How do you choose between g4dn, g5, and g6 instances for simulation?

Picking the wrong GPU instance does not just slow your simulation. It can terminate the job entirely, or run it at a fraction of the speed a better instance would have delivered. GPU memory, throughpu