Benchmark performance of GPU computing of Monte-Carlo simulation

I wanted to quantify what performance improvement I may see using the GPU to solve systems, Ax=b, on a mobile workstation. The Parallel Computing Toolbox supplies the benchmarking script, paralleldemo_gpu_backslash.m, which I ran to see what improvement GPU computing would have over the CPU. In the documentation of this benchmark script the following graphical results are given:

On my computer the graphs look like this:

In the documentation the GPU outperforms the CPU for both double precision and single precision.

Why on my computer did the GPU computation perform worse than CPU computation for double-precision but performed better than CPU for single precision?

[DaveF]

The computer I have been evaluating this on is a mobile workstation, HP ZBook running 64 bit windows 7. The gpu is NVIDIA Quadro K3100M. The specs for this gpu seem to indicate that it is not designed for double precision.

Answer 1

[Walter Roberson]

Which GPU do you have?

There are overheads to communicating with the GPU, and it could be the case that only your single precision performance is enough to overcome those overheads. Entry-level NVIDIA GPU are better on single precision than double precision because graphics is usually done in single precision and the GPU cards in entry systems are usually there to support graphics rather than to give a platform for high performance computation.

Answer 2

[Edric Ellis]

Walter's answer covers most of the salient points here. You could also try GPUBench to see how your GPU performs.