Description
The NVIDIA DLProf Plugin for TensorBoard makes it easy to visualize the performance of your models by showing Top 10 operations that took the most time, eligibility of Tensor Core operations and Tensor Core usage as well as interactive iteration reports.
Requirements
- Requires GPU Event files generated by DLProf v0.9.0 / r20.02
- Ensure you have access and are logged into NGC. For step-by-step instructions, see the NGC Getting Started Guide.
- Install Docker and nvidia-docker. For DGX users, see Preparing to use NVIDIA Containers. For users other than DGX, see nvidia-docker installation documentation.
New Features
The key features of DLProf Plugin for TensorBoard v0.1 / r20.02 are:
- Enhanced plugin for visualizing GPU profile data on Tensorboard
- New Summary page in a panelized dashboard arrangement
- Introduction of two new key metrics in the first summary panel
- TC Utilization % shows the percentage of GPU time spent executing Tensor Core kernels over the total GPU time spent by Tensor Core eligible operations.
- GPU Idle % shows the amount of time the GPU spent not executing any kernel over the entire aggregated time range for the GPU.
- Pie charts showing Node and Kernel summaries for quick visual evaluation.
- Interactive Tables in all reports that allow a user to sort, filter, and paginate the results.
- Remaining reports are equivalent to the previous GPU Summary plugin for TensorBoard.
Known Issues
- This software is only accessible in the NGC TensorFlow container.
- This software is only supported for TensorFlow 1.15 and TensorBoard 1.15.
- TensorBoard will only display the default "tensorflow-core" domain in this release.
Resolved Issues
- None