Execution of Resource Intensive Tasks on a Heterogeneous SoC for Low-Latency Embedded Compute

Execution of resource-intensive tasks, such as artificial intelligence (AI), big-data algorithms, video processing, etc. is a common requirement in distributed embedded systems today. The typical solution is to execute such algorithms on some cloud computing service, as embedded systems do not have the computing resources for such tasks. However, this approach is not viable in Real-Time Cyber-Physical Systems (RT-CPS) due to the high-latency, non-real-time nature of cloud computing. The latency requirements can be only satisfied with an edge compute solution in the form of an edge data center or by a heterogeneous System on Chip (SoC) based implementation, which has the required high-performance computing resources on the chip. In the paper, we investigate the viability of the heterogeneous SoC compute approach on the Texas Instruments AM5729 SoC, which was specifically designed for such applications. The paper introduces the hardware capabilities of the SoC, the software development, and the run-time (Linux-based) environment, and presents the results of a detailed performance evaluation on relevant computing tasks.