Streaming dataflow based recurrent neural network accelerator

Abstract

A hardware accelerator system processes recurrent neural networks (RNNs) on edge devices using a streaming architecture that minimizes memory bottlenecks. The system used a matrix multiplication unit for matrix multiplications with an LSTM/GRU accelerator for scalar and gate computations. The matrix multiplication unit processes interleaved weight matrices to compute multiple gate operations in a single pass, while the accelerator includes specialized processing units for arithmetic, activation functions, and recurrent operations. A switching matrix transmits data between components to maintain continuous dataflow. The system supports various RNN architectures including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and minimal GRU (minGRU) variants through configurable datapaths.

Dependencies are inferred from the claim text.