My apologies. I was just beginning to give my answer (in my longwinded fashion) when a pressing issue forced me to cut the communication short. Having only what I wrote to work with, your response is very reasonable. I will now continue what I was hoping to say last time with more brevity...
The reason for the system RAM is for the purpose of loading the scene onto the GPU. Insufficient system RAM will create slow downs in the rendering process.
The GPU memory is imperative because if the entire scene does not 'fit' in the GPU memory the scene reverts to the system RAM and the render time slows significantly, sometimes by an order of magnitude (read hours). From your reply it looks like you already have an understanding of this. So... my answer concerning GPU memory would be in the form of a question. How complex are the scenes you plan to create? The benefits gained by the cuda cores become moot if you do not have sufficient GPU memory to load your entire scene.
The short answer to your question is: Make sure that you have enough GPU memory for your planned scenes and then go for as much cuda core benefit as you can afford. Again, I apologize for not getting back to you sooner. The distraction was unavoidable.
Thank you for your patience, Jack.