Host memory (CUDA DLLs loaded into ram or pagefiles). If I have 20 processes and each uses 5GB of ram, I would need 100GB of RAM+pagefiles to be able to run. See picture below.
.
