Please need help : Access to an NVMe SSD inside CUDA kernel through RDMA


I plugged a Samsung V-NAND SSD 970 Pro in my Xavier AGX’ M.2 Key M. My SSD is well recognized and I and I followed the steps of this site ( to make that my SSD get mounted automatically at boot of the Xavier.

Given the large amount of data on which my kernels will work, I would like to be able to store my results directly in the SSD using RDMA. In other words, I would like to be able to allocate memory in the SSD (using a DMA), pass the pointer pointing to the area allocated to my CUDA kernel so that the threads can write to it.

Is this doable?
Could you suggest some ideas for me?

I saw the sample on the FPGA (, but I don’t see how to apply it to my use case.


Hi roseMink,

This sample is not relevant for your use-case, that all code related to GPUDirect RDMA is not supported NVME SSDs, since this is the first time seeing this use case, I have no idea if it can work, may other developers share experiences with you if they done similar things.