GPU Plotter is not using my GPU memory.



  • For some reason GPU Plotter its not wanting to use all my gpu memory. I have this configuration

    0 0 2048 128 2048
    0 1 2048 128 2048

    My PC Specs...
    Intel i5-6500
    2x nVidia Quadro M6000 24gb (SLI)
    32gb DDR4 2666MHz (4x 8gb)

    For the Global Work Size
    globalWorkSize: The amount of nonces to process at the same time (determine the required amount of GPU memory). Should be a power of 2.

    Both of my Quadros have 24gb of memory but if i set the Global Work Size over 2048 it throw me this error... [ERROR][-36][CL_INVALID_COMMAND_QUEUE] Error in step2 kernel finish

    Can someone tell me if there's a way to use all my GPU memory for Plotting?


  • admin

    @Wolf I'd change the 600 setting to a larger power of two - eg 2048, 4096, 8192 etc - but I've had issues getting nVidias to plot - have stuck with AMD.



  • @haitch
    0 0 2048 128 2048
    0 1 2048 128 2048

    this configuration give me these results

    0_1480207905303_upload-3ee25d41-7638-4e3d-8e8b-770b6dd40c6e

    The Plotter uses 26gb of CPU memory but only 2gb each of both Quadros
    There a way that we can configure this to use more GPU memory?


  • admin

    @Wolf increase the last number - try 65536 - 16GB RAM



  • @haitch the 65536 give me this [ERROR][-36][CL_INVALID_COMMAND_QUEUE] Error in step2 kernel finish


  • admin

    @Wolf try powers of 2 between the 2048 and 65536 - drop down to 32768, 16384, 8192 ....



  • @haitch You are talking about the [globalWorkSize] right? Can I get an example?


  • admin

    @Wolf I was mis-remembering the settings - set the last number to 8192, vary the third number - global worksize



  • @haitch If i set the last number to 8192 it throws me an error
    Even if i set the configuration to this

    example..
    0 0 1024 128 8192
    0 1 1024 128 8192

    This is the error for this configuration.
    [ERROR][-5][CL_OUT_OF_RESOURCES] Error in synchronous read

    The highest that I can configure the last number with no problems is 7168. But I have to leave the third number to 1024

    example..
    0 0 1024 128 7168
    0 1 1024 128 7168

    But if I leave the last number to 7168 and go up on the third number to 2048 it throws me a another error.

    example..
    0 0 2048 128 7168
    0 1 2048 128 7168

    This is the error for this configuration.
    [ERROR][-36][CL_INVALID_COMMAND_QUEUE] Error in step2 kernel finish



  • btw.. The highest I can configure the third number with no problems is 4096. But I have to leave the last number to 1024

    example..
    0 0 4096 128 1024
    0 1 4096 128 1024

    But with this configuration I lose 20000 nonces/minutes,.



  • @Wolf said in GPU Plotter is not using my GPU memory.:

    btw.. The highest I can configure the third number with no problems is 4096. But I have to leave the last number to 1024

    example..
    0 0 4096 128 1024
    0 1 4096 128 1024

    But with this configuration I lose 20000 nonces/minutes,.

    You lose 20K nonces with this method? Meaning the last number is more important? What speed are you getting on this config vs the other?

    I have a 1060 3gb and average about 20K a min, I must not be near maximized.



  • This post is deleted!


  • So it is actually slower for me to plot 1 hard drive at a time? Do I just start up another instance of GPUPlotter or do I put multiple lines in the bat file?



  • This post is deleted!


  • I'm running a GTX 1070 for plotting and in my experience, the hashesNumber crashes my driver if the value is above 500 or so. I just set it to 128 and tweak the other numbers instead.

    Here are some other things I noticed with the GTX 1070 and my 16GB RAM:

    A globalWorkSize of 4096 translates to 1GB of VRAM AND 1GB of RAM usage.
    The localWorkSize doesn't want to go above 512. Higher values throw errors.
    Asking for more than 11GB/16GB total RAM for the plotter throws memory allocation errors.

    Here are some working examples:

    10 GB RAM
    10,485,760 KB / 256 = 40,960 Stagger

    makeplots.bat
    gpuPlotGenerator generate buffer D://Burst/plots/9999999_0_81920_40960
    pause

    devices.txt
    1 0 4096 512 128

    Throughput is about 22,000 nonces/minute.

    If you want to increase the VRAM usage above 1 GB, then the value must be a multiple of your Stagger size or OpenCL will throw memory allocation errors. The gpuPlotGenerator setup process recommended 16384 for me when I ran it, however, I didn't realize that it was also going to mirror that 4 GB of VRAM in the system RAM as well, so I had to reduce the Stagger size to avoid the aforementioned memory allocation errors:

    4 GB RAM
    4,194,304 KB / 256 = 16,384 Stagger

    makeplots.bat
    gpuPlotGenerator generate buffer D://Burst/plots/9999999_0_65536_16384
    pause

    devices.txt
    1 0 16384 512 128

    Throughput is about 25,000 nonces/minute.

    Here are some other tests I ran:

    Keep the 4 GB Stagger and reduce to 2 GB VRAM. Throughput is about 23,000 nonces/minute.
    Use an 8 GB Stagger with 2 GB VRAM. Throughput is about 21,000 nonces/minute.
    Use a 7 GB Stagger with 1.768 GB VRAM. Throughput is about 22,000 nonces/minute.

    So what about writing to two drives in parallel? This was a bit tricky due to my 11GB memory limitation, but still feasible:

    5 GB RAM
    5,242,880 KB / 256 = 20,480 Stagger

    makeplots.bat
    gpuPlotGenerator generate buffer D://Burst/plots/9999999_0_61440_20480 F://Burst/plots/9999999_61440_61440_20480
    pause

    devices.txt
    1 0 4096 512 128

    Throughput is about 30,000 nonces/minute.