Exploring Power Usage and Inference Speed Evaluation of Low-power Clustered Many-core Platforms

High-performance platforms capable of running deep neural network (DNN)-based applications are necessary for embedded systems such as autonomous-driving systems. These systems must be compact and power-efficient, rather than relying on rich-computational power platforms such as graphics processing units (GPUs). Additionally, platforms capable of executing multiple applications in parallel in such large-scale systems are required. Clustered many-core platforms, such as the Kalray massively parallel processor array (MPPA) 3-80 Coolidge, have been designed to meet these requirements. Coolidge employs more computing cores compared with single-core or multi-core processors, allowing for reduced clock frequencies per core. Consequently, Coolidge is a high-performance platform with low-power usage. Additionally, in Coolidge, cores are grouped into clusters, each of which can independently run different applications. This enables a single Coolidge platform to support multiple applications simultaneously. In the realm of cyber-physical systems (CPS), which bridge the physical and digital domains, these platforms become crucial. CPS relies on real-time embedded systems, like those in autonomous vehicles, which necessitate low-power, high-performance platforms that can perform complex computations like DNNs for object detection. In our evaluation, we examine DNN inference speed and explore the performance of the Coolidge platform. The DNN task for the evaluation employs object detection, which is commonly used in autonomous-driving systems. The acceptable speed threshold for real-time applications in object detection is 30 frames per second. Our study reveals that the “you only look once” v5 model can exceed performance benchmarks with only one cluster in Coolidge for int8 data type and two clusters for the higher precision fp16 data type. Moreover, when two clusters are utilized for DNN tasks, the remaining clusters are available for other non-DNN applications, underlining the platform’s flexibility and resource efficiency.