Library from George Hotz focused on creating the smallest possible useful deep learning library.
Check the examples dir, for ex the mnist example that only uses 145 lines
# from the root dir of the repo python examples/mnist_gan.py
on my computer each epoch takes about 3 minutes, so training the whole net (300 epochs) would take about 15 hours.
I saw in the source that there is a metal interface, but no idea if it's being used.