A Cascaded ReRAM-based Crossbar Architecture for Transformer Neural Network Acceleration

Emerging resistive random-access memory (ReRAM) based processing-in-memory (PIM) accelerators have been increasingly explored in recent years because they can efficiently perform in-situ matrix-vector multiplication (MVM) operations involved in a wide spectrum of artificial neural networks. However, there remain significant challenges to apply existing ReRAM-based PIM accelerators to the most popular Transformer neural networks. Since Transformers involve a series of matrix-matrix multiplication (MatMul) operations with data dependencies, they should write intermediate results of MatMuls to ReRAM crossbar arrays for further processing. Conventional ReRAM-based PIM accelerators often suffer from high latency of ReRAM writes and intra-layer pipeline stalls. In this paper, we propose ReCAT, a ReRAM-based PIM accelerator designed particularly for Transformers. ReCAT exploits transimpedance amplifiers (TIAs) to cascade a pair of crossbar arrays for MatMul operations involved in the self-attention mechanism. The intermediate result of a MatMul generated by one crossbar array can be directly mapped to another crossbar array, avoiding costly analog-to-digital conversions. In this way, ReCAT allows MVM operations to overlap with the corresponding data mapping, hiding the high latency of ReRAM writes. Furthermore, we propose an analog-to-digital converter (ADC) virtualization scheme to dynamically share scarce ADCs among a group of crossbar arrays, and thus significantly improve the utilization of ADCs to eliminate the performance bottleneck of MVM operations. Experimental results show that ReCAT achieves 207.3×, 2.11×, and 3.06× performance improvement on average compared with other Transformer acceleration solutions—GPUs, ReBert, and ReTransformer, respectively.