An Approximate DRAM Design with an Adjustable Refresh Scheme for Low-power Deep Neural Networks

—A DRAM device requires periodic refresh operations to preserve data integrity, which incurs significant power consumption. Slowing down the refresh rate can reduce the power consumption; however, it may cause a loss of data stored in a DRAM cell, which affects the correctness of computation. This paper proposes a new memory architecture for deep learning applications, which reduces the refresh power consumption while maintaining accuracy. Utilizing the error-tolerant property of deep learning applications, the proposed memory architecture avoids the accuracy drop caused by data loss by flexibly controlling the refresh operation for different bits, depending on their criticality. For data storage in deep learning applications, the approximate DRAM architecture reorganizes the data so that these data are mapped to different DRAM devices according to their bit significance. Critical bits are stored in more frequently refreshed devices while non-critical bits are stored in less frequently refreshed devices. Compared to the conventional DRAM, the proposed approximate DRAM requires only a separation of the chip select signal for each device in a DRAM rank and a minor change in the memory controller. Simulation results show that the refresh power consumption is reduced by 66.5 % with a negligible accuracy drop on state-of-the-art deep neural networks.