Cool! It's too bad about the cuDNN comparison bit people mention (as with dgacmu, I contend that all frameworks now devolve into the low level kernel used), but

Cool! It's too bad about the cuDNN comparison bit people mention (as with dgacmu, I contend that all frameworks now devolve into the low level kernel used), but I'm both impressed by the release and appreciate all the Microsoft folks clearly disclaiming their affiliation. Disclosure: I work on Google Cloud (but not directly in/on TensorFlow).