Peer-To-Peer (P2P) CUDA Transfers#

Peer-to-Peer (P2P) CUDA transfers enable device memory between vGPUs on different GPUs that are assigned to the same VM to be accessed from within CUDA kernels. NVLink is a high-bandwidth interconnect that enables fast communication between such vGPUs.

P2P CUDA transfers over NVLink are supported only on a subset of vGPUs, hypervisor releases, and guest OS releases.

For multi-vGPU VM assignment without the P2P focus, see Multi-vGPU. For the shared hypervisor support matrix that covers both Multi-vGPU and P2P, see Multi-vGPU and P2P.

Peer-to-Peer CUDA Transfers Known Issues and Limitations#

  • Only time-sliced vGPUs are supported. MIG-backed vGPUs are not supported.

  • P2P transfers over PCIe are not supported.

Warning

On NVIDIA A100 (Ampere) GPUs, P2P over NVLink and NVSwitch is disabled when Unified Virtual Memory (UVM) is enabled on a multi-vGPU VM. Inside the guest, nvidia-smi topo -m reports PIX (PCIe) instead of NV12 (NVLink). To restore NVLink P2P between vGPUs, disable UVM on the vGPU devices (for example, pciPassthru<n>.cfg.enable_uvm = 0 on VMware vSphere). This restriction is specific to the NVIDIA Ampere A100 architecture and does not apply to NVIDIA Hopper (H100/H200) or newer architectures.

vGPU Support for P2P#

Only NVIDIA vGPU for Compute time-sliced vGPUs allocated all of the physical GPU framebuffer on physical GPUs supporting NVLink are supported.

Table 29 vGPU Support for P2P on the NVIDIA Blackwell GPU Architecture#

Board

vGPU

NVIDIA HGX B300 279 GB

NVIDIA B300X-279C

NVIDIA HGX B200 180 GB

NVIDIA B200X-180C

Table 30 vGPU Support for P2P on the NVIDIA Hopper GPU Architecture#

Board

vGPU

NVIDIA H800 PCIe 94 GB (H800 NVL)

H800L-94C

NVIDIA H800 PCIe 80 GB

H800-80C

NVIDIA H200 PCIe 141 GB (H200 NVL)

H200-141C

NVIDIA H200 SXM5 141 GB

H200X-141C

NVIDIA H100 PCIe 94 GB (H100 NVL)

H100L-94C

NVIDIA H100 SXM5 94 GB

H100XL-94C

NVIDIA H100 PCIe 80 GB

H100-80C

NVIDIA H100 SXM5 80 GB

H100XM-80C

NVIDIA H100 SXM5 64 GB

H100XS-64C

NVIDIA H20 SXM5 141 GB

H20X-141C

NVIDIA H20 SXM5 96 GB

H20-96C

Table 31 vGPU Support for P2P on the NVIDIA Ampere GPU Architecture#

Board

vGPU

  • NVIDIA A800 PCIe 80 GB

  • NVIDIA A800 PCIe 80 GB liquid-cooled

  • NVIDIA AX800

A800D-80C

NVIDIA A800 PCIe 40 GB active-cooled

A800-40C

NVIDIA A800 HGX 80 GB

A800DX-80C [1]

  • NVIDIA A100 PCIe 80 GB

  • NVIDIA A100 PCIe 80 GB liquid-cooled

  • NVIDIA A100X

A100D-80C

NVIDIA A100 HGX 80 GB

A100DX-80C [1]

NVIDIA A100 PCIe 40 GB

A100-40C

NVIDIA A100 HGX 40 GB

A100X-40C [1]

NVIDIA A40

A40-48C

  • NVIDIA A30

  • NVIDIA A30X

  • NVIDIA A30 liquid-cooled

A30-24C

NVIDIA A16

A16-16C

NVIDIA A10

A10-24C

NVIDIA RTX A6000

A6000-48C

NVIDIA RTX A5500

A5500-24C

NVIDIA RTX A5000

A5000-24C