I’m Givin’ Her All She’s Got Captain!

The hardware sets we chose to test were:

1. The neural networks run on a standard x86 processor
2. The algorithms run on an nVIDIA GPU
3. The algorithms run on a Computational Storage NVMe SSD (also called a CSD).

The 5 algorithms we chose to run were:

1. YOLO (You Only Look Once) Lite – used for real time object tracking on portable devices
2. GOTURN – very fast CNN based object tracker
3. KCF (Kernelized Correlation Filters) – fast tracker often used in multiple object tracking
4. MOSSE (Minimum Output Sum of Squared Error) – very fast tracker but lower accuracy
5. WiSARD – Filter based Weightless neural network

You can find the surprising summary of results from our research in Figure 2 below.

Figure: 2 – Object Recognition and Detection at the Edge CPU vs. GPU vs. Computational Storage

What you see in these results is the amount of energy consumed in processing each video frame using the 5 different Object Detection Algorithms in each of the different compute environments. We also wanted to measure accuracy in terms of IoU (Intersection over Union) for each of the runs.

The bottom line is that in almost all cases, using an NVMe Computational Storage SSD to run the Object Detection and tracking algorithms is significantly more energy efficient than either a CPU or a GPU based approach. This efficiency comes WITHOUT sacrificing accuracy or increasing system cost.

So, if you “need more power” for your edge application, the most energy efficient way to increase your system compute capabilities is through the deployment of Computational Storage enabled SSD’s. More storage, more compute, less power, and no additional cost. Something worth looking at.