Profiler User Guide

Users familiar with TensorBoard who develop with TensorFlow or PyTorch can start using the respective profiler similar to GPU or TPU. The only change required is substituting HPU for GPU in your sources. No extra guide nor additional readings are required. You will enjoy near GPU-like experience when training your model on HPU. Most users will use the TensorBoard with TensorFlow or PyTorch and utilize the frameworks for both model development and profiling.

A short outline of the basic steps for TensorFlow and PyTorch can be found in Profiling with TensorFlow and Profiling with Pytorch.

Users interested in low-level focused profiling as well as those writing their models in SynapseAI® language without framework can refer to SynapseAI Profiler User Guide.

For details on the profiling control flow, refer to Profiling Architecture. For further training acceleration, see Profiling Tips and Tricks.

• Profiling with TensorFlow
• Profiling with Pytorch
• Profiling with SynapseAI
• Profiling Architecture
• Tips and Tricks to Accelerate the Training

Note

In this chapter, the terms step and batch are used interchangeably unless told otherwise.