Announcement

Collapse
No announcement yet.

Mediatek Posts 8k Lines Of New Linux Kernel Driver Code For AI Processing Unit Support

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

  • Mediatek Posts 8k Lines Of New Linux Kernel Driver Code For AI Processing Unit Support

    Phoronix: Mediatek Posts 8k Lines Of New Linux Kernel Driver Code For AI Processing Unit Support

    For a number of months Mediatek engineers have been posting some Linux kernel driver code for bringing up the AI Processing Unit (APU) within the MT8192 SoC while out this weekend is the complete patch series at more than eight thousand lines of code...

    Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite

  • #2
    Interesting. Is this something that Tensorflow (for instance) could eventually support, to the point where once could just fire up an existing TF project with minor changes to the code? Or is it something that will have to be very specifically targeted?

    Comment


    • #3
      Originally posted by brucethemoose View Post
      Interesting. Is this something that Tensorflow (for instance) could eventually support, to the point where once could just fire up an existing TF project with minor changes to the code? Or is it something that will have to be very specifically targeted?
      It should be already supported in Android via NNAPI
      Last edited by RedEyed; 23 October 2021, 11:08 PM.

      Comment


      • #4
        This would start to make the prospect of running Linux on them very appealing. We need these SoCs on a SBC or in a NUC-type mini desktop chassis!

        Comment


        • #5
          > 2.4 TOPs
          what kind of TOPS? int8/16/32/64 or float8/16/32/64 ?

          Comment


          • #6
            Yet people keep saying MediaTek never shares driver code………

            Comment


            • #7
              Originally posted by RedEyed View Post
              > 2.4 TOPs
              what kind of TOPS? int8/16/32/64 or float8/16/32/64 ?
              Take any such statistic with a grain of salt. MediaTek is one of the companies known to cheat on benchmarks. In fact, MediaTek is so awful they provide it as a service to their partners. https://www.anandtech.com/show/15703...ating-mediatek

              Comment


              • #8
                Originally posted by stormcrow View Post

                Take any such statistic with a grain of salt. MediaTek is one of the companies known to cheat on benchmarks. In fact, MediaTek is so awful they provide it as a service to their partners. https://www.anandtech.com/show/15703...ating-mediatek
                > Take any such statistic with a grain of salt
                That's what I'm doing...

                Comment


                • #9
                  APUs are a mystery to me. Are there benefits to general purpose computing? For instance, in gaming or image processing? Will we see APU-accelerated builds of ImageMagick or Pillow, or will the functionality be exposed in ways where a library chooses to use an APU instead of the CPU?

                  Comment


                  • #10
                    Originally posted by mangeek View Post
                    APUs are a mystery to me. Are there benefits to general purpose computing? For instance, in gaming or image processing? Will we see APU-accelerated builds of ImageMagick or Pillow, or will the functionality be exposed in ways where a library chooses to use an APU instead of the CPU?
                    APU is just a GPU and graphics together like you see in most laptops. Are there benefits to this approach? Yes because it becomes easier for the CPU & GPU to share memory and move data between them. The performance of Apples M1 is partially attributable to how integrated every thing is. APUs are the future as AMD has stated that their GPUs at least for desktops will all have integrated graphics starting I think late next year. Intel has already started moving in that direction.

                    Comment

                    Working...
                    X