At-Scale Assessment of Weight Clustering for Energy-Efficient Object Detection Accelerators

DNN-based object detection operates on large data volumes to fetch images andDNNweights, which leads to high power and bandwidth demands. Solutions to mitigate those demands, such as weight clustering, are normally studied in limited examples of a much smaller scale than target applications, which poses difficulties to determine the best tradeoff to implement. This paper performs an atscale (using a real life application) assessment of weight clustering for a DNN-based object detection system - You Only Look Once (YOLO) - considering real driving videos. Our case study shows that an Output Stationary accelerator (e.g. a systolic array) restricting weights to only between 32 (5-bit) and 256 (8-bit) different values allows preserving the accuracy of the original 32-bit weights of YOLO while decreasing bandwidth requirements to around 30%40% of the original bandwidth, and overall energy consumption to around 45% of the original consumption. Overall, our case study provides key insights on which to take design decisions for an accelerator for camera-based object detection.