Using Approximate DRAM for Enabling Energy-Efficient, High-Performance Deep Neural Network Inference

The effectiveness of deep neural networks (DNN) has prompted a tremendous demand for energy-efficient high-performance DNN inference systems. Due to the increasing memory intensity of most DNN workloads, main memory can dominate the system’s energy consumption and stall time. One effective way to reduce the energy consumption and increase the performance of DNN inference systems is by using approximate memory, which operates with reduced supply voltage and reduced access latency parameters, at the cost of introducing higher bit error rates. Fortunately, neural networks have an intrinsic capacity to tolerate increased bit errors. This can enable energy-efficient and high-performance neural network inference using approximate DRAM devices. This chapter discusses EDEN by Koppula S et al ((2019) EDEN: Enabling energy-efficient, high-performance deep neural network inference using approximate DRAM. In: Proceedings of the 52nd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO)), Koppula S et al ((2019) EDEN: Enabling energy-efficient, high-performance deep neural network inference using approximate DRAM. arXiv), a recent work that uses this observation to realize higher energy efficiency and performance on DNN inference workloads. We describe how EDEN reduces DNN energy consumption and DNN evaluation latency by using approximate DRAM devices, while strictly meeting a user-specified target DNN accuracy. We evaluate EDEN on multi-core CPUs, GPUs, and DNN accelerators with error models obtained from real approximate DRAM devices. We also discuss how the core concepts of EDEN have inspired state-of-the-art, and the new directions, opportunities, and challenges that EDEN has opened for the approximate computing research community.