High performance AXI-4.0 based interconnect for extensible smart memory cubes

The recent technological breakthrough represented by the Hybrid Memory Cube is on its way to improve bandwidth, power consumption, and density. This is while heterogeneous 3D integration has provided another opportunity for revisiting near memory computation to fill the gap between the processors and memories even further. In this paper, we take the first step towards a “Smart Memory Cube (SMC)”, a fully backward compatible and modular extension to the standard HMC, supporting near memory computation on its Logic Base (LoB), through a high performance interconnect designed for this purpose. The main feature of SMC is the high bandwidth, low latency, and AXI-4.0 compatible interconnect, designed to serve the huge bandwidth demand by HMC´s serial links, and to provide extra bandwidth to a processor-in-memory (PIM) embedded in the Logic Base (LoB). Our results obtained from cycle accurate simulation demonstrate that this interconnect can easily meet the demands of current and future projections of HMC (Up to 87GB/s READ bandwidth with 4 serial links and 16 memory vaults, and 175GB/s with 8 serial links and 32 memory vaults, for injected random traffic). Moreover, the interference between the PIM traffic and the main links was found to be negligible with execution time increase of less than 5%, and average memory access time increase of less than 15% when 56GB/s bandwidth is requested by the main links and 15GB/s bandwidth is delivered to the PIM port. Moreover, preliminary logic synthesis with Synopsys Design Compiler confirms that our interconnect is implementable and realistic.