Next we will explain why TensorFlow enthusiasts are so excited about this new feature of just-in-time compilation via XLA.
The following the key strengths for this feature:
- Server side speedups: Through this JIT compilation and specialization that we have mentioned, we see that some in-house models that TensorFlow team has win up to 60%. Also, SyntaxNet gets latency reductions from around 200 microseconds to 5 microseconds. And the reason for this was SyntaxNet had a lot of small operations in its graph so the interpreter has to go grab each little operation and this process of going and grabbing these small operations and running those causes some latency overhead but by compiling you're actually able to ...