Announcement

Collapse
No announcement yet.

Intel Extension For TensorFlow Released - Provides Intel GPU Acceleration

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

  • Intel Extension For TensorFlow Released - Provides Intel GPU Acceleration

    Phoronix: Intel Extension For TensorFlow Released - Provides Intel GPU Acceleration

    Intel has published the Intel Extension for TensorFlow that makes use of TF's PluggableDevice mechanism to now provide an Intel GPU back-end for TensorFlow that works with the Data Center GPU Flex Series as well as Arc Graphics discrete GPUs...

    https://www.phoronix.com/news/Intel-...For-TensorFlow

  • #2
    Typo:

    Originally posted by phoronix View Post
    TensorFlow's PluggableDevice interfac to allow

    Comment


    • #3
      Excited to see benchmarks of Intel’s(XMX engines) tensor accelerators compared to Nvidia’s(Tensor cores). The Arc A770 has a bit over 4.5x the amount of tensor accelerators when compared to the RTX 3060.

      Intel also confirmed Intel Arc can be used in the data center unlike Nvidia’s consumer GPUs.

      Comment


      • #4
        Would be nice if support for AMD hip meant working with rusticl, but it seems that is not to be.

        Comment


        • #5
          Originally posted by WannaBeOCer View Post
          Excited to see benchmarks of Intel’s(XMX engines) tensor accelerators compared to Nvidia’s(Tensor cores). The Arc A770 has a bit over 4.5x the amount of tensor accelerators when compared to the RTX 3060.

          Intel also confirmed Intel Arc can be used in the data center unlike Nvidia’s consumer GPUs.
          Would be nice if we could get a concrete yes or no on SR-IOV on consumer cards, because I'm seeing both answers scattered about.

          Comment


          • #6
            Originally posted by Snaipersky View Post

            Would be nice if we could get a concrete yes or no on SR-IOV on consumer cards, because I'm seeing both answers scattered about.
            SR-IOV isn’t support on Intel Arc GPUs but is supported on their Flex GPUs. There isn’t licensing cost which Nvidia charges for GRID. I think AMD only support SR-IOV on MxGPUs, I suggest buying a Mi25 for $99 used if you really need it. Aside from that Intel Flex GPUs will be your cheapest new option which will probably cost $1000-$2000.
            Last edited by WannaBeOCer; 29 October 2022, 06:12 AM.

            Comment


            • #7
              Originally posted by WannaBeOCer View Post
              Intel also confirmed Intel Arc can be used in the data center unlike Nvidia’s consumer GPUs.
              They said their software license won't forbid it. That's different than a statement about suitability for that purpose.

              Comment


              • #8
                Originally posted by coder View Post
                They said their software license won't forbid it. That's different than a statement about suitability for that purpose.
                Agreed, hopefully we’ll see blower style models from some AIB that are suitable. The current axial designs aren’t. Slapping in that many tensor accelerators it won’t be surprising regarding the use in the data center. Along with the quick support for OneAPI makes the 16GB model more appealing.

                Comment


                • #9
                  Originally posted by Developer12 View Post
                  Would be nice if support for AMD hip meant working with rusticl, but it seems that is not to be.
                  HIP and OpenCL are two different APIs.

                  Comment


                  • #10
                    Originally posted by agd5f View Post

                    HIP and OpenCL are two different APIs.
                    That's the problem. While ROCm offers an OpenCL interface, that doesn't mean that stuff built on "ROCm" are using OpenCL.

                    Comment

                    Working...
                    X