Header

UZH-Logo

Maintenance Infos

Energy-Efficient Recurrent Neural Network Accelerators for Real-Time Inference


Gao, Chang. Energy-Efficient Recurrent Neural Network Accelerators for Real-Time Inference. 2022, University of Zurich, Faculty of Science.

Abstract

Over the past decade, Deep Learning (DL) and Deep Neural Network (DNN) have gone through a rapid development. They are now vastly applied to various applications and have profoundly changed the life of hu- man beings. As an essential element of DNN, Recurrent Neural Networks (RNN) are helpful in processing time-sequential data and are widely used in applications such as speech recognition and machine translation. RNNs are difficult to compute because of their massive arithmetic operations and large memory footprint. RNN inference workloads used to be executed on conventional general-purpose processors including Central Processing Units (CPU) and Graphics Processing Units (GPU); however, they have un- necessary hardware blocks for RNN computation such as branch predictor, caching system, making them not optimal for RNN processing. To accelerate RNN computations and outperform the performance of conventional processors, previous work focused on optimization methods on both software and hardware. On the software side, previous works mainly used model compression to reduce the memory footprint and the arithmetic operations of RNNs. On the hardware side, previous works also designed domain-specific hardware accelerators based on Field Pro- grammable Gate Arrays (FPGA) or Application Specific Integrated Circuits (ASIC) with customized hardware pipelines optimized for efficient pro- cessing of RNNs. By following this software-hardware co-design strategy, previous works achieved at least 10X speedup over conventional processors. Many previous works focused on achieving high throughput with a large batch of input streams. However, in real-time applications, such as gaming Artificial Intellegence (AI), dynamical system control, low latency is more critical. Moreover, there is a trend of offloading neural network workloads to edge devices to provide a better user experience and privacy protection. Edge devices, such as mobile phones and wearable devices, are usually resource-constrained with a tight power budget. They require RNN hard- ware that is more energy-efficient to realize both low-latency inference and long battery life. Brain neurons have sparsity in both the spatial domain and time domain. Inspired by this human nature, previous work mainly explored model compression to induce spatial sparsity in RNNs. The delta network algorithm alternatively induces temporal sparsity in RNNs and can save over 10X arithmetic operations in RNNs proven by previous works.
In this work, we have proposed customized hardware accelerators to exploit temporal sparsity in Gated Recurrent Unit (GRU)-RNNs and Long Short-Term Memory (LSTM)-RNNs to achieve energy-efficient real-time RNN inference. First, we have proposed DeltaRNN, the first-ever RNN accelerator to exploit temporal sparsity in GRU-RNNs. DeltaRNN has achieved 1.2 TOp/s effective throughput with a batch size of 1, which is 15X higher than its related works. Second, we have designed EdgeDRNN to accelerate GRU-RNN edge inference. Compared to DeltaRNN, EdgeDRNN does not rely on on-chip memory to store RNN weights and focuses on reducing off-chip Dynamic Random Access Memory (DRAM) data traffic using a more scalable architecture. EdgeDRNN have realized real-time inference of large GRU-RNNs with submillisecond latency and only 2.3 W wall plug power consumption, achieving 4X higher energy efficiency than commercial edge AI platforms like NVIDIA Jetson Nano. Third, we have used DeltaRNN to realize the first-ever continuous speech recognition sys- tem with the Dynamic Audio Sensor (DAS) as the front-end. The DAS is a neuromorphic event-driven sensor that produces a stream of asyn- chronous events instead of audio data sampled at a fixed sample rate. We have also showcased how an RNN accelerator can be integrated with an event-driven sensor on the same chip to realize ultra-low-power Keyword Spotting (KWS) on the extreme edge. Fourth, we have used EdgeDRNN to control a powered robotic prosthesis using an RNN controller to replace a conventional proportional–derivative (PD) controller. EdgeDRNN has achieved 21 μs latency of running the RNN controller and could maintain stable control of the prosthesis. We have used DeltaRNN and EdgeDRNN to solve these problems to prove their value in solving real-world problems. Finally, we have applied the delta network algorithm on LSTM-RNNs and have combined it with a customized structured pruning method, called Column-Balanced Targeted Dropout (CBTD), to induce spatio-temporal sparsity in LSTM-RNNs. Then, we have proposed another FPGA-based accelerator called Spartus, the first RNN accelerator that exploits spatio- temporal sparsity. Spartus achieved 9.4 TOp/s effective throughput with a batch size of 1, the highest among present FPGA-based RNN accelerators with a power budget around 10 W. Spartus can complete the inference of an LSTM layer having 5 million parameters within 1 μs.

