1# Performance 2 3Performance is often a significant issue when training a machine learning 4model. This section explains various ways to optimize performance. Start 5your investigation with the @{$performance_guide$Performance Guide} and then go 6deeper with techniques detailed in @{$performance_models$High-Performance Models}: 7 8 * @{$performance_guide$Performance Guide}, which contains a collection of best 9 practices for optimizing your TensorFlow code. 10 11 * @{$performance_models$High-Performance Models}, which contains a collection 12 of advanced techniques to build highly scalable models targeting different 13 system types and network topologies. 14 15 * @{$performance/benchmarks$Benchmarks}, which contains a collection of 16 benchmark results. 17 18XLA (Accelerated Linear Algebra) is an experimental compiler for linear 19algebra that optimizes TensorFlow computations. The following guides explore 20XLA: 21 22 * @{$xla$XLA Overview}, which introduces XLA. 23 * @{$broadcasting$Broadcasting Semantics}, which describes XLA's 24 broadcasting semantics. 25 * @{$developing_new_backend$Developing a new back end for XLA}, which 26 explains how to re-target TensorFlow in order to optimize the performance 27 of the computational graph for particular hardware. 28 * @{$jit$Using JIT Compilation}, which describes the XLA JIT compiler that 29 compiles and runs parts of TensorFlow graphs via XLA in order to optimize 30 performance. 31 * @{$operation_semantics$Operation Semantics}, which is a reference manual 32 describing the semantics of operations in the `ComputationBuilder` 33 interface. 34 * @{$shapes$Shapes and Layout}, which details the `Shape` protocol buffer. 35 * @{$tfcompile$Using AOT compilation}, which explains `tfcompile`, a 36 standalone tool that compiles TensorFlow graphs into executable code in 37 order to optimize performance. 38 39And finally, we offer the following guide: 40 41 * @{$quantization$How to Quantize Neural Networks with TensorFlow}, which 42 can explains how to use quantization to reduce model size, both in storage 43 and at runtime. Quantization can improve performance, especially on 44 mobile hardware. 45 46