Questions about PPCIe Mode Behaviors and CUDA IPC Support

Accelerated Computing Confidential Computing
1

According to https://docs.nvidia.com/570TRD1-trusted-computing-solutions-release-notes.pdf, , there are specific limitations in the Hopper PPCIe Mode.

  • In the PPCIe mode, when the source or destination operand are imported, GPU memory allocations on a device that is not visible to the process, the host-to-device, or device-to-host copies might fail asynchronously with cudaErrorLaunchFailure.
  • In the PPCIe mode, using cooperative_groups::multi_grid_group::sync in kernels launched with cudaLaunchCooperativeKernelMultiDevice results in the kernel failing with cudaErrorIllegalAddress.
  • CUDA Interprocess Communication (IPC) is not supported in PPCIe mode.

I would like to seek clarification on the reasoning behind these limitations:

  • Why the first and second points mentioned are forbitten in PPCIe mode. Shouldn’t the GPU and Switch belonging to the same VM trust each other?

  • Why isn’t CUDA IPC supported in PPCIe mode? Even sharing video memory on the same GPU isn’t allowed. And isn’t it supported in SPT CC either?