A ton of GPU workloads require leaving large amounts of RAM resident on the GPU and running computation with some new data from the CPU.