Dear Janice,
>
> I am test running eddy_cuda (5.0.11 pre-release) on a Linux machine with an NVIDIA 510 GPU card. It brought the processing time down from ~6h on 4 CPUs to ~3h on 1 GPU. Is this the expected speedup for my GPU card?
>
> I recall Jesper mentioning somewhere on the forum that he had a 20 minute processing time for eddy running on a Mac equipped with GPU. What GPU hardware was it run on? Should I expect the same processing time? Is there something additional that I need to configure for my GPU to run efficiently with eddy_cuda (I am a beginner at GPU computing)?
it is hard to say exactly what to expect in terms of speedup. The 20 minutes was for a “standard” 64 direction, 2mm cubed data set acquired here at FMRIB. As for an HCP data set I haven’t even attempted that on my laptop. In general I only use the laptop to test small snippets of GPU code, and not for analysis of data.
It sounds from above like you have ~8 times speedup compared to a single CPU. The speedups I have seen on our system are more like ~20 times. This is for “standard” eddy. I know that the GPU code for the slice-to-vol achieves lower occupancy (it uses the GPU less efficiently). It is always hard to know how much effort one should put into optimising these things. It takes a lot of time and effort.
As for the slight discrepancy between your speedup and what I see, I can’t really say what that is caused by. It could just be as simple as you having faster CPUs and/or slower GPUs than we have. It sounds like you are in the right ballpark though. A factor of ~2 different isn’t inexplicable.
Jesper
>
> Cheers,
> Jan
|