It will depend on the multi-GPU scaling properties of these applications. It would probably be best to look at benchmark data provided by the teams who maintain these applications. Many report simulation performance in ns/day for a variety of different systems across a large-ish universe of hardware platforms. If you cannot find appropriate data, you can also ask in the user forums / mailing lists for these applications. You are more likely to find domain experts there than in these forums.
My gut instinct is informed by Seymour Cray’s famous quip “It is easier to plow a field with a pair of oxen than 1024 chickens”: in the absence of data suggesting otherwise, using fewer more powerful processors is usually the best way to go. Communication and coordination overhead tends to be expensive (time, power).
The amount of memory on each individual GPU may also be a limiting factor to the size of simulations; you might want to look into that as well.
The CPU requirements of molecular dynamics applications differ substantially, with some using the GPU almost exclusively, while others purposefully try to split the work between GPU and CPU. If you have advanced visualization requirements, that may shift the balance. Perform due diligence to ensure your system doesn’t become bottlenecked by a slow CPU or small system memory.