Megh and Intel® performed a performance and validation assessment of Megh Video Analytics Solution (VAS) running on Intel® NUC kits with 11th or 12th generation Intel® Core™ processors. The kits are cost-effective, small-form factor hardware that provide customers with the performance, power, and accuracy they need to run Megh VAS at the edge. Results demonstrate that VAS enables customers to leverage NUC kits to scale solutions with long-term stability. It also defines the specifications required for CPU and iGPU hardware to be deployed for scaling VAS with NUC kits.
The total number of streams supported by a system depends on camera resolution (e.g., FHD 1080p or HD 720p) and video stream frame rate (e.g., 30 FPS or lower). The resolution and frame rate required for a solution depend on the use cases deployed. Most physical security use cases, like intrusion and loiter detection, only require 720p resolution at 5 FPS. Other use cases, like license plate recognition, may require higher resolutions and frame rates. We demonstrated that by utilizing both the CPU and iGPU simultaneously to analyze video streams, we can scale the solution by varying the resolution and frame rate of the camera streams.
VAS supports down-sampling of video stream FPS at the inference engine, which allows for more streams to be analyzed on a system by configuring the video stream to a lower FPS than the original. For instance, a 25 or 30 FPS input stream can be configured to process at 5 FPS after decoding of the video streams. Also, independent of the input resolution, images are resized to 640×480 pixels before they are analyzed by the inference engine to detect objects. Thus, the change in resolution from 1080p to 720p has less impact on throughput compared to reducing the frame rate to 5 or 10 FPS.
By managing both resolution and frame rate in the processing pipeline, VAS is able to deliver very high throughput on the NUC kits.
Summary of key results:
- The maximum FPS using only CPUs to implement the complete video analytics pipeline scales linearly with the number of cores, from 35 FPS on an Intel® Core™ i3 processor to 140/150 FPS on an Intel® Core™ i9 processor, while overall CPU utilization remains low.
- The maximum FPS using iGPU to implement the inferencing while using the CPU to decode the video streams and perform analytics after detection scales linearly from 83 FPS on an Intel® Core™ i3 processor to 167 on an Intel® Core™ i7 processor, but drops to 87 FPS for the Intel® Core™ i9 processor (because of the fewer number of iGPU execution units available). CPU utilization remains low in this configuration also.
- On an Intel® NUC 12 Pro Kit NUC12WSHi7 with an Intel® Core™ i7- 1260P processor, CPU-only throughput is 66.5 FPS and iGPU-only throughput is 112.0 FPS. This translates to a maximum of 13 streams and 22 streams, respectively, at 5 FPS. By utilizing both the CPU and iGPU, up to 35 video streams can be analyzed with VAS on the NUC12WSHi7 kit.
Megh VAS and Intel® NUC kits with 11th or 12th generation Intel® Core™ processors work exceptionally well together.
For a deeper look, read the full performance and validation report.
