Effective compensation of IR drop in reram-based DNN inference accelerators

Abstract

As artificial intelligence (AI) performance improves, the complexity of AI has increased exponentially in recent years. To solve the ‘memory bottleneck’ problems caused by data movement in the Von-Neumann computer architecture, the Processing-in memory (PIM) technology, which processes data in the memory unit and transfers only necessary data, is attracting attention. Among PIM Memory, ReRAM is receiving more attention because Matrix-Vector Multiplication (MVM) operations performed on crossbar arrays play a similar role to the multiplication and accumulation operation of Neural Network. However, the ReRAM has non-ideal characteristics such as IR Drop, thermal noise, shot noise, and Stuck-At-Fault (SAF), which negatively affects the use of ReRAM as a neural network accelerator. In this work, we study the IR Drop. We analyzed the effect of IR Drop on the accuracy of neural network in ReRAM PIM and point out that IR Drop is a non-negligible factor. The previous study uses a re-training by reflecting the expected value of IR Drop to the weight value. However, it requires many resources such as memory, time, and power consumption for re-training, and a change in the weight values always accompanies it. To solve this problem, we propose methods that can minimize the change in neural network accuracy due to IR Drop by using two aspects of the algorithm and the hardware without performing re-training even in an inference environment. Our compensation experimental results show an accuracy drop within -1.4% compared to the non-IR Drop at the worst-IR Drop of 35%.