k-Wave

1. marcelr
   

   Member
   Joined: May '14
   Posts: 2

   In their recent cuda tookits, NVidia have published plugin replacement libs for FFTW and BLAS. I want to have a look at these libs, see if they are any good in speeding up things related to k-Wave. I wonder if forum members have experience with those, in particular in combination with k-Wave?

   best,

   Marcel

   Posted 9 years ago #

2. Jiri Jaros
   

   Developer
   Joined: Feb '12
   Posts: 115

   Hi marcelr,

   I've performed two simple benchmarks and the results are not surprising :-)
   

   size     iCore7 920 (4c/8T)     GTX 580        2xIntel E5-2670  (16T)
   128^3   0.09918s/step        0.0967s/step        0.025s/step
   256^3   1.10s/step           0.7347s/step        0.42s/step

   Offloading the FFT on the GPU is not going to speed up k-Wave substantially because the data has to be moved back and forth 14 times per time step. This virtually kills all the benefit gained from the GPU. The only way to get a reasonable speed-up is to run k-Wave entirely on the GPU (this version is being tested).

   The maximum size of the simulation domain I was able to run with a GTX580/1.5GB VRAM was 256^3 (cufft is incredibly memory voracious).

   However, if you have a slow CPU and a powerful GPU, the speed-up is measurable. If interested, I can send you a patch file to be applied before compiling.

   Jiri

   Posted 9 years ago #

3. marcelr
   

   Member
   Joined: May '14
   Posts: 2

   Hi Jiri,

   Thanks for the update. I have the exact same configuration CPU/GPU you tested on. No big speed-ups to be reckoned with, then. I would love to get the patch file anyway, might learn a thing or two.

   Good to hear that a GPU-only version of k-Wave is in the making. Looking forward to it.

   Marcel

   Posted 9 years ago #

RSS feed for this topic

Reply

You must log in to post.