I have few clarification on RDMA using GPUDirect. I am primarily interested in copying data across devices within the same machine (For e.g. GPU and a NIC/storage card). I understand the idea of pinning GPU memory (given the virtual address) but I am not sure on how to use the GPU physical address to transfer data. Do I need to use the address to program the DMA controller in the device that I am interested in (NIC/storage) to copy to/from GPU? If this the case, any reference on how to achieve this would be really helpful? Are there any sample driver code that makes use of GPUDirect for zero-copy to/from GPU available?
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| GPUDirect RDMA for Windows | 2 | 1193 | October 4, 2022 | |
| RDMA from local host memory to remote GPU memory? | 1 | 702 | April 10, 2019 | |
| Python Cupy and DirectGPU | 2 | 428 | April 10, 2025 | |
| How to memcpy in linux driver directly to GPU? | 3 | 1806 | October 28, 2022 | |
| Copy to CUDA GPU Memory from a PCI Device | 2 | 1007 | June 12, 2013 | |
| I have a few questions about GPU Direct RDMA | 3 | 920 | December 1, 2022 | |
| GPUDirect RDMA on proprietary interconnect | 0 | 405 | March 24, 2020 | |
| How to force CUDA to use DMA for memcpy | 3 | 3813 | November 27, 2018 | |
| GPU Direct 2.0 (GPU to external device) GPUDirect | 0 | 9338 | June 8, 2011 | |
| First using the gpu-direct RDMA | 0 | 458 | October 19, 2021 |