Updated 11/28/2018
Here's my experience of installing the NVIDIA CUDA kit 9.0 on a fresh install of Ubuntu Desktop 16.04.4 LTS. Scroll down to the bottom if you wish to only install NVIDIA drivers and run tensorflow via docker container.
Table of contents generated with markdown-toc
Do not use the CUDA run file to install your driver. Use apt-get
instead. This way you do not need to worry about the Nouveau stuff you read about on StackOverflow.
As of 04/11/2018, the latest version of NVIDIA driver for Ubuntu 16.04.4 LTS is 384. To install the driver, excute
sudo apt-get update
sudo apt-get install nvidia-384 nvidia-modprobe -y
Reboot the machine.
Afterwards, you can check the Installation with the nvidia-smi
command, which will report all your CUDA-capable devices in the system.
ERROR: Unable to load the 'nvidia-drm' kernel module.
- One probable reason is that the system is boot from UEFI but Secure Boot option is turned on in the BIOS setting. Turn it off and the problem will be solved.
nvidia-smi -pm 1
can enable the persistent mode, which will save some time from loading the driver. It will have significant effect on machines with more than 4 GPUs.
nvidia-smi -e 0
can disable ECC on TESLA products, which will provide about 1/15 more video memory. Reboot is reqired for taking effect. nvidia-smi -e 1
can be used to enable ECC again.
nvidia-smi -pl <some power value>
can be used for increasing or decrasing the TDP limit of the GPU. Increasing will encourage higher GPU Boost frequency, but is somehow DANGEROUS and HARMFUL to the GPU. Decreasing will help to same some power, which is useful for machines that does not have enough power supply and will shutdown unintendedly when pull all GPU to their maximum load.
-i <GPUID>
can be added after above commands to specify individual GPU.
These commands can be added to /etc/rc.local
for excuting at system boot.
Installing CUDA from runfile is much simpler and smoother than installing the NVIDIA driver. It just involves copying files to system directories and has nothing to do with the system kernel or online compilation. Removing CUDA is simply removing the installation directory. So I personally does not recommend adding NVIDIA's repositories and install CUDA via apt-get
or other package managers as it will not reduce the complexity of installation or uninstallation but increase the risk of messing up the configurations for repositories.
The CUDA runfile installer can be downloaded from NVIDIA's websie, or using wget in case you can't find it easily on NVIDIA:
cd
wget https://developer.nvidia.com/compute/cuda/9.0/Prod/local_installers/cuda_9.0.176_384.81_linux-run
What you download is a package the following three components:
- an NVIDIA driver installer, but usually of stale version;
- the actual CUDA installer;
- the CUDA samples installer;
I suggest extracting the above three components and executing 2 and 3 separately (remember we installed the driver ourselves already). To extract them, execute the runfile installer with --extract
option:
cd
chmod +x cuda_9.0.176_384.81_linux-run
./cuda_9.0.176_384.81_linux-run --extract=$HOME
You should have unpacked three components:
NVIDIA-Linux-x86_64-384.81.run
(1. NVIDIA driver that we ignore),
cuda-linux.9.0.176-22781540.run
(2. CUDA 9.0 installer), and
cuda-samples.9.0.176-22781540-linux.run
(3. CUDA 9.0 Samples).
Execute the second one to install the CUDA Toolkit 9.0:
sudo ./cuda-linux.9.0.176-22781540.run
You now have to accept the license by scrolling down to the bottom (hit the "d" key on your keyboard) and enter "accept". Next accept the defaults.
To verify our CUDA installation, install the sample tests by
sudo ./cuda-samples.9.0.176-22781540-linux.run
Please make sure that
- PATH includes /usr/local/cuda-9.0/bin
- LD_LIBRARY_PATH includes /usr/local/cuda-9.0/lib64, or, add /usr/local/cuda-9.0/lib64 to /etc/ld.so.conf and run ldconfig as root
After the installation finishes, configure the runtime library.
sudo bash -c "echo /usr/local/cuda/lib64/ > /etc/ld.so.conf.d/cuda.conf"
sudo ldconfig
It is also recommended for Ubuntu users to append string /usr/local/cuda/bin
to system file /etc/environments
so that nvcc
will be included in $PATH
. This will take effect after reboot. To do that, you just have to
sudo vim /etc/environment
and then add :/usr/local/cuda/bin
(including the ":") at the end of the PATH="/blah:/blah/blah" string (inside the quotes).
After a reboot
, let's test our installation by making and invoking our tests:
cd /usr/local/cuda-9.0/samples
sudo make
It's a long process with many irrelevant warnings about deprecated architectures (sm_20
and such ancient GPUs). After it completes, run deviceQuery
and p2pBandwidthLatencyTest
:
cd /usr/local/cuda/samples/bin/x86_64/linux/release
./deviceQuery
The result of running deviceQuery
should look something like this:
./deviceQuery Starting...
