I don't know why that NVIDIA document is wrong, but the established term for doing DMA from eg. an NVMe SSD to a GPU within a single system without the CPU initiating the transfer is peer to peer DMA. RDMA is when your data leaves the local machine's PCIe fabric.