Since I upgraded to the CUDA 4.0 development drivers, I started getting
“Failed to initialize NVML: Unknown Error” whenever I run nvidia-smi. Is anyone else getting this?
I haven’t had that one, but nvidia-smi from the 4.0rc dev drivers is basically 100% broken on consumer GPUs at the moment. Tim Murray said it will be fixed in the next 4.0 release.
Thanks. Sounds like I missed the memo, then.
I get the same error with Cuda 3.2 on GTX 580. Only solution is to reboot the machine.
Hi,
I have the same problem on a system with two NVIDA Quadro 6000 GPUs:
NVIDIA: could not open the device file /dev/nvidia1 (Input/output error).
Failed to initialize NVML: Unknown Error
The weird thing is that we have several identical machines and the errors occurs only on a couple of them, the others run just fine (same driver, same toolkit).
Driver: 275.09.07
Cuda Toolkit: 4.0
Cheers, Sandra
CUDA 4.0 final solved this issue for me with GTX 480. But when I added a GTX 580 to the system, the problem reappeared. I have a ticket open with NVIDIA.
Could you let us know as soon as you get feedback (any kind)? Thanks!
Do you see this message all the time, even right after reboot ?
Yes.
Using driver 275.19
cudatoolkit_4.0.17_linux_64_rhel6.0
Failed to initialize NVML: Unknown error
on GTX285s & GTX480s compute nodes 2 GPUs/node
using SYS V init.d script
But functions fine on my headnode using xorg & 1 Quadro FX 380
NVIDIA suggests trying the just posted beta drivers 280.13. I’ll report my results when the problematic machine is back online (currently down waiting for a replacement part to arrive).
Thanks DrAnderson42 for the advisory,
I now have nvidia-smi on all my nodes: 480’s and 285’s
I have also a previous ver. 275.21 working on computes with C2050s
I will soon replace the 285s with 580s.
If I have problems I will report here.
Driver: 280.13 and 275.89
Card : Quadro 4000
Windows 7 Enterprise and Windows 2008 HPC Server
In all mentioned configurations I get the same error:
Failed to initialize NVML: Unknown Error
Please help!
Oh, sorry, I didn’t see that this was Linux specific.
Anyway, I solved the problem. I need a second vga card by nvidia. The display adapter must not be by another manufacturer if you want to use that program. I can accept this, but why not return an appropriate Error-Message which would tell me that.
Nevertheless CUDA rocks!
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.