qinchun/cpu-vs-gpu-benchmark-on-mnist

compare training duration of CNN with CPU (i7 8550U) vs GPU (mx150) with CUDA depending on batch size

CPU-vs-GPU-benchmark-on-MNIST

System: i7 8550U (4 cores), 16 GB ram, Geforce MX150 (2GB), windows 10 using CUDA toolkit 8.0.16, CuDNN 8.0, python 3.5, keras 2.1.2 with tensorflow 1.4.0, visual studio 2015. The CUDA installation is easiest done by following the instructions in link .

Training of neural networks can be accelerated by using the parallelism of calculations on GPUs. But by how much, for a typical small convolutional neural network (4 convolutional layers, one fully connected layer)? Let's find out! Here I compare training duration of a CNN with CPU or GPU for different batch sizes (see ipython notebook in this repo). The GPU load is monitored in an independent program (GPU-Z).

Here's the result:

We can see that the GPU calculations with Cuda/CuDNN run faster by a factor of 4-6 depending on the batch sizes (bigger is faster).

Edit: I tried training the same notebook on a Tesla K80 in the cloud, which can be accessed for free (!!!) via google colab (https://colab.research.google.com/notebooks/welcome.ipynb). The setup is quick and easy, tensorflow is already installed etc. The training ran another factor of 2 faster than my local GPU, and avoid driving my poor laptop so hard, so I'll probably switch to that. Thanks google!