Dear PI Folks,
as a software developer I am also concerned about GPU computing.
While scanning through the forum and the announcements of CUDA computing in future PI Releases, I have
a technical question:
What is the reason for the decision to support only the vendor specific NVidia / CUDA platform ?
What if there will be some other vendors or platforms that could be much more efficient in the future ?
I know that CUDA seems to be much more easier to implement into an existing software project.
But on the other side - you could adress much wider computing platforms if you would take the OpenCl/OpenGl standards.
The current development will - in future - no longer make a difference between CPU, GPU or a HPC computing cluster.
Are the current KHRONOS papers missing something that PI must rely on ?
By the way: I found, that - currently still on many systems - the advantages of GPU computing can be inhibited by the time of data transfer into and from GPU cards,
but this could be eliminated in the future by newer technologies.
Kind regards,
Christoph
