Announcement

Collapse
No announcement yet.

An Early Benchmark Of The NVIDIA CUDA GPU Performance On WSL2

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

  • oiaohm
    replied
    Originally posted by bug77 View Post
    Did you post the wrong link? Because I visit, it tells me about "productivity and collaboration applications like Skype and YouTube and even PowerPoint". Nothing about compute
    https://developer.nvidia.com/cuda-toolkit-31-downloads

    Compute does not stop exactly where you expect. You do run into excel and other desktop applications running compute. It was not exactly the wrong link. Grid from Nvidia is also for compute under hypervisor.

    https://aws.amazon.com/ec2/instance-types/
    If you look under Accelerated Computing here you will find under P3.

    • High frequency Intel Xeon E5-2686 v4 (Broadwell) processors for p3.2xlarge, p3.8xlarge, and p3.16xlarge.
    • High frequency 2.5 GHz (base) Intel Xeon P-8175M processors for p3dn.24xlarge.
    Xeon E5-2686 is a 18 core part with out hyper threading and Xeon P-8175M is a 24 core chip. p3.2xlarge only gives you 8 vcpu or 4 real cores of usage yes this is being sliced out by hypervisor. So yes there is a lot of compute on AWS that performs quite well that is done under a hypervisor so being under hypervisor is not excuse for bad performance at least not on the GPU side if it done well. WSL2 is not really done well in this regard.

    Leave a comment:


  • bug77
    replied
    Originally posted by oiaohm View Post

    There are a lot of compute heavy workloads not run on bare metal.
    https://www.nvidia.com/en-us/data-ce...rtual-pc-apps/

    Grid for example is used by AMS with hypervisor instances there are also fpga passed though as well. This is really about server density.
    Did you post the wrong link? Because I visit, it tells me about "productivity and collaboration applications like Skype and YouTube and even PowerPoint". Nothing about compute.
    Originally posted by AndyChow View Post
    I've run lot of CUDA things in docker containers. Not really pure-virtualization, but not bare-metal either. There is a small (2-8%) penalty I've noticed, probably because the io on the docker filesystem is really slow.

    To be able to run on a vm on windows without a passthrough, it's impressive that they can even use CUDA through such a layer. It's good for testing maybe. But I wouldn't do any serious dev on windows in the first place, but many people do and they like it just fine.
    Docker runs just an isolated process, it doesn't virtualize hardware

    Leave a comment:


  • AndyChow
    replied
    I've run lot of CUDA things in docker containers. Not really pure-virtualization, but not bare-metal either. There is a small (2-8%) penalty I've noticed, probably because the io on the docker filesystem is really slow.

    To be able to run on a vm on windows without a passthrough, it's impressive that they can even use CUDA through such a layer. It's good for testing maybe. But I wouldn't do any serious dev on windows in the first place, but many people do and they like it just fine.

    Leave a comment:


  • oiaohm
    replied
    Originally posted by bug77 View Post
    Can't say I'm surprised. Compute is one workload you really, really want to run on bare metal. There's a reason specialized hardware was created for it.
    There are a lot of compute heavy workloads not run on bare metal.
    https://www.nvidia.com/en-us/data-ce...rtual-pc-apps/

    Grid for example is used by AMS with hypervisor instances there are also fpga passed though as well. This is really about server density.

    Leave a comment:


  • blacknova
    replied
    Is it possible to get native Windows 10 numbers for these tests?

    Leave a comment:


  • shmerl
    replied
    What is even the point of running CUDA in a virtual machine? I wouldn't even ask the question of what's the point of using CUDA in the first place

    Leave a comment:


  • Mario Junior
    replied
    What a shit...

    Leave a comment:


  • Michael
    replied
    Originally posted by tildearrow View Post
    Did you just...



    Wait did I just steal their name for my recording software?! D:
    It's a new test, yes. Added to PTS/OpenBenchmarking a few days ago when scrolling through CUDA projects on GitHub. And being a fan of a few breweries in Darmstadt from a SUSE'r that lives there, I couldn't resist as I love the town.

    But it's short name of the upstream benchmark is DAPHNE.

    Leave a comment:


  • tildearrow
    replied
    Did you just...

    Darmstadt Automotive Parallel Heterogeneous Suite
    Wait did I just steal their name for my recording software?! D:

    Leave a comment:


  • pipe13
    replied
    Yabut it's a start, and at least some bits work. Whether or not the current performance penalty is inherent is up for debate. The qemu boys seem to do pretty well, at least with video passthrough: https://wiki.gentoo.org/wiki/GPU_pas...bvirt_qemu_kvm

    Leave a comment:

Working...
X