Compilers for Machine Learning

Currently the main thing I do is working on a compiler for machine learning. It's taking your TensorFlow model and getting it running on a TPU. So now that we have TPUs and phones too, it's like, how can we get it even potentially to compile fast enough you could compile it on device? Compilers are pretty much always not Iobound. That would be amazing. But that's not the world we live in.