Troubleshooting Faster RCNN object detection performance on AWS EC2

Question

I am running a Faster RCNN object detection system written in Python on an AWS EC2 instance (T2.large). The input is a video feed from a simple security camera. Usually the object detection system takes about 3 second to process a 512x512 video frame. However, from time to time, the system slows to a crawl, taking 20 sec or more per frame.

I think it may be related to CPU credit per hour limit imposed by AWS. T2.large is allocated 36 CPU credits per hour, equivalent to 36 minutes of 100% CPU utilization per hour.

When I run the command "top -i", it shows the python program is pushing CPU utilization to 200%. If this corresponds to the instance's 2 vCPUs, then it means the python program is using all the CPU resources, then AWS is forced to throttle the performance after about 36 minutes.

While I know training an object detection model would require GPU, I do not believe running a model in inference mode would require high end computing platform as I have seen android based real time object detection system running on cell phones.

Any thoughts on why my python program is using 100% of the CPU capacity? I have added a time.sleep(0.01) statement for all the while loops, and it has no impact on CPU utilization rate.