To assess the total capacity of an A16 GPU, we added tasks for processing inference rules with camera streams until the system queue for new tasks could not be cleared. Each test run performed a mixture of use cases since there are differences in the characteristics that can impact CPU and GPU utilization.
The following simulated camera setup was used for every test run to ensure consistency:
- 1080p HD video resolution
- 3.8 - 4.0 Mb/s streaming bit range
- 70 total cameras for each use case
The maximum camera stream count for a single use case per A16 is listed below:
| Use case | Description | Max cameras per A16 |
| People Counting | Count the number of people entering and exiting a stadium complex access point. | 60 |
| Queue Capacity | Count the number of people waiting in a line. Raise an alert if a threshold count is exceeded. | 60 |
| Overcrowding Detection | Set a limit on the number of people permitted to be in an area. Raise an alert if that limit is reached. | 60 |
| Virtual Perimeter | Define an area on a map of the stadium that unauthorized people are not permitted to enter. | 60 |
| Lost Child | The CPU and GPU utilization did not exceed 70% during any of the above tests. The application is designed so that if scene complexity increases or spikes in load occur, the system has some spare capacity to operate reliably. | 60 |
The CPU and GPU utilization did not exceed 70% during any of the above tests. The application is designed so that if scene complexity increases or spikes in load occur, the system has some spare capacity to operate reliably.