Example: with an Nvidia GPU, I usually only have to change one line of my code.
If I want to use a TPU, I have to uninstall 2 different libraries, re-install everything, reset my runtime, downgrade my ML framework... it is just a pain. TPUs would see much wider adoption if they were as easy to use as GPUs.
Example: with an Nvidia GPU, I usually only have to change one line of my code. If I want to use a TPU, I have to uninstall 2 different libraries, re-install everything, reset my runtime, downgrade my ML framework... it is just a pain. TPUs would see much wider adoption if they were as easy to use as GPUs.