Performance Metrics - 2023.2 English

AI Engine Tools and Flows User Guide (UG1076)

Document ID
UG1076
Release Date
2023-12-04
Version
2023.2 English

The Performance Metrics view shows the percentage of each type of stall against total simulation time or total stall count.

Figure 1. Performance Metrics View
  1. Click on the Performance Metrics view.
  2. Choose the stall type from the drop-down list. The metrics separate different types of stalls. It will show the existing stalls in the simulation result.
    Lock Stall Time (ps)
    Percentage of lock stall time during which buffers in the AI Engine tile are being acquired.
    Lock Stall Count
    The number of lock stalls in a specific tile.
    Memory Stall Time (ps)
    Percentage of memory stall time that memory accesses in the AI Engine tile have conflicts.
    Memory Stall Count
    The number of memory stalls in a specific tile.
    Stream Stall Time (ps)
    Percentage of stream stall time due to full or empty stream.
    Stream Stall Count
    The number of stream stalls in a specific tile.
    Cascade Stall Time (ps)
    Percentage of cascade stream stall time that either due cascade stream is full or empty.
    Cascade Stall Count
    The number of cascade stream stalls in a specific tile.
  3. Each AI Engine tile is shown as a bar in the Performance Metrics view. The higher the stall percentage, the more towards the right the bar will be. More attention should be payed to the highest bars. Click one of the bars to select the tile to focus.
  4. The Tiles view at the bottom lists all the AI Engine tiles with the information about column, row, kernels, buffers, and all the stall times, percentage, as well as count. You can click on the title row to sort on a specific column.
    Tip: When the number or information in the view is in green, it can be cross-probed with other views.
  5. There is a % button beside the drop-down list. By clicking on this button, you can switch between displaying the stall in percentage or absolute value.
  6. While exploring the information in the Performance Metrics view, it is usually helpful to cross-probe with other views. For example:
    Figure 2. Trace View