Try to get some time to play with PyCUDA

Nvidia's got a GPGPU library called CUDA.  There's a Python binding.  Thinking it would be interesting to try it out if I can get some time.  It's a C++ wrapper, looks reasonably elegant and seems to integrate nicely with numpy (just going by the documentation).

Not that I actually have anything to do with the GPGPU at the moment, I just think it would be interesting to learn.


  1. Rene Dudfield

    Rene Dudfield on 06/23/2008 10:57 p.m. #

    yeah, CUDA is cool - except it's only nvidia.

    Otherwise GLSL is probably the way to go.

    The latest nvidia gpus have 260 cores, and the latest ATI GPUs have 320.

    You can put 4 of them in one machine... That's around 1000 cores -- in a desktop box.

  2. Mike Fletcher

    Mike Fletcher on 06/25/2008 6:42 a.m. #

    I was thinking of using it to reverse engineer requirements for PyOpenGL. Since I don't use GPGPU stuff myself I don't know what kind of support people need to get GPGPU tasks done.

  3. berlinguyinca

    berlinguyinca on 07/29/2008 10:30 p.m. #

    well the current pycuda git version

    ( (fork by me, with several enhancements)

    works very well and makes it pretty easy to work with cuda. It provides by now the complete math module to run on the gpu and can have reach speed ups of a factor of 5k, depending on array size.

Comments are closed.


Pingbacks are closed.