nmf: resource usage on CPU #336
Description
If we are just running on CPU (which seems pretty fine), then we are basically using one process, one CPU core.
cNMF can be easily sharded over replicates. Would it be faster to shard replicates across devices and have lightning use multiple devices? I'm not sure lightning will automatically do this for us in a nice way. Basically this would be the same as us just using multiprocessing for the steps that have a for-loop over replicates, and run those in different processes...
Look into this.