[CP2K:1891] Re: Energy is NaN with CUFFT

[Ondrej Marsalek]

On Tue, Mar 17, 2009 at 17:12, Axel <akoh... at gmail.com> wrote:
>
>>
>> > there is no support for CUDA with parallel compiles.
>>
>> Oh, that explains things. Would perhaps be nice to mention it
>> somewhere near the CUDA stuff in CP2K, let's say the example arch file
>> or the cuda_tools directory.
>
> please feel free to submit a patch that changes it
> accordingly. cp2k is an open source project and
> contributions from _everybody_ are welcome.

I am well aware of that, but I don't think that I am able to
contribute in a systematic way at the moment. Therefore I at least try
to provide some feedback. If someone with write access to the
repository and actively involved in the development chooses to act on
it, it is their own choice.

> i thought it was clear from our previous discussions,
> that CUDA support in cp2k is in pre-alpha status and
> that you cannot expect anything to work other than the
> test runs that we did on our machines and that
> documentation is likely to be incomplete, misleading or
> non-existing.

Indeed, this was clear. This, by itself, does not mean that users are
not supposed to comment on it or suggest changes. If this is the case,
please let me know and I will shut up.

> due to the reasons explained before, CUDA
> support in cp2k is not exactly a high priority project
> for us. the gains are just too little (compared to, e.g.
> coarse grained classical MD where we get >20x speedups).

This is understood as well. Even if nobody worked on it at all, that
would be fine with me, I don't expect any guarantees. I try to offer
my opinion on bits and pieces that I encounter and would like to see
changed, but that is all.

> for cp2k you are much better off buying a couple of
> intel nehalem cpus than some GPUs.

And it seems like this will be the direction for us as well.

Best regards,
Ondrej