced vGPUs.
All vGPUs communicating peer-to-peer must be assigned to the same VM.
On GPUs based on the NVIDIA Hopper™ and Blackwell GPU architectures, multicast is supported when unified memory (UVM) is enabled.
On NVIDIA systems that include NVSwitch, GPU passthrough with VMware vSphere is supported only as a full-node configuration. Sub-node configurations, such as 1-GPU or 2-GPU VMs, are not supported on NVIDIA HGX systems with VMware vSphere.
On NVIDIA HGX B200 and B300 with VMware vSphere, NVSwitch support requires the Fabric Manager CRX component. This component is not included in NVIDIA AI Enterprise Infra 9.0.
On NVIDIA HGX B200 and HGX B300 with VMware vSphere, NVIDIA vGPU for Compute supports 1:1 vGPU VMs and multi-vGPU VMs only. Fractional vGPU VMs are not supported.
Hypervisor Platform Support for NVSwitch#
Consult the documentation from your hypervisor vendor for information about which generic Linux with KVM hypervisor software releases supports NVIDIA NVSwitch.
All supported Red Hat Enterprise Linux KVM and Ubuntu KVM releases support NVIDIA NVSwitch.
The earliest VMware vSphere Hypervisor (ESXi) release that supports NVIDIA NVSwitch depends on the GPU architecture.
VMware vSphere ESXi release by GPU architecture
GPU Architecture |
Earliest Supported VMware vSphere Hypervisor (ESXi) Release |
|---|---|
NVIDIA Blackwell |
VMware vSphere Hypervisor (ESXi) 9 Update 1 [1] |
NVIDIA Hopper |
VMware vSphere Hypervisor (ESXi) 8 update 2 |
NVIDIA Ampere |
VMware vSphere Hypervisor (ESXi) 8 update 1 |
vGPU Support for NVSwitch#
Only the following vGPU for Compute time-sliced vGPUs allocated all of the physical GPU’s framebuffer are supported:
NVIDIA A800
NVIDIA A100 HGX
NVIDIA B300 HGX
NVIDIA B200 HGX
NVIDIA H800
NVIDIA H200 HGX
NVIDIA H100 SXM5
NVIDIA H20
Board |
vGPU |
|---|---|
NVIDIA HGX B300 288 GB |
B300X-269C |
NVIDIA HGX B200 180 GB |
B200X-180C |
Board |
vGPU |
|---|---|
NVIDIA H800 SXM5 80 GB |
H800XM-80C |
NVIDIA H200 SXM5 141 GB |
H200X-141C |
NVIDIA H100 SXM5 80 GB |
H100XM-80C |
NVIDIA H20 SXM5 141 GB |
H20X-141C |
NVIDIA H20 SXM5 96 GB |
H20-96C |
Board |
vGPU |
|---|---|
NVIDIA A800 HGX 80 GB |
A800DX-80C |
NVIDIA A100 HGX 80 GB |
A100DX-80C |
NVIDIA A100 HGX 40 GB |
A100X-40C |
Guest OS Releases Support for NVSwitch#
Linux only. NVIDIA NVSwitch is not supported on Windows.
NVLink Multicast (on NVSwitch)#
NVLink™ multicast is an NVSwitch fabric capability that distributes data from one GPU to many peers over the NVSwitch fabric.
Important
NVLink multicast support requires that unified memory is enabled. For more information about enabling unified memory, refer to the Enabling Unified Memory for a vGPU documentation, or Unified Virtual Memory (UVM) for in-repo coverage.
Only full-sized, time-sliced NVIDIA® vGPU for Compute support NVLink multicast.
Board |
vGPU |
|---|---|
NVIDIA HGX™ B300 288 GB |
NVIDIA B300X-269C |
NVIDIA HGX B200 180 GB |
NVIDIA B200X-180C |
Board |
vGPU |
|---|---|
NVIDIA H800 SXM5 80 GB |
NVIDIA H800XM-80C |
NVIDIA H200 SXM5 141 GB |
NVIDIA H200X-141C |
NVIDIA H100 SXM5 80 GB |
NVIDIA H100XM-80C |
NVIDIA H20 SXM5 141 GB |
NVIDIA H20X-141C |
NVIDIA H20 SXM5 96 GB |
NVIDIA H20-96C |
NVLink Multicast Limitations#
NVLink multicast is supported only with full-sized, time-sliced NVIDIA vGPU for Compute profiles. MIG-backed vGPU profiles and fractional time-sliced profiles are not supported.
Unified memory must be enabled for the vGPU. Refer to the Enabling Unified Memory for a vGPU documentation.
NVLink multicast requires NVSwitch-connected GPU boards, such as NVIDIA HGX baseboards. PCIe-only GPU configurations do not support NVLink multicast.
Only the specific board and vGPU profile combinations listed in the support tables above are supported.