GPU / CPU Benchmarks for tensorflow and keras

The benchmarks run on all the devices that tensorflow finds. If that should include a GPU, make sure to install the python egg tensorflow-gpu, also make sure the nvidia kernel module has the version supported by your cuda installation, in my case I needed the ubuntu-package nvidia-384==384.111-0ubuntu1 (the 111-part also has to match). You'll see an error message in the console (not the notebook) when importing tensorflow and the versions mismatch. Also make sure cuda is in your LD_LIBRARY_PATH, in my case /usr/local/cuda-8.0/targets/x86_64-linux/lib/ for the ubuntu package cuda-8.0, the folder should contain object files like libcudnn.so (updatedb and locate libcuda.so to find such a folder on your linux system).

Run the benchmarks preferrably in a virtualenv with python>=3.4,<4

virtualenv --python=/usr/bin/python3 gpubenchmark
source gpubenchmark/bin/activate
pip3 install numpy scipy tensorflow-gpu keras jupyter
jupyter-notebook # starts a http-server

instead of running jupyter-notebook, you can also make python-scripts.

Experiments

Benchmark for matrix multiplication in tensorflow

Matrix multiplication benchmark

Benchmark for training and predicting on a 5-layer neural network in keras+tensorflow

4-layer Dense Neural Network

Benchmarks for predictions using ResNet50, Inception v3, VGG16 and VGG19

Popular deep learning models