This is a question for hardware experts, and I am not one of those
Nevertheless, it may depend very much on the type of simulation running, if the system has to switch all the time between CPU and GPU, for example. In this case, simulation with a larger amount of particles would benefit more. Also, tools measuring OpenGL or DirectX benchmark values would not be appropriate, because only CUDA-calculations count here. In my case with the melting simulation and around 700k particles, the "GPU load" shown in the GPU-Z tool oscillated always between 0 and 30%. Memory consumption is relatively low with only 1.6GB.
https://photos.app.goo.gl/miy3BS6ReXbfjh228