Watt: A Write-Optimized RRAM-Based Accelerator for Attention

Attention-based models, such as Transformer and BERT, have achieved remarkable success across various tasks. However, their deployment is hindered by challenges such as high memory requirements, long inference latency, and significant power consumption. One potential solution for accelerating attention is the utilization of resistive random access memory (RRAM), which exploits process-in-memory (PIM) capability. However, existing RRAM-based accelerators often grapple with costly write operations. Accordingly, we exploit a write-optimized RRAM-based accelerator dubbed Watt for attention-based models, which can further reduce the number of intermediate data written to the RRAM-based crossbars, thereby effectively mitigating workload imbalance among crossbars. Specifically, given the importance and similarity of tokens in the sequences, we design the importance detector and similarity detector to significantly compress the intermediate data K-T and V written to the crossbars. Moreover, due to pruning numerous vectors in K-T and V, the number of vectors written to crossbars varies across different inferences, leading to workload imbalance among crossbars. To tackle this issue, we propose a workload-aware dynamic scheduler comprising a top-k engine and remapping engine. The scheduler first ranks the total write counts of each crossbar and the write counts for each inference using the top-k engine, then assigns inference tasks to the crossbars via the remapping engine. Experimental results show that Watt averagely achieves 6.5x, 4.0x, and 2.1x speedup compared to the state of the art accelerators Sanger, TransPIM, and Retransformer. Meanwhile, it averagely achieves 18.2x, 3.2x, and 2.8x energy savings with respect to the three accelerators.