Abstract
Long Short-Term Memory (LSTM) recurrent networks are frequently used for tasks involving time sequential data such as speech recognition. However, it is difficult to deploy these networks on hardware to achieve high throughput and low latency because the fully-connected structure makes LSTM networks a memory-bounded algorithm. Previous work in LSTM accelerators either exploited weight spatial sparsity or temporal sparsity. In this paper, we present a new accelerator called "Spartus" that exploits spatio-temporal sparsity to achieve ultra-low latency inference. The spatial sparsity was induced using our proposed pruning method called Column-Balanced Targeted Dropout (CBTD) that leads to structured sparse weight matrices benefiting workload balance. It achieved up to 96% weight sparsity with negligible accuracy difference for an LSTM network trained on a TIMIT phone recognition task. To induce temporal sparsity in LSTM, we create the DeltaLSTM by extending the previous DeltaGRU method to the LSTM network. This combined sparsity saves on weight memory access and associated arithmetic operations simultaneously. Spartus was implemented on a Xilinx Zynq-7100 FPGA. The per-sample latency for a single DeltaLSTM layer of 1024 neurons running on Spartus is 1 us. Spartus achieved 9.4 TOp/s effective batch-1 throughput and 1.1 TOp/J energy efficiency, which are respectively 4X and 7X higher than the previous state-of-the-art.
Abstract (translated)
URL
https://arxiv.org/abs/2108.02297
