A 8.93-TOPS/W LSTM Recurrent Neural Network Accelerator Featuring Hierarchical Coarse-Grain Sparsity With All Parameters Stored On-Chip

Long short-term memory (LSTM) networks are widely used for speech applications but pose difficulties for efficient implementation on hardware due to large weight storage requirements. We present an energy-efficient LSTM recurrent neural network (RNN) accelerator, featuring an algorithm-hardware co-optimized memory compression technique called hierarchical coarse-grain sparsity (HCGS). Aided by HCGS-based block-wise recursive weight compression, we demonstrate LSTM networks with up to 16x fewer weights while achieving minimal accuracy loss. The prototype chip fabricated in 65-nm LP CMOS achieves 8.93/7.22 TOPS/W for 2-/3-layer LSTM RNNs trained with HCGS for TIMIT/TED-LIUM corpora.