Since RAM doesn’t play an active role in the deep learning computation, the best strategy is to have just enough RAM for your GPUs to perform. If there’s not enough RAM the you’ll find your system ending up disk swapping and slowing down your performance.
The rule of thumb is to match RAM size to the largest GPU’s RAM. For example, if you’re using an GPU with 8 GB of RAM, then you’ll want to add 8 GB of RAM to your motherboard to prevent swapping data from your disk. If you still find disk swapping, you can save money by only add enough RAM as needed to prevent it from happening.