When I am assembling a machine learning workstation, what is the most appropriate ratio between my host memory and GPU memory?
