Memory

The overall memory usage trend is displayed in a line chart. You can select and zoom in on the peak area to precisely identify the processes or operators with high memory consumption. For operators with abnormal memory allocation and release, jump to the Timeline to locate the code.

Memory optimization roadmap: Increase the batch size to maximize the NPU memory usage. Observe the memory usage trends, mitigate spikes, and balance the peak and off-peak hours.

Figure 1 shows that the NPU is underutilized and memory spikes occur.

Figure 1 Typical example

Select operators within the memory spike period. In the memory allocation/release details, sort the operators by requested memory size in descending order. Click the operator ranked first to go to the Timeline and locate the specific code, as shown in Figure 2. Then, communicate with the model development personnel based on the code location to evaluate potential tuning opportunities.

Figure 2 Jumping to Timeline

The Memory page also supports the comparison between two cards. For details, see "Instructions" in MindStudio Insight User Guide.