IPOCIM: Artificial Intelligent Processor with Adaptive Ping-pong Computing-in-Memory Architecture

Abstract

Computing-in-memory (CIM) architecture is a promising solution toward energy-efficient artificial intelligent (AI) processor. Practically, the AI processor with CIM engine induces a series of issues including data updating and flexibility. For instance, in AI-oriented applications, the weight stored in the CIM must be reloaded due to the huge gap between limited capacity of CIM and growing weight parameter, which greatly reduces the computation efficiency of the AI processor. Moreover, the natural parallelism of CIM leads to the mismatch of various convolution kernel sizes in different networks and layers, which reduces hardware utilization efficiency. In this work, we explore a CIM engine with a ping-pong strategy as an alternative to traditional CIM macro and weight buffer, hiding the data update latency to enhance data reuse. In addition, we proposed a flexible CIM architecture adapting to different neural networks, namely IPOCIM, with a fine-grained data-flow mapping strategy. Based on the evaluation, IPOCIM achieves 1.4-7.1× performance improvement, and 2.2-6.1× energy efficiency, compared to baseline.