Abstract

Over the past decade, Deep Learning (DL) and Deep Neural Network (DNN) have gone through a rapid development. They are now vastly applied to various applications and have profoundly changed the life of hu- man beings. As an essential element of DNN, Recurrent Neural Networks (RNN) are helpful in processing time-sequential data and are widely used in applications such as speech recognition and machine translation. RNNs are difficult to compute because of their massive arithmetic operations and large memory footprint. RNN inference workloads used to be executed on conventional general-purpose processors including Central Processing Units (CPU) and Graphics Processing Units (GPU); however, they have un- necessary hardware blocks for RNN computation such as branch predictor, caching system, making them not optimal for RNN processing. To accelerate RNN computations and outperform the performance of conventional processors, previous work focused on optimization methods on both software and hardware. On the software side, previous works mainly used model compression to reduce the memory footprint and the arithmetic operations of RNNs. On the hardware side, previous works also designed domain-specific hardware accelerators based on Field Pro- grammable Gate Arrays (FPGA) or Application Specific Integrated Circuits (ASIC) with customized hardware pipelines optimized for efficient pro- cessing of RNNs. By following this software-hardware co-design strategy, previous works achieved at least 10X speedup over conventional processors. Many previous works focused on achieving high throughput with a large batch of input streams. However, in real-time applications, such as gaming Artificial Intellegence (AI), dynamical system control, low latency is more critical. Moreover, there is a trend of offloading neural network workloads to edge devices to provide a better user experience and privacy protection. Edge devices, such as mobile phones and wearable devices, are usually resource-constrained with a tight power budget. They require RNN hard- ware that is more energy-efficient to realize both low-latency inference and long battery life. Brain neurons have sparsity in both the spatial domain and time domain. Inspired by this human nature, previous work mainly explored model compression to induce spatial sparsity in RNNs. The delta network algorithm alternatively induces temporal sparsity in RNNs and can save over 10X arithmetic operations in RNNs proven by previous works.
In this work, we have proposed customized hardware accelerators to exploit temporal sparsity in Gated Recurrent Unit (GRU)-RNNs and Long Short-Term Memory (LSTM)-RNNs to achieve energy-efficient real-time RNN inference. First, we have proposed DeltaRNN, the first-ever RNN accelerator to exploit temporal sparsity in GRU-RNNs. DeltaRNN has achieved 1.2 TOp/s effective throughput with a batch size of 1, which is 15X higher than its related works. Second, we have designed EdgeDRNN to accelerate GRU-RNN edge inference. Compared to DeltaRNN, EdgeDRNN does not rely on on-chip memory to store RNN weights and focuses on reducing off-chip Dynamic Random Access Memory (DRAM) data traffic using a more scalable architecture. EdgeDRNN have realized real-time inference of large GRU-RNNs with submillisecond latency and only 2.3 W wall plug power consumption, achieving 4X higher energy efficiency than commercial edge AI platforms like NVIDIA Jetson Nano. Third, we have used DeltaRNN to realize the first-ever continuous speech recognition sys- tem with the Dynamic Audio Sensor (DAS) as the front-end. The DAS is a neuromorphic event-driven sensor that produces a stream of asyn- chronous events instead of audio data sampled at a fixed sample rate. We have also showcased how an RNN accelerator can be integrated with an event-driven sensor on the same chip to realize ultra-low-power Keyword Spotting (KWS) on the extreme edge. Fourth, we have used EdgeDRNN to control a powered robotic prosthesis using an RNN controller to replace a conventional proportional–derivative (PD) controller. EdgeDRNN has achieved 21 μs latency of running the RNN controller and could maintain stable control of the prosthesis. We have used DeltaRNN and EdgeDRNN to solve these problems to prove their value in solving real-world problems. Finally, we have applied the delta network algorithm on LSTM-RNNs and have combined it with a customized structured pruning method, called Column-Balanced Targeted Dropout (CBTD), to induce spatio-temporal sparsity in LSTM-RNNs. Then, we have proposed another FPGA-based accelerator called Spartus, the first RNN accelerator that exploits spatio- temporal sparsity. Spartus achieved 9.4 TOp/s effective throughput with a batch size of 1, the highest among present FPGA-based RNN accelerators with a power budget around 10 W. Spartus can complete the inference of an LSTM layer having 5 million parameters within 1 μs.

Statistics

Downloads

333 downloads since deposited on 03 Aug 2022
153 downloads since 12 months
Detailed statistics

Additional indexing

Item Type:Dissertation (monographical)
Referees:Liu Shih-Chii, Delbruck Tobias, Indiveri Giacomo, Han Song
Communities & Collections:07 Faculty of Science > Institute of Neuroinformatics
UZH Dissertations
Dewey Decimal Classification:570 Life sciences; biology
Language:English
Date:2022
Deposited On:03 Aug 2022 11:12
Last Modified:09 Aug 2022 15:10
Number of Pages:213
OA Status:Green
  • Content: Accepted Version
  • Language: English