Confused about system RAM and GPU RAM

[frozen]

I read that DL uses the memory in the GPU. Thus, a computer with 8-16GB is good enough. However, in Kaggle, I heard people using 64, 128 or even 256GB RAM in the computer. This seems to be contradictory. Can anybody please clarify this? Is it recommended to get 64 or even 128GB RAM? Thanks

 

[i386]

I think the more ram you have the less you will need to access the persistent storage (still way slower than ram).
For homeusers running one machine probably less important than companies running racks full with machines or clusters with shared datasets.

 

[frozen]

What is the recommended amount of RAM? 32? 64 or 128GB? I guess there is a threshold above which would be better to use cloud services.