FBsim and the Fully Buffered DIMM Memory System Architecture

Abstract

As DRAM device data rates increase in chase of ever increasing memory request rates, parallel bus limitations and cost constraints require a sharp decrease in load on the multi-drop buses between the devices and the memory controller, thus limiting the memory system's scalability and failing to meet the capacity requirements of modern server and workstation applications. A new technology, the Fully Buffered DIMM architecture is currently being introduced to address these challenges. FB-DIMM uses narrower, faster, buffered point to point channels to meet memory capacity and throughput requirements at the price of latency. This study provides a detailed look at the proposed architecture and its adoption, introduces an FB-DIMM simulation model - the FBSim simulator - and uses it to explore the design space of this new technology - identifying and experimentally proving some of its strengths, weaknesses and limitations, and uncovering future paths of academic research into the field.