Parallelizing Bank-level Fine Granularity Refresh with Column Access Operation using Split Row Buffer

Abstract

DRAM requires refresh operations to maintain data integrity. As the density of DRAM increases, system performance degradation caused by refresh is getting more serious because column access commands are blocked during the refresh operation. In this paper, we propose a novel refresh method called Bank-Level Fine Granularity Refresh (BFGR) that minimizes the performance degradation due to refresh. In BFGR, the refresh granularity is finer by applying Fine Granularity Refresh (FGR) of DDR4 at the bank level to reduce the stall time that column access commands are blocked by a refresh operation. If a special row buffer called split row buffer is used, a column access operation and a refresh operation can be simultaneously carried out, and BFGR employs this split row buffer to parallelize column access operations with a refresh operation. Our evaluation for DRAM devices with a 64ms retention time shows that the performance improvement of the proposed scheme over the conventional DRAM system on memory-intensive applications is 13.6% for 16Gb devices and 17.1% for 32Gb devices on a multi-core system, respectively.