Configurable emulated shared memory architecture for general purpose MP-SOCs and NOC regions

Emulated shared memory (ESM) multiprocessor systems on chip (MP-SOC) and network on chip (NOC) regions are efficient general purpose computing engines for future computers and embedded systems running applications unknown at the design phase. While they provide programmer a synchronous, unified, and constant time accessible shared memory, the existing ESM architectures have been shown to be inefficient with workloads having low parallelism. In this paper we outline a configurable emulated shared memory (CESM) architecture that retains the advantages of the ESM architectures for parallel enough code but is also able to execute applications with low parallelism efficiently. This happens by allowing multiple threads to join as a single nonuniform memory access (NUMA) bunch and organizing memory system to support NUMA-like behavior for thread-local data if parallelism is limited. Performance simulations as well as silicon area and power consumption estimations of CESM MP-SOC/ NOC regions are provided.