We can do the same for the GPU build of the executable. This allows us to compare how long an epoch takes to train through the model. As our model is not complex, we don't expect to see that much of a difference:

2018/07/21 23:54:31 Using CUDA build
2018/07/21 23:54:32 Batches 600
2018/07/21 23:54:56 Epoch 0 | cost -0.6914807096357707
2018/07/21 23:55:19 Epoch 1 | cost -0.6917470871356043
2018/07/21 23:55:42 Epoch 2 | cost -0.6918343739257966
2018/07/21 23:56:05 Epoch 3 | cost -0.6918777292080605
2018/07/21 23:56:29 Epoch 4 | cost -0.6919036464362168
2018/07/21 23:56:52 Epoch 5 | cost -0.69192088335746
2018/07/21 23:57:15 Epoch 6 | cost -0.6919331749749763
2018/07/21 23:57:39 Epoch 7 | cost -0.691942382545885
2018/07/21 23:58:02 Epoch 8 | cost -0.6919495375223687
2018/07/21 23:58:26 Epoch 9 | cost -0.691955257565567
2018/07/21 23:58:26 Run Tests
2018/07/21 23:58:32 Epoch Test | cost -0.6896057773382677

On this GPU (an NVIDIA Geforce GTX960), we can see that this is marginally faster for this simple task, at 23–24 seconds.