Stream-Thru Processing Scheme for IoT Computing Platform
Shuji Sannomiya, Senri Yoshikawa, Makoto Iwata, Hiroaki Nishikawa · 2025
This paper presents a processor architectural scheme named Stream- Thru processing scheme (STPS) that makes it possible to achieve a guaranteed throughput which is essential for making IoT devices robust against any dynamic load fluctuation. Data-Driven processing scheme (DDPS), which is categorized as non-von Neumann architecture, exhaustively exploits the parallelism inherent in data streams without any runtime overheads such as context-switching and process scheduling and thus is in principle a promising architectural candidate for im-plementing IoT computing platforms in which processing power is scarce. DDPS passively initiates the execution of operations whenever their operands have arrived; therefore, in practice, the number of data flowing in a processor may exceed a given capacity, and an expected throughput may become unachievable. In this paper, we point out the architectural issue that the exhaustive exploitation of parallelism may become excessive, especially for data streams whose latent parallelism is generous. To innovatively resolve this issue, we propose STPS that equips DDPS with a virtual in-lining capability that faithfully buffers any unwanted increase in the flowing data amount. Experimental results show that STPS can guarantee the expected throughput precisely and an area-performance efficiency of more than three times higher can resultantly become achievable with the virtual in-lining capability than without it.