| Description |
The demand for main memory capacity has been increasing for many years and will continue to do so. In the past, Dynamic Random Access Memory (DRAM) process scaling has enabled this increase in memory capacity. Along with continued DRAM scaling, the emergence of new technologies like 3D-stacking, buffered Dual Inline Memory Modules (DIMMs), and crosspoint nonvolatile memory promise to continue this trend in the years ahead. However, these technologies will bring with them their own gamut of problems. In this dissertation, I look at the problems facing these technologies from a current delivery perspective. 3D-stacking increases memory capacity available per package, but the increased current requirement means that more pins on the package have to be now dedicated to provide Vdd/Vss, hence increasing cost. At the system level, using buffered DIMMs to increase the number of DRAM ranks increases the peak current requirements of the system if all the DRAM chips in the system are Refreshed simultaneously. Crosspoint memories promise to greatly increase bit densities but have long read latencies because of sneak currents in the cross-bar. In this dissertation, I provide architectural solutions to each of these problems. We observe that smart data placement by the architecture and the Operating System (OS) is a vital ingredient in all of these solutions. We thereby mitigate major bottlenecks in these technologies, hence enabling higher memory densities. |
