Intelligent Co-operative Processor-in-memory Architectures

Advances in VLSI technology are enabling the processor-memory integration to bridge the processor-memory performance gap. It is also a key driver in the innovation of a new concept called Processor-In-Memory (PIM). The work described in this paper capitalises on the extensive work carried out on PIMs in general and develops a road map for an intelligent revision of a PIM architecture referred to as Co-operative Intelligent Memory (CIM). The journey made to reach the goal of achieving a CIM is taken via the route of developing a Cooperative Pseudo Intelligent Memory (CPIM), as proof of concept and mid point in the ratification of the intelligence needed for a full CIM implementation. Both architectures use a hierarchical two level CPU structure referred to as major and minor CPUs. By partitioning computation through dividing workload between major and minor CPUs in an intelligent manner and without any pre-processor compilation or kernel task scheduling, the PIM system can be made more efficient and co-operative for class of tasks, which are heavily reliant on memory-to-memory iterative processes. The proposed architectures exploit the key feature in the iterative process by using vectors that characterize the iteration. The process of identifying intelligently these vectors is described in this paper. In addition, the performance of the proposed architectures has been evaluated