[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums

backgammon bots and GPUs?

Posted By: eXtreme Gammon
Date: Monday, 12 July 2010, at 2:25 p.m.

In Response To: backgammon bots and GPUs? (Chuck Bower)

It looks appealing so I made some experiment using CUDA and also directCompute (DirectX) a few month ago.

Unfortunately with the size of matrix used by the Neural Network (256x128) it is not efficient Bench for 10000 Matrix multiplication

  • CPU=0.480 sec
  • GPU=0.519 sec

    But the CPU code is not SSE and use only 1 thread so it could be up to 15 times faster if fully optimized.

    I can find huge improvement with CUDA if using very large matrix (2048x2048) 10000 trials

  • CPU=62.76 sec
  • GPU=5.38 sec

    and for (8192x8192) 2500 trials

  • CPU=250.51 sec
  • GPU= 10.50 sec

    To summarize: For backgammon purposes the overhead of sending the data to the GPU, initializing and retrieving is, by far, too costly to be viable.

    Last note: for XG the time spent in the NN is only 34% of the total CPU.

    Tested on

  • CPU: Core i7 920 @1.86Hhz
  • GPU: GeForce GTS 250 (128 shaders)

    • Messages In This Thread

     

    Post Response
    Your Name:
    Your E-Mail Address:
    Subject:
    Message:

    If necessary, enter your password below:

    Password:
     

     

    [ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

    BGonline.org Forums is maintained by Stick with WebBBS 5.12.