Adaptive Early Termination Refresh to Reduce Refresh Energy for Low-Power DRAM

Abstract

The main memory of modern IT devices mainly uses DRAM to store data. DRAM consists of a MOS transistor and a capacitor. The transistor acts as a switch and the capacitor serves to store the data. It is ideal to keep the data permanently while power is maintained, but data is lost because DRAM uses capacitors. Due to the specific of the capacitor, the charged charge leaks over time, so periodic charge recharging is necessary to maintain the reliability of the data. This operation is called refresh operation, and access to the memory cell is blocked during the refresh operation. The deterioration of DRAM performance caused by memory blocking phenomenon is one of the major problems of modern DRAM. Therefore, it is important to increase the efficiency of the refresh operation.

In this thesis, we propose adaptive early termination refresh based on per-bank refresh to improve the efficiency of refresh operation. In the per-bank refresh method a row group which is refresh unit is limited to a single bank in the DRAM. Therefore, memory access is possible for banks other than the refresh target bank. On the other hand, since the refresh operation is sequentially performed for each bank, the time and energy consumption for completing the refresh operation of all the banks increases. Therefore, in order to prevent increase of energy consumption and to increase refresh efficiency, adaptive early termination refresh method is used in which size is specified according to retention time of row group and refresh cycle is newly designated for each row group. Flag bits which are located at the top of each row group are added as an indicator of size and retention time. Therefore, the DRAM reads the flag bits of the first row through refresh and determines whether to refresh afterwards. By using this method, compared to the basic per-bank refresh model, the average weighted speed increases by about 3.5% and the energy consumption is reduced by about 59%.