XLA accelerates both inference and training computations. It can, for example, fuse together many operations to perform them more efficiently and with less memory traffic. It's a compiler and doesn't really care what kind of computation it's being asked to optimize.

→ View original post on X — @jeffdean,