Hi,
I have 2 questions I hope some of you might be able to help me with.
I am currently simulating QLMs using DMRG. I have recently made the switch to GPU (this was very easy, thanks to the devs for that) but due to an A100 only having 80gb of vRam, need to use virtual unified memory since my MPS get pretty large. This however gives me a severe bottleneck in the bandwith between GPU and RAM. For small systems without vunified memory i get a ~40 times speedup, with unified memory, more like 3 times.
However the node i am running this on has 4 A100s with nvlink and I am currently only using one.
I now have 2 questions:
Can I in my current setup easily utilize all 4 A100s?
Can I use unified memory in such a way that only the 4*80=320 gb from the 4 A100 are used?