GPU

Introduction to GPU

Vector processing units in graphic cards (GPU) allow for very fast processing, provided the program makes the most of its parallel features. The most popular platform for GPU computing is Nvidia CUDA.

Enable GPU in Scipion

If you foresee that the end user will have access to a GPU when using the image, then you should prepare all the GPU-related stuff.

First, NVidia driver:

# Download the latest stable driver.
# Disable nouveau:
echo "blacklist nouveau" > /etc/modprobe.d/disable-nouveau.conf
# Update init disk
update-initramfs -u
# reboot
# install dependencies
apt-get install linux-headers-$(uname -r) build-essential
# run the binary installer
chmod u+x NVIDIA-Linux-x86_64-361.42.run
./NVIDIA-Linux-x86_64-361.42.run

Then, install CUDA. We recommend CUDA 5.5 (which requires gcc 4.6)

./cuda_5.5.22_linux_64.run -toolkit -silent

The last requisite is OpenCV, which can be installed with scipion install script (you may need to specify gcc/g++ 4.6 in CMakeCache.txt).

The CUDA switch will be enabled later in scipion.conf.

# GPU features require manual change in config/scipion.conf:
CUDA = True

(see How to install)

EM packages that benefit from GPU

Motioncorr

./scipion install -j 4  motioncorr

GPU purchase guide