I would give the framework developers some credit, they shipped and the field made a lot of progress on their tools. Getting good performance out of GPUs (CUDA or OpenCL) is quite difficult, cuDNN has historically been the only good low-level deep learning library, and NVIDIA makes great hardware. My company is building a fully open source low-level framework called PlaidML bringing OpenCL support and other benefits to the existing frameworks. We're starting with Keras, first code was posted last Friday: https://github.com/plaidml/plaidml