Announcement

Collapse
No announcement yet.

RADV External Memory Patches Revised, Needed For SteamVR Support

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • RADV External Memory Patches Revised, Needed For SteamVR Support

    Phoronix: RADV External Memory Patches Revised, Needed For SteamVR Support

    Bas Nieuwenhuizen has taken to revising David Airlie's work on external memory support for the RADV Vulkan driver, which are among the extensions needed by SteamVR on Linux...

    http://www.phoronix.com/scan.php?pag...nal-Memory-Bas

  • #2
    I wondering if it's possible to use this extension to implement RDMA operations between gpus on different hosts like GPUDirect and DirectGMA but with consumers GPUs.
    Last edited by sp82; 05 June 2017, 08:07 AM.

    Comment


    • #3
      Originally posted by sp82 View Post
      I wondering if it's possible to use this extension to implement RDMA operations between gpus on different hosts like GPUDirect and DirectGMA but with consumers GPUs.
      The biggest obstacle is that the Linux DMA API does not support access to device memory.

      Comment


      • #4
        Originally posted by sp82 View Post
        I wondering if it's possible to use this extension to implement RDMA operations between gpus on different hosts like GPUDirect and DirectGMA but with consumers GPUs.
        I'm a bit worried about security implications of this. I know that there is IOMMU, but still...

        Comment


        • #5
          Originally posted by agd5f View Post

          The biggest obstacle is that the Linux DMA API does not support access to device memory.
          Suppose I want to create a big cluster of gpus, not for mining!, to do distributed realtime raytracing global illumination and want to return the result back to a VR workstation (thanks Valve to make linux VR ready). Do you think is possible to RDMA using infiniband nics?

          Comment


          • #6
            Originally posted by sp82 View Post

            Suppose I want to create a big cluster of gpus, not for mining!, to do distributed realtime raytracing global illumination and want to return the result back to a VR workstation (thanks Valve to make linux VR ready). Do you think is possible to RDMA using infiniband nics?
            The lack of support for device memory in the DMA API is only a problem for cross device local memory access (e.g., device A DMAing directly to device B's local memory or vice versa). If lots of devices are working on their own local memory, it's not a problem.

            Comment

            Working...
            X