计算机硬件
算术
宏
计算机科学
数字信号处理
并行计算
数学
程序设计语言
作者
Chuan-Tung Lin,Dewei Wang,Bo Zhang,Gregory K. Chen,Phil Knag,Ram Krishnamurthy,Mingoo Seok
出处
期刊:IEEE Journal of