0

I run a Debian Jessie on a Tegra-K1 board (Toradex Apalis-TK1), installed the CUDA toolkit 7.0 and the nvidia_drivers.tbz2 from the Jetson TK1 21.5 package.

Now I get

$ nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2015 NVIDIA Corporation
Built on Fri_Oct_16_15:23:46_CDT_2015
Cuda compilation tools, release 7.0, V7.0.70

$ cat /etc/ld.so.conf.d/nvidia-tegra.conf 
/usr/lib/tegra

$ cat /etc/nv_tegra_release 
# R21 (release), REVISION: 5.0, GCID: 7273100, BOARD: ardbeg,  EABI: hard, DATE: Wed Jun  8 04:19:09 UTC 2016
50e7e42bde409d5ba0eb410d803b991cd6c4c32b */usr/lib/xorg/modules/extensions/libglx.so
cc612c5afcb8a76a5b847fc500d2354582d32abb */usr/lib/xorg/modules/drivers/nvidia_drv.so
e96a78d9c3947a980f065d96557f98c70500db17 */usr/lib/arm-linux-gnueabihf/tegra/libnvomxilclient.so
... and so on ...

I was able to build the NVIDIA_CUDA-7.0_Samples (using make SMS=53 to avoid linking problems), but

1_Utilities/deviceQuery/deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

cudaGetDeviceCount returned 35
-> CUDA driver version is insufficient for CUDA runtime version
Result = FAIL

How can I narrow the cause of the problem?

0

After consulting the developer forum I found out: The error message is correct (this happens sometimes). Cuda toolkit 7 requires the R23 release of the cuda driver, but the latest release available for Tegra K1 is R21.

So you need an older toolkit version, but only version 7 is available for download on the nvidia site. Older versions are only accessible via JetPack, but JetPack is Ubuntu-only (fails to install on Debian).

But they gave me instructions to work around the problem:

  • download JetPack 2.3.1
  • extract the content of the archive: bash JetPack-L4T-2.3.1-linux-x64.run –noexec
  • go to the _installer directory where you will find a Chooser binary
  • ./Chooser -> it may ask for a libpng12 so I installed this lib in my home then: export LD_LIBRARY_PATH+=:path_to_libpng/lib then retry ./Chooser
  • the GUI appears but this is not what is important: you may notice that a repository.json file is created which contains all the link to the cuda, etc packages. Just find the link corresponding to your CUDA version and that's it.

Download the .deb package from the URL given in the json file with wget, install it with dpkg -i, then

apt update
apt search cuda
apt install ...

Obviously, Nvidia wants us to head for TX1 and TX2, not stick to TK1 ...

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.