Announcement

Collapse
No announcement yet.

Linux's New Compute Accelerator Framework Quickly Taking Shape

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

  • Linux's New Compute Accelerator Framework Quickly Taking Shape

    Phoronix: Linux's New Compute Accelerator Framework Quickly Taking Shape

    Towards the end of October there finally came about a patch series fleshing out the "accel" subsystem for the Linux kernel in preparing this new subsystem/framework that builds atop the Direct Rendering Manager (DRM) code and is designed for all the up and coming AI accelerator drivers for the kernel. Given the number of accelerator drivers from different vendors eyeing mainline kernel adoption, this new compute accelerator framework is quickly being formed...

    https://www.phoronix.com/news/Linux-...Accelerator-v3

  • #2
    Not familiar with this framework. Would this be a way to unify things like cuda/opencl etc into a common "language" that will work across different "subplatforms/accelerator cards"?!

    http://www.dirtcellar.net

    Comment


    • #3
      Originally posted by waxhead View Post
      Not familiar with this framework. Would this be a way to unify things like cuda/opencl etc into a common "language" that will work across different "subplatforms/accelerator cards"?!
      No.


      5 character limit.

      Comment


      • #4
        Originally posted by waxhead View Post
        Not familiar with this framework. Would this be a way to unify things like cuda/opencl etc into a common "language" that will work across different "subplatforms/accelerator cards"?!
        when they say "framework" here it's not a unified language, it's just the kernel development framework... it refers to the kernel parts for all accelerators being put in the same kernel subsystem to make it easier to colaborate, share bits of code between drivers for different accelerators and define upstream maintaners that will work reviewing and accepting code focused on this instead of mixed with a bunch of unrelated stuff... probably some experts in the subject will be put up and front for the task and that's already a win

        each accellerator will still have their own drivers and can still offer their own APIs and compute languages over the same subsystem... though maybe at some point this does helps catalyze something in that direction

        Comment

        Working...
        X