Open ninnghazad opened 4 years ago
After testing this, by just assigning multiple domain-parts to a single GPU, i'd say the answer is no. Source comments are a bit ambiguous about this, but it will fail trying to link up the domains. So i'll just close this.
Hi,
Sorry -- forgot to reply to this sooner. It isn't something that's ever been tested, and as you say it doesn't work at the minute. However, I see the potential applications and I suspect it would only need a few tweaks to get it working.
Happy to look at making it work if you want. I'll reopen this for now.
L
Hm, nah, i don't need it at the moment - the current solution is to throw more GPUs at problems. It was more of a what-if idea. For cases where adding hardware is not possible, it would be nice addition i assume - if one would be willing to accept probably substantially longer simulation times due to a lot of IO and transferring data from/to GPUs.
is it possible to use a split/multi-domain on a single GPU or is the domain-to-device relation a fixed one? while using multiple gpus works very well, and all the swapping in/out this would result in would probably a performance no-go, i am wondering.