A large-scale in-memory computing for deep neural network with trained quantization

Cheng，Yuan1; Wang，Chao2; Chen，Hai Bao1; Yu，Hao2

2019

10.1016/j.vlsi.2019.08.004

INTEGRATION-THE VLSI JOURNAL   影响因子和分区

0167-9260

1872-7522

There is a grand challenge to develop energy-efficient yet high throughput accelerator for deep learning. This paper shows an in-memory deep learning accelerator with trained low-bitwidth quantization method. Firstly, we show that a large-scale deep residual network (ResNet) can be trained with 4-bit quantization constraints for both weights and features under high accuracy. Next, we show the quantized ResNet can be mapped to resistive random access memory (ReRAM) with in-memory computing architecture, which can achieve significantly higher parallelism and higher energy-efficient than the state-of-art works. Experiment results using the benchmark of ImageNet show that the training on the ReRAM-crossbar with 4-bit quantization achieves 88.1% accuracy, and only losses 2.5% compared to the full-precision one. Moreover, the proposed accelerator can achieve 432 times speed-up and six-magnitude more energy-efficiency than a CPU-based implementation; 1.30 times speed-up and four-magnitude more energy-efficiency when compared to a GPU-based implementation; and 15.21 times speed-up and 498 times more energy-efficiency than a CMOS-ASIC-based implementation.

Deep neural network Large-scale in-memory computing ReRAM-crossbar Trained quantization

SCI ; CPCI-S ; EI

英语

其他

National Natural Science Foundation of China[61604095] ; Shanghai Jiao Tong University[JCYJ20180504165652917] ; Shanghai Jiao Tong University[]

Computer Science ; Engineering

Computer Science, Hardware & Architecture ; Engineering, Electrical & Electronic

WOS:000500051400031

Elsevier B.V.

20193607403860

Arts computing ; Computer architecture ; Energy efficiency ; RRAM ; Scales (weighing instruments)

Energy Conservation:525.2 ; Data Processing and Image Processing:723.2 ; Special Purpose Instruments:943.3

COMPUTER SCIENCE

2-s2.0-85071689026

Scopus

1.Department of Micro/NanoelectronicsShanghai Jiao Tong University,Shanghai,200240,China
2.Department of Electrical and Electronic EngineeringSouthern University of Science and Technology,Shenzhen,518055,China

Cheng，Yuan,Wang，Chao,Chen，Hai Bao,et al. A large-scale in-memory computing for deep neural network with trained quantization[J]. INTEGRATION-THE VLSI JOURNAL,2019,69:345-355.

Cheng，Yuan,Wang，Chao,Chen，Hai Bao,&Yu，Hao.(2019).A large-scale in-memory computing for deep neural network with trained quantization.INTEGRATION-THE VLSI JOURNAL,69,345-355.

Cheng，Yuan,et al."A large-scale in-memory computing for deep neural network with trained quantization".INTEGRATION-THE VLSI JOURNAL 69(2019):345-355.