Straightforward data transfer in a blockwise dataflow for an analog RRAM-based CIM system

Abstract

Analog resistive random-access memory (RRAM)-based computation-in-memory (CIM) technology is promising for constructing artificial intelligence (AI) with high energy efficiency and excellent scalability. However, the large overhead of analog-to-digital converters (ADCs) is a key limitation. In this work, we propose a novel LINKAGE architecture that eliminates PE-level ADCs and leverages an analog data transfer module to implement inter-array data processing. A blockwise dataflow is further proposed to accelerate convolutional neural networks (CNNs) to speed up compute-intensive layers and solve the unbalanced pipeline problem. To obtain accurate and reliable benchmark results, key component modules, such as straightforward link (SFL) modules and Tile-level ADCs, are designed in standard 28 nm CMOS technology. The evaluation shows that LINKAGE outperforms the conventional ADC/DAC-based architecture with a 2.07×∼11.22× improvement in throughput, 2.45×∼7.00× in energy efficiency, and 22%–51% reduction in the area overhead while maintaining accuracy. Our LINKAGE architecture can achieve 22.9∼24.4 TOPS/W energy efficiency (4b-IN/4b-W) and 1.82 ∼4.53 TOPS throughput with the blockwise method. This work demonstrates a new method for significantly improving the energy efficiency of CIM chips, which can be applied to general CNNs/FCNNs.