Per-bank refresh with adaptive early termination for high density DRAM

Abstract

DRAM, which is mainly used as main memory, requires a refresh operation to maintain the integrity of stored data. Since memory read and write operations to a bank are not allowed while the bank is being refreshed, a lot of memory accesses may be blocked due to refresh, which may lead to significant performance degradation. Therefore, a lot of active studies to minimize this negative performance impact of refresh have been conducted. In a refresh scheme called per-bank refresh, the refresh unit will be one bank rather than all banks in a rank, allowing memory access to other banks while a certain bank in the same rank is refreshed. However, the per-bank refresh consumes more power than all-bank refresh. In this paper, we propose a per-bank refresh method with adaptive early termination, which allows both the refresh period and the size of each row group to be non-uniformly determined, to increase the efficiency of per-bank refresh and reduce energy consumption. By using this method, compared to the basic all-bank refresh model, the average weighted speed increases by about 6.4% and the energy consumption is reduced by about 51%.