An option would be to do something similar to Einstein@Home where GPU utilization factor can be set in user preferences, instead of having to make our own `app_info.xml` file. In this way you offload the problem of resource scheduling to the BOINC client, for which it's quite well designed to do, and for free.

Agree!!! That is why I do not run GPU work here very often as I do not customize my config. files for individual projects. The manner in which Einstein does it makes it very easy to tune # wu's/GPU. My 2 cents.
Cheers

For the record, I don't know where the rumor of a minimum 2xxx series card started, but its not true.

We'll be bound by the minimum version of cuda that pytorch supports. For pytorch 1.6, this is 9.2 I think.. so any card that's supported by cuda 9.2 should be supported for this. A high end card shouldn't be necessary, especially since (at the moment) the goal isn't to train start of the art over-parameterized networks on gigabytes of data. A simple 970 or 1050 should be plenty.

I also want to re-iterate that, at the moment, the networks in dataset 1 and 2 (and probably 3, though not actually tested) actually train slower on GPUs than CPUs (maybe with a substantial client rewrite it could be better). Sometimes the networks are so small the overhead of using/transferring data to/from the GPU dwarfs the speedup gained in the actual matrix calculations.