author: ricardo chaves

... Copy memory paradigm allows a GPU to directly access host memory over PCIe or NVLink interconnect by pinning a memory region in host memory and mapping it to the GPU. A memory region allocated with Unified Memory via cudaMallocManaged ...