MolSoft has tested RIDE on two GPUs:
Surprisingly, the RIDE performance on RTX was even better ~10%-15% than on the more expensive Tesla. It can be explained because we don't use double precision arithmetic in
RIDE superposition, so P100 doesn't give you any advantage and the RTX
newer model has higher GPU clock rate.
The algorithm performance linearly depends on the size of your template (number of bonds). 0.5 million conformers/sec performance refers the ligand with ~22 bonds. The GPU RAM required for that ligand size is ~200Mb. For multiple templates everything (speed/RAM requirements) is also scaled linearly.
The other important requirement for the fast RIDE performance is that conformer DB should be located on SDD drive, otherwise the bottleneck will be just reading data from the disk.