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

BGonline.org Forums

time considerations -

Posted By: Frank Berger
Date: Wednesday, 29 August 2012, at 7:22 p.m.

In Response To: time considerations - (alex99)

Currently the programming modell is *vastly* different to a CPU. This is not only awful lot of work, but the problem simply doesn't match.

1) E.g. 32 cores share one program counter. If you have a misaligned memory access or a swivth with different outcomes etc. you code will get serialzed and then a CPU is much better.

2) GPU's shine if you move just few data (memory transfer is very expensive) and calculate like mad on the GPU. Matrix multiplication or FFT is an perfect example. The problem is: with a neural net you don't need a matrix multiplication but Matrix-Vector multiplication. Here happens a lot less calculation but it still pays off if you have very large vectors.... unfortunately I (and Xavier and Gnu will be similar) have around 250 to 300 inputs so in fact using a middle class graphic card is slower. At least that was my experience with OpenCL and IIRC it was Xaviers experience with CUDA. Has been posted here sometimes in the past.

ciao
Frank

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.