Demo of running CUDA programs on GPU and potential speed-up over CPU ITCS 6/8010 CUDA Programming,...

9
Demo of running CUDA programs on GPU and potential speed-up over CPU ITCS 6/8010 CUDA Programming, UNC-Charlotte, B. Wilkinson, Jan 10, 2011
  • date post

    20-Dec-2015
  • Category

    Documents

  • view

    215
  • download

    0

Transcript of Demo of running CUDA programs on GPU and potential speed-up over CPU ITCS 6/8010 CUDA Programming,...

Demo of running CUDA programs on GPU and

potential speed-up over CPU

ITCS 6/8010 CUDA Programming, UNC-Charlotte, B. Wilkinson, Jan 10, 2011

2

Xclock running on client PC

Xclock running on coit-

grid01.uncc.edu

Xclock running on coit-

grid06.uncc.edu

Xterm running on client PC, logged onto coit-grid06.uncc.edu

Typical user

interface (using a Windows

PC)

WinSCP running on client PC connected to

grid01.uncc.edu

To make sure all X servers running

3

Heat distribution problem(Solving Laplace’s equation)

800 x 800 points with 2000 iterationsSpeed-up = 21.2(Not sufficiently converged)

Fireplace

4

800 x 800 points50000 iterations

Different GPU block structure

Speed-up = 16.57

Fireplace

5

200 x 200 points with 20000 iterations

Different GPU block structure

Speed-up = 3.9

Fireplace

6

Potential speed-upSpeed-up factor = Execution time on CPU

Execution using GPU

One can get one or two orders of magnitude speed up just by using a single GPU!!

But it will take care to achieve large speed-ups.

Algorithm used on GPU may be different to that used on CPU because of constraints on GPU, so should really compare best sequential version on CPU with algorithm used on GPU

7

N Body problem

8

Questions