A tool for visualizing the Tenstorrent Neural Network model (TT-NN)
TT-NN Visualizer can be installed from PyPI:
pip install ttnn-visualizer
After installation run ttnn-visualizer to start the application.
It is recommended to do this within a virtual environment. The minimum Python version is 3.10.
Please see the getting started guide for further information on getting up and running with TT-NN Visualizer.
If you want to test out TT-NN Visualizer you can try some of the sample data. See loading data for instructions on how to use this.
For the latest updates and features, please see releases.
- Comprehensive list of all operations in the model
- Interactive graph visualization of operations
- Detailed and interactive L1, DRAM, and circular buffer memory plots
- Filterable list of tensor details
- Overview of all buffers for the entire model run
- Visualization of input and output tensors with core tiling and sharding details
- Visualize inputs/outputs per tensor or tensor allocation across each core
- Detailed insights into L1 peak memory consumption, with an interactive graph of allocation over time
- Navigate a tree of device operations with associated buffers and circular buffers
- Operation flow graph for a holistic view of model execution
- Load reports via the local file system or through an SSH connection
- Supports multiple instances of the application running concurrently
- BETA: Network-on-chip performance estimator (NPE) for Tenstorrent Tensix-based devices
Visualiser-Demo.v4.mp4
| L1 Summary with Tensor highlight | Operation inputs and outputs |
|---|---|
 |
 |
| Device operations with memory consumption | DRAM memory allocation |
|---|---|
 |
 |
| Operation graph view | Model buffer summary |
|---|---|
 |
 |
| Per core allocation details | Per core allocation details for individual tensors |
|---|---|
 |
 |
| Tensor details list | Performance report |
|---|---|
 |
 |
| Performance charts | |
|---|---|
 |
 |
| NPE | |
|---|---|
 |
 |
You may test the application using the following sample reports.
Unzip the files into their own directories and select them with the local folder selector, or load the NPE data on the /npe route.
Segformer encoder memory report
Segformer decoder memory report
Llama mlp memory + performance report
T3K synthetic synthetic_t3k_small.json.zip
How to run TT-NN Visualizer from source.