Questions about NVMEs and the GPU

Hello all, I have a few questions related to NVME SSD storage and how it relates to the GPU on the Jetson Orin Nano. So far, I’ve figured out that GPUDirect Storage is how you “connect” the GPU and the NVME.

  1. What NVME’s can be used with the Orin Nano? If anyone has had specific models work with the Orin Nano, it’d be appreciated if you could post them.
  2. Is there any use from a read/write speed perspective to using a NVME with more than PCIe 3 capability with the Orin Nano, since the Orin Nano only supports PCIe 3?
  3. What form does the NVME provide extra memory to the GPU? Is it just like adding extra CUDA global memory? Does it have to be interfaced specially (C++/CUDA)?
  4. Can/should the Jetson OS be flashed onto the NVME? I currently have a flashed 256GB microSD card, can I use that in conjunction with the NVME?

I’m planning to use the NVME and GPUDirect Storage to expand the memory available to the GPU, as I’m hoping to store the depth map of the past frames of a stereo camera and utilise that to do processing on the GPU, and if I had to transfer those frames between the CPU and GPU it would be a huge bottleneck.

Thanks in advance for any help!

This is not supported in GPU. The disk storage can be set to swap space so that CPU can use it as memory extension. But GPU cannot access it. If your use-case require significant memory, please consider Orin Nano 8GB.

So I take it that this document (1. Overview — GPUDirect RDMA 12.4 documentation) stating that GPUDirect can be used from CUDA 11.4 and onwards on Jetson Orin platforms is wrong? Or am I misunderstanding it?

The solution is for FPGA devices instead of NVMe SSD storage. Please check
GitHub - NVIDIA/jetson-rdma-picoevb: Minimal HW-based demo of GPUDirect RDMA on NVIDIA Jetson AGX Xavier running L4T

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.