I attempted yesterday to do my first ever installation of CUDA.
The only machine I have available with Nvidia Graphics has been standing in a closet for quite some time. I dusted of the worst cobwebs and was somewhat surprised to see it start at all. Some drives are dead, and the fans on the GFX needed quite a bit of love to start spinning - but you run with what you got.
Server is running Ubuntu 16.04 x86_64, which upgraded to “current” without problem.
I started installing cuda using the instructions here: CUDA Toolkit 11.7 Update 1 Downloads | NVIDIA Developer
It was all good until I rebooted after the ‘apt install cuda’ step. After that reboot the GFX was no longer starting as it should.
Looking through dmesg I saw alot of lines stating that the nvidia-440 driver is unable to start the NVIDIA Corporation GF104 [GeForce GTX 460] card. The error message also states that nvidia-390 is the way to go.
I nuked some 440-packages and installed 390-packages instead, which made the machine boot with working graphics, but broke the cuda install.
./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
Looks to me like the cuda meta package has a bit narrow minded dependency chain:
# for package in cuda cuda-10-2 cuda-runtime-10-2 cuda-drivers; do apt-cache depends $package; done
cuda
Depends: cuda-10-2
cuda-10-2
Depends: cuda-toolkit-10-2
Depends: cuda-runtime-10-2
Depends: cuda-demo-suite-10-2
cuda-runtime-10-2
Depends: cuda-drivers
Depends: cuda-libraries-10-2
Depends: cuda-license-10-2
cuda-drivers
Depends: nvidia-440
Depends: nvidia-440-dev
Depends: libcuda1-440
Depends: nvidia-opencl-icd-440
Depends: <libopencl1>
ocl-icd-libopencl1
nvidia-libopencl1-304
nvidia-libopencl1-340
nvidia-libopencl1-384
nvidia-libopencl1-387
nvidia-libopencl1-390
nvidia-libopencl1-396
nvidia-libopencl1-410
nvidia-libopencl1-418
nvidia-libopencl1-440
Depends: nvidia-modprobe
Depends: nvidia-settings
Conflicts: cuda-drivers-diagnostic
Replaces: cuda-drivers-diagnostic
Much like nvidia-libopencl1 is accepted in every shape and color, should it not be possible to satisfy the driver dependencies with 390-packages?
How else am I supposed to do the installation to get CUDA runtime that works with 390 driver?
CUDA Zone - Library of Resources | NVIDIA Developer lists GTX 460 as supported.
current software does not support that GF104 Fermi card. It is too old.
The last GPU driver version that supported that card was 390.xx and the last CUDA version that supported it was CUDA 8.0
Go back to the setup where you had the R390 driver installed. Then install CUDA 8 using the runfile install method, while deselecting the driver install.
You can get CUDA 8 runfile installer from CUDA Toolkit 11.7 Update 1 Downloads | NVIDIA Developer by following the legacy releases button
1 Like
Ah, ok, thanks Robert!
I’ll try the runfile, but my previous remark on the apt dependency chain still feels valid. CUDA8 is in the repo, so is the older driver, but apt is actively prevented from doing a working install:
# for package in cuda-8-0 cuda-runtime-8-0 cuda-drivers; do apt-cache depends $package; done
cuda-8-0
Depends: cuda-toolkit-8-0
Depends: cuda-runtime-8-0
Depends: cuda-demo-suite-8-0
cuda-runtime-8-0
Depends: cuda-nvrtc-8-0
Depends: cuda-nvgraph-8-0
Depends: cuda-cusolver-8-0
Depends: cuda-cublas-8-0
Depends: cuda-cufft-8-0
Depends: cuda-curand-8-0
Depends: cuda-cusparse-8-0
Depends: cuda-npp-8-0
Depends: cuda-cudart-8-0
Depends: cuda-drivers
Depends: cuda-license-8-0
cuda-drivers
Depends: nvidia-440
Depends: nvidia-440-dev
Depends: libcuda1-440
Depends: nvidia-opencl-icd-440
Depends: <libopencl1>
ocl-icd-libopencl1
nvidia-libopencl1-304
nvidia-libopencl1-340
nvidia-libopencl1-384
nvidia-libopencl1-387
nvidia-libopencl1-390
nvidia-libopencl1-396
nvidia-libopencl1-410
nvidia-libopencl1-418
nvidia-libopencl1-440
Depends: nvidia-modprobe
Depends: nvidia-settings
Conflicts: cuda-drivers-diagnostic
Replaces: cuda-drivers-diagnostic
Once you have your 390 driver install done and working, the apt command that I would use to install CUDA at that point is
sudo apt-get install cuda-toolkit-8-0
You may still run into dependency issues. You can push through those if you want. There are plenty of questions on this forum discussing various dependency issues and how to work through them. The runfile installer may offer a quicker path to get things set up. Given that the last and probably only CUDA version you would use on this GPU is 8.0, the runfile install method seems to offer few downsides from my perspective.
1 Like
Installing using the script worked fine!
I ran
# sh cuda_8.0.61_375.26_linux-run
# sh cuda_8.0.61.2_linux-run
and created:
# cat /etc/ld.so.conf.d/cuda-8-0.conf
/usr/local/cuda-8.0/lib64
And after building a new deviceQuery it seem to work fine:
$ ./deviceQuery
./deviceQuery Starting...
CUDA Device Query (Runtime API) version (CUDART static linking)
Detected 1 CUDA Capable device(s)
Device 0: "GeForce GTX 460"
CUDA Driver Version / Runtime Version 9.1 / 8.0
CUDA Capability Major/Minor version number: 2.1
Total amount of global memory: 960 MBytes (1006108672 bytes)
( 7) Multiprocessors, ( 48) CUDA Cores/MP: 336 CUDA Cores
GPU Max Clock rate: 1430 MHz (1.43 GHz)
Memory Clock rate: 1800 Mhz
Memory Bus Width: 256-bit
L2 Cache Size: 524288 bytes
Maximum Texture Dimension Size (x,y,z) 1D=(65536), 2D=(65536, 65535), 3D=(2048, 2048, 2048)
Maximum Layered 1D Texture Size, (num) layers 1D=(16384), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(16384, 16384), 2048 layers
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total number of registers available per block: 32768
Warp size: 32
Maximum number of threads per multiprocessor: 1536
Maximum number of threads per block: 1024
Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
Max dimension size of a grid size (x,y,z): (65535, 65535, 65535)
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Concurrent copy and kernel execution: Yes with 1 copy engine(s)
Run time limit on kernels: Yes
Integrated GPU sharing Host Memory: No
Support host page-locked memory mapping: Yes
Alignment requirement for Surfaces: Yes
Device has ECC support: Disabled
Device supports Unified Addressing (UVA): Yes
Device PCI Domain ID / Bus ID / location ID: 0 / 1 / 0
Compute Mode:
< Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >
deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 9.1, CUDA Runtime Version = 8.0, NumDevs = 1, Device0 = GeForce GTX 460
Result = PASS
Still using the 390 driver:
$ nvidia-smi
Sun Feb 2 11:42:00 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 390.116 Driver Version: 390.116 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 GeForce GTX 460 Off | 00000000:01:00.0 N/A | N/A |
| 40% 33C P12 N/A / N/A | 479MiB / 959MiB | N/A Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 0 Not Supported |
+-----------------------------------------------------------------------------+
Thanks again Robert!