High Throughput Neural Network based Embedded Streaming Multicore Processors
With power consumption becoming a critical processor design issue, specialized architectures for low power processing are becoming popular. Several studies have shown that neural networks can be used for signal processing and pattern recognition applications. This study examines the design of memristor based multicore neural processors that would be used primarily to process data directly from sensors. Additionally, we have examined the design of SRAM based neural processors for the same task. Full system evaluation of the multicore processors based on these specialized cores were performed taking I/O and routing circuits into consideration. The area and power benefits were compared with traditional multicore RISC processors. Our results show that the memristor based architectures can provide an energy efficiency between three and five orders of magnitude greater than that of RISC processors for the benchmarks examined.