nvidia-smi problems with 4.0rc


Just tried installing CUDA 4.0rc on a machine with a bunch of 580GTX cards in it. nvidia-smi no longer seems to work neither for getting GPU/Memory utilization or setting GPU compute modes:

Under CUDA 3.2, nvidia-smi -L produced:

GPU 0: (108010DE:83791043) GeForce GTX 580 (S/N: 679871489)
GPU 1: (108010DE:83791043) GeForce GTX 580 (S/N: 679874625)
GPU 2: (108010DE:83791043) GeForce GTX 580 (S/N: 679870561)
GPU 3: (108010DE:83791043) GeForce GTX 580 (S/N: 679872641)
GPU 4: (108010DE:83791043) GeForce GTX 580 (S/N: 868290561)
GPU 5: (108010DE:83791043) GeForce GTX 580 (S/N: 679814145)
GPU 6: (108010DE:83791043) GeForce GTX 580 (S/N: 679871633)
GPU 7: (108010DE:83791043) GeForce GTX 580 (S/N: 868289537)

With 4.0rc, we get:

Unable to determine the product UUID for gpu 0: N/A

nvidia-smi -q -a will list all GPU’s, but will produce N/A for most fields, including “Utilization” and “Compute mode”.

Also, trying to set compute modes will fail:

nvidia-smi -c 1 -i 0

Compute mode is not supported for GPU 0:12:0.

What’s going wrong here? I hope this nvidia-smi functionality hasn’t been limited to Tesla and Quadro cards only!


I am seeing the same thing.

From the CUDA 4.0RC nvidia-smi man page “NVSMI provides diagnostic information for NVIDIA Tesla and Quadro devices.” This is truly a sad thing.