Hi,
Instead of using the Jacket toolbox for GPU computing in the k-Wave version b04, could I use the Parallel Computing Toolbox to do the same thing?
Thanks,
Chao
Hi,
Instead of using the Jacket toolbox for GPU computing in the k-Wave version b04, could I use the Parallel Computing Toolbox to do the same thing?
Thanks,
Chao
Hi Chao,
In principle, yes. However, the current release of the parallel computing toolbox (2011b) only supports a small number of GPU functions. As this expands to include bsxfun
and fftn
we will definitely add support for k-Wave to utilise this toolbox. We also plan to release an optimised C/C++ version of the 3D simulation function (which will work in conjunction with the MATLAB codes). On a modern multicore CPU, this runs almost as fast as MATLAB + GPU.
Brad.
Hi Chao,
It seems that R2012a has included bsxfun
and fftn
in the list of GPU functions supported by the Parallel Computing Toolbox. We'll work towards supporting 'DataCast'
set to 'GPUArray'
for the next release.
Brad.
I've done a lot of benchmarking of Jacket vs the Parallel Computing Toolbox, and Jacket always ends up much faster. PCT GPU rarely even beats the CPU in my benchmarks. Jacket on a laptop is faster than PCT on a Tesla, http://www.walkingrandomly.com/?p=4062
I say forget the PCT GPU and stick with what is actually fastest.
Hi soundmaster,
Thanks for your comment. I agree that the GPU capabilities of the Parallel Computing Toolbox are currently in their infancy. However, many institutions have site-wide licenses to MATLAB that include access to this toolbox. This is particularly relevant for students who may not be able to purchase a separate license to the Accelereyes product. We are not planning to deprecate compatibility with Jacket (it is a nice product and integrates well with k-Wave). Rather, we plan to add additional compatibility with GPUArray
. I'll post some benchmarking results back here when we start working on this!
Kind regards,
Brad.
You must log in to post.