k-Wave

1. mabdinur85
   

   Member
   Joined: Jun '13
   Posts: 4

   Hello to all;

   I was looking around for OpenCL based alternatives to GPUMat and saw ViennaCL as an option ... has anyone tried ViennaCL or try to use it in conjunction with MATLAB in general or even with the K-Wave Toolbox?

   CUDA is too GPU specific (only for Nvidia based graphics cards); making OpenCL the platform agnostic approach which is more preferable for a general compute setup although from the benchmarks I've seen it would seem to indicate the specificity found in CUDA setups makes it a high performer.

   ViennaCL website:
   http://viennacl.sourceforge.net/

   Posted 11 years ago #

2. Jiri Jaros
   

   Developer
   Joined: Feb '12
   Posts: 115

   Hi mabdinur85,

   We've got an OpenCL and CUDA version under development and we are likely to release it with the next version of k-Wave. The main problem of GPUs is a small amount of memory they offer. Even Tesla cards with 6GB do not allow us to solve domains bigger than 512x256x256 or so.

   You're right CUDA is GPU specific, however, it's very fast and supports much more features than OpenCL (especially for clusters). It's always a trade off between generality and speed.

   Anyway, thanks for a tip. I'll take a look.
   Cheers,
   Jiri

   Posted 11 years ago #

3. mabdinur85
   

   Member
   Joined: Jun '13
   Posts: 4

   Glad to hear there is work being done on this front. As for ViennaCL I think they implement 3 solvers in MATLAB and provide a library to build against as well. The hardware designers from these manufacturers seemed to be more interested in faster memory as opposed to larger memory. I think you can find older cards with 5GB DDR3 right now at reasonably low rates.

   I found a GT630 model with 4GB DDR3 and 96 CUDA Cores for $77 on newegg. Whereas a Titan series with 6GB GDDR5 and 2688 CUDA Cores goes for roughly $1000 on newegg. I guess more trade offs to consider regarding cores and memory where one bottlenecks the other. I think as you mentioned the need for big memory is essential especially when slower memory is accessed after the graphics memory fills up making the speed up moot at some certain point. Man I wish I had the budget to do a real study on this.

   Posted 11 years ago #

RSS feed for this topic

Reply

You must log in to post.