GPU Load

This topic contains 4 replies, has 2 voices, and was last updated by  hiendang 1 year, 5 months ago.

Viewing 5 posts - 1 through 5 (of 5 total)
  • Author
    Posts
  • #53614

    hiendang
    Member

    Hi,

    I see that there is ‘GPU task load: 3D/TA core’ in some figures of PVRTune GUI and ‘Shader Load’ in Monitor window (http://cdn.imgtec.com/sdk-documentation/PVRTune.User%20Manual.Developer.pdf Figure 11: Monitor Window). However, I can not find those metrics in my PVRTune which is included in the latest SDK.
    Could you please show me how to enable those counters or tell me why they disappear?

    My purpose is simply stress the GPU and confirm its workload status whether it reaches 100%.
    Currently, I am executing 4 glmark2 apps in parallel and use PVRTune to observe GPU stats.
    The problem is I do not know which counter should be based on to judge whether the GPU is overloaded.
    I would appreciate if you can guide me to the correct path.

    I am sorry if this forum is not the place for my question!

    Thank you!

    #53615

    Joe Davis
    Member

    The Monitor widget names are short hands for the full counter names. In the “PVRTune Counter List and Description” document, the description for the total shader load counter can be found under “Shader processing load”.

    My purpose is simply stress the GPU and confirm its workload status whether it reaches 100%.

    In most counters displayed in PVRTune, 100% is a theoretical maximum performance of a given hardware unit. As the performance of many unit depends on how quickly the preceding block can feed work to them, it’s unlikely you will be able to create a benchmark that causes a given unit to be 100% loaded.

    In PVRTune, there is no counter for overall GPU utilization. We have considered adding such a counter before, but the data it gives could be misleading. For example, a load of 100% could indicate that any block in the GPU was busy during the time period, or it could indicate that every block is 100% utilized.

    The timing data and associated counters (e.g. Tiler active) will give you the clearest indication of how busy the GPU is. However, these tasks execute in parallel and the resources required to execute one can impact the performance of the other.

    #53626

    hiendang
    Member

    Hi Joe,

    Thank you very much for your explanation!

    I still have some unclear points.
    There are several versions of PVRTune out there which are different from the one (v14.1xx) integrated in the latest SDK.
    For example:
    – v14.59 in the first image of https://www.imgtec.com/blog/powervr-developers/powervr-graphics-sdk-tools-explained-pvrtune-gui-basics
    – v14.73 in 10th page of http://cdn.imgtec.com/sdk-presentations/gdc2014_pvrsdkProfilingAndDebugging.pdf

    So my questions are:
    1. Are they kind of prototype or RC of the next release of SDK?
    2. They have ‘GPU task load’ counter which is not available in the current version. The number/name of counter depends on SDK version or GPU architecture?

    Best regards,
    Hien

    #53628

    Joe Davis
    Member

    They have ‘GPU task load’ counter which is not available in the current version. The number/name of counter depends on SDK version or GPU architecture?

    Yes – counter names can change between SDK releases. Since the blog post and presentations you have linked to, the timing data names were changed along with the corresponding counters:

    Timing data
    “TA”–>”Tiler”
    “3D”–>”Renderer”.

    Counters
    “GPU task load: TA core”–>”Tiler active”
    “GPU task load: 3D core”–>”Renderer active”

    Regards,
    Joe

    #53648

    hiendang
    Member

    Hi Joe,

    Thank you very much!

    Best regards,
    Hien

Viewing 5 posts - 1 through 5 (of 5 total)
You must be logged in to reply to this topic.