Announcement

Collapse
No announcement yet.

Limited SR-IOV support for Vega (and maybe others)? (proposal)

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

  • Limited SR-IOV support for Vega (and maybe others)? (proposal)

    I'm not sure if this would be possible, but I think AMD could gain a lot of positive feedback if they would enable this.

    A lot of people would like (or at least consider) to ditch Windows as their main system, but are also gamers, or use other software that only runs good under Windows with full 3D hardware support.

    Some good progress has been made recently with enabling virtual displays (no need for additional displays, or constantly switching inputs anymore) for PCI pass-through GPUs in KVM (see Looking Glass at level1tech for example - I think my ability to link is still limited). I also think I have seen some AMD driver patches in the past, with something about virtual displays, which could maybe help with that? Maybe it would be even possible to enable Freesync for virtual displays using PCI pass-through?

    But the main part which is currently missing, is the ability to use a single GPU in the host system, and in a virtual machine at the same time. Currently you always have to use multiple GPUs to use PCI pass-through, which is some real hassle which could potentially be avoided. I think it should be possible to give people the ability with enabling SR-IOV on RX Vega, and other compatible hardware. As far as I can tell currently it is only enabled via BIOS and driver on the Radeon Instinct series.

    So my proposal would be to enable a limited SR-IOV support for other Vega series, like allow one hardware accelerated GPU instance on RX Vega, and maybe two, or three, on the Pro WX 9100, and keep the hardware max. limit on the Radeon Instinct MI25. This would give you enough room to keep the price differences, but could potentially win over a lot new customers, and produce some really good press.

    To make it short: AMD please enable us to use a single GPU with full hardware acceleration in the main system, and in a virtual machine, at the same time.

    Thanks,
    BoMbY

  • #2
    I hope that a decision maker from AMD listens to this and considers this.
    May be have a mail sent to this email id ?
    For more information, contact the AMD Professional Graphics Team by email: [email protected]
    Source : https://www.amd.com/en-us/press-rele...2016feb01.aspx

    Originally posted by agd5f View Post
    SR-IOV is only enabled on SR-IOV capable parts (generally workstation/server SKUs).
    Source: https://www.phoronix.com/forums/foru...n-linux-status

    Comment


    • #3
      Definitely AMD shot themselves into the foot by disabling SR-IOV almost everywhere, because it has uses outside workstation/server market. Cross-platform development of GPGPU applications for example. Even Vega FE, a card aimed squarely at developers, doesn't support it.

      But: If I understand correctly, SR-IOV is not possible to use together with display outputs. That means you would need a second GPU (possibly iGPU) to drive your outputs.

      Also I think there is no way to fine-grained control of SR-IOV capabilities in polyfuses. Either it is all on for 8(16?) users, or it is all off.

      Comment


      • #4
        So, activating SR-IOV disables all physical output capabilities of the GPU?

        Comment


        • #5
          Originally posted by BoMbY View Post
          So, activating SR-IOV disables all physical output capabilities of the GPU?
          The display hardware does not support SR-IOV so it's not available in the virtual functions as such I think all SR-IOV capable cards are headless.

          Comment


          • #6
            Radeon Pro V340 appears to have a single mini-DP connector.

            Comment


            • #7
              Originally posted by chithanh View Post
              Radeon Pro V340 appears to have a single mini-DP connector.
              It would be nice to know what that connector is.
              I happened to check the datasheet & it seems to indicate one "circuitry".

              Remote Management
              The AMD Radeon™ Pro V340 datacenter graphics
              card adds out-of-band management of circuitry for
              simplified GPU querying in large scale systems.
              Accessibility is via I2C and will return data regardless of
              the state of the GPU providing advanced monitoring
              for a range of static and dynamic GPU information.
              Using PMCI compliant data structures, the V340 gives
              system administrators the information they need to
              manage their environments.

              Comment

              Working...
              X