I have connected a NVIDIA T4 GPU on a Jetson Orin Nano via a PCIex4 interface. Orin is running on JetPack 6 (Ubunt 22.04). This discrete GPU is detected automatically and listed on the lspci. I tried to install the NVIDIA Data Center Driver for ARM64, where T4 is on the Supported Products list. Installing driver via the Debian package didn’t seem to generate any errors. However, the nvidia-smi command always only listed the iGPU from the Orin. Later I updated the CUDA toolkit to 12.6, which went rather smooth, too.
If I have to rely on nvidia-smi
to return an available GPU or to set CUDA_VISIBLE_DEVICE
, it seems I cannot send any code to run on the T4 as it is yet to be visible to nvidia-smi
.
What driver shall I use? How do I make this dGPU visible to CUDA? Any suggestions are welcome.