CUDA Device Query (Runtime API) version (CUDART static linking)
Detected 1 CUDA Capable device(s)
Device 0: "GeForce GTX 1060"
CUDA Driver Version / Runtime Version 9.0 / 9.0
CUDA Capability Major/Minor version number: 6.1
Total amount of global memory: 6073 MBytes (6367739904 bytes)
(10) Multiprocessors, (128) CUDA Cores/MP: 1280 CUDA Cores
GPU Max Clock rate: 1671 MHz (1.67 GHz)
Memory Clock rate: 4004 Mhz
Memory Bus Width: 192-bit
L2 Cache Size: 1572864 bytes
Maximum Texture Dimension Size (x,y,z) 1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)
Maximum Layered 1D Texture Size, (num) layers 1D=(32768), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(32768, 32768), 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: 65536
Warp size: 32
Maximum number of threads per multiprocessor: 2048
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): (2147483647, 65535, 65535)
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Concurrent copy and kernel execution: Yes with 2 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
Supports Cooperative Kernel Launch: Yes
Supports MultiDevice Co-op Kernel Launch: 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 = 9.0, CUDA Runtime Version = 9.0, NumDevs = 1
Result = PASS
Cleanup: if ./deviceQuery works, remember to rm
the 4 files (1 downloaded and 3 extracted).
The recommended way for installing cuDNN is to
-
Download the "cuDNN Library for Ubuntu 16" (need to register for an Nvidia account) and select the right version compatible with cuda 9
-
you should have following files on your system:
libcudnn7_7.4.1.5-1+cuda9.0_amd64.deb
libcudnn7-dev_7.4.1.5-1+cuda9.0_amd64.deb
libcudnn7-doc_7.4.1.5-1+cuda9.0_amd64.deb
- install these:
sudo dpkg -i libcudnn7_7.4.1.5-1+cuda9.0_amd64.deb
sudo dpkg -i libcudnn7-dev_7.4.1.5-1+cuda9.0_amd64.deb
sudo dpkg -i libcudnn7-doc_7.4.1.5-1+cuda9.0_amd64.deb
- Finally, execute
sudo ldconfig
to update the shared library cache.
Select GPU tarball and save it.
- Extract .so files and move to system path.
tar -zxvf libtensorflow-gpu-linux-x86_64-<tab>
sudo chown -R root:root lib
sudo mv lib/lib* /usr/local/lib
sudo ldconfig
- Check if binary that links against libtensorflow is able to find all dynamic dependencies
ldd matrix-inversion-benchmark-tf`
Example output from ldd command:
$ ldd matrix-inversion-benchmark-tf
linux-vdso.so.1 => (0x00007ffdf776e000)
libtensorflow.so => /usr/local/lib/libtensorflow.so (0x00007efe88ebc000)
libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007efe88c9f000)
libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007efe888d5000)
libtensorflow_framework.so => /usr/local/lib/libtensorflow_framework.so (0x00007efe8797a000)
libcublas.so.9.0 => /usr/local/cuda-9.0/lib64/libcublas.so.9.0 (0x00007efe84544000)
libcusolver.so.9.0 => /usr/local/cuda-9.0/lib64/libcusolver.so.9.0 (0x00007efe7f949000)
libcudart.so.9.0 => /usr/local/cuda-9.0/lib64/libcudart.so.9.0 (0x00007efe7f6dc000)
libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007efe7f4d8000)
libgomp.so.1 => /usr/lib/x86_64-linux-gnu/libgomp.so.1 (0x00007efe7f2b6000)
libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007efe7efad000)
librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007efe7eda5000)
libstdc++.so.6 => /usr/lib/x86_64-linux-gnu/libstdc++.so.6 (0x00007efe7ea23000)
libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007efe7e80d000)
/lib64/ld-linux-x86-64.so.2 (0x00007efe972d5000)
libcuda.so.1 => /usr/lib/x86_64-linux-gnu/libcuda.so.1 (0x00007efe7d98f000)
libcudnn.so.7 => /usr/lib/x86_64-linux-gnu/libcudnn.so.7 (0x00007efe6b3df000)
libcufft.so.9.0 => /usr/local/cuda-9.0/lib64/libcufft.so.9.0 (0x00007efe6333e000)
libcurand.so.9.0 => /usr/local/cuda-9.0/lib64/libcurand.so.9.0 (0x00007efe5f3da000)
libnvidia-fatbinaryloader.so.384.130 => /usr/lib/nvidia-384/libnvidia-fatbinaryloader.so.384.130 (0x00007efe5f188000)
- Finally run the benchmark to see awesome power of GPU. A random 10k by 10k matrix inverted in under 8 seconds! Same Go binary linked against CPU tensorflow library on a laptop takes anywhere from 5-10 minutes!
$ ./matrix-inversion-benchmark-tf 2> /dev/null # this runs on machine with GPU
[100 100] 354.559305ms
[100 100] 402.992636ms
[200 200] 7.406717ms
[200 200] 8.793598ms
[500 500] 17.260441ms
[500 500] 27.268701ms
[1000 1000] 49.058466ms
[1000 1000] 88.828322ms
[2000 2000] 159.050976ms
[2000 2000] 333.588065ms
[5000 5000] 1.229218361s
[5000 5000] 2.00629059s
[10000 10000] 4.162459538s
[10000 10000] 7.302393948s
Same binary running without GPU
$ ./matrix-inversion-benchmark-tf 2> /dev/null
[100 100] 27.505568ms
[100 100] 6.989513ms
[200 200] 6.123381ms
[200 200] 9.456749ms
[500 500] 14.438066ms
[500 500] 46.444771ms
[1000 1000] 39.278103ms
[1000 1000] 282.240379ms
[2000 2000] 148.83378ms
[2000 2000] 2.016059554s
[5000 5000] 1.113634783s
[5000 5000] 28.653253206s
[10000 10000] 6.156776647s
[10000 10000] 3m57.371907035s
Do not use the CUDA run file to install your driver. Use apt-get
instead. This way you do not need to worry about the Nouveau stuff you read about on StackOverflow.
As of 04/11/2018, the latest version of NVIDIA driver for Ubuntu 16.04.4 LTS is 384. To install the driver, excute
sudo apt-get install nvidia-384 nvidia-modprobe
Reboot the machine.
Afterwards, you can check the Installation with the nvidia-smi
command, which will report all your CUDA-capable devices in the system.
# Add the package repositories
curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | \
sudo apt-key add -
distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | \
sudo tee /etc/apt/sources.list.d/nvidia-docker.list
sudo apt-get update
# Install nvidia-docker2 and reload the Docker daemon configuration
sudo apt-get install -y nvidia-docker2
sudo pkill -SIGHUP dockerd
# Test nvidia-smi with the latest official CUDA image
sudo docker run --runtime=nvidia --rm nvidia/cuda:9.0-base nvidia-smi
sudo docker pull tensorflow/tensorflow:1.12.0-gpu
This does not have libtensorflow_framework.so
and libtensorflow.so
, so these two dynamic dependencies need to be installed from here (select the GPU supported tarball and extract contents of lib
folder into /usr/local/lib
sudo docker run --runtime=nvidia -v /home/sdeoras:/home/sdeoras --entrypoint /bin/bash -it tensorflow/tensorflow:1.12.0-gpu
Running the matrix inversion binary from within the container is now possible
# ldd ./matrix-inversion-benchmark-tf
linux-vdso.so.1 => (0x00007ffc0b1fe000)
libtensorflow.so => /usr/local/lib/libtensorflow.so (0x00007fb593ed1000)
libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007fb593cb4000)
libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007fb5938ea000)
libtensorflow_framework.so => /usr/local/lib/libtensorflow_framework.so (0x00007fb59298f000)
libcublas.so.9.0 => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcublas.so.9.0 (0x00007fb58ed12000)
libcusolver.so.9.0 => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcusolver.so.9.0 (0x00007fb58a117000)
libcudart.so.9.0 => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcudart.so.9.0 (0x00007fb589eaa000)
libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007fb589ca6000)
libgomp.so.1 => /usr/lib/x86_64-linux-gnu/libgomp.so.1 (0x00007fb589a84000)
libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007fb58977b000)
librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007fb589573000)
libstdc++.so.6 => /usr/lib/x86_64-linux-gnu/libstdc++.so.6 (0x00007fb5891f1000)
libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007fb588fdb000)
/lib64/ld-linux-x86-64.so.2 (0x00007fb5a22ea000)
libcuda.so.1 => /usr/lib/x86_64-linux-gnu/libcuda.so.1 (0x00007fb58815d000)
libcudnn.so.7 => /usr/lib/x86_64-linux-gnu/libcudnn.so.7 (0x00007fb576c56000)
libcufft.so.9.0 => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcufft.so.9.0 (0x00007fb56ebb5000)
libcurand.so.9.0 => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcurand.so.9.0 (0x00007fb56ac51000)
libnvidia-fatbinaryloader.so.384.130 => /usr/lib/x86_64-linux-gnu/libnvidia-fatbinaryloader.so.384.130 (0x00007fb56a9ff000)
Finally execute the binary that links against TF libraries:
$ ./matrix-inversion-benchmark-tf 2> /dev/null
[100 100] 369.287241ms
[100 100] 478.682456ms
[200 200] 8.72297ms
[200 200] 9.57128ms
[500 500] 19.873331ms
[500 500] 31.00355ms
[1000 1000] 56.71972ms
[1000 1000] 99.010054ms
[2000 2000] 189.614255ms
[2000 2000] 358.203112ms
[5000 5000] 1.334024582s
[5000 5000] 2.202555466s
[10000 10000] 4.340356657s
[10000 10000] 8.713958014s