An 8.93 TOPS/W LSTM Recurrent Neural Network Accelerator Featuring Hierarchical Coarse-Grain Sparsity for On-Device Speech Recognition
Deepak Kadetotad, Shihui Yin, Visar Berisha, Chaitali Chakrabarti, Jae-sun Seo
Abstract
Long short-term memory (LSTM) is a type of recurrent neural networks (RNNs), which is widely used for time-series data and speech applications, due to its high accuracy on such tasks. However, LSTMs pose difficulties for efficient hardware implementation because they require a large amount of weight storage and exhibit computation complexity. Prior works have proposed compression techniques to alleviate the storage/computation requirements of LSTMs but elementwise sparsity schemes incur sizable index memory overhead and structured compression techniques report limited compression ratios. In this article, we present an energy-efficient LSTM RNN accelerator, featuring an algorithm-hardware co-optimized memory compression technique called hierarchical coarse-grain sparsity (HCGS). Aided by the HCGS-based blockwise recursive weight compression, we demonstrate LSTM networks with up to 16× fewer weights while achieving minimal error rate degradation. The prototype chip fabricated in 65-nm LP CMOS achieves up to 8.93 TOPS/W for real-time speech recognition using compressed LSTMs based on HCGS. HCGS-based LSTMs have demonstrated energy-efficient speech recognition with low error rates for TIMIT, TED-LIUM, and LibriSpeech data sets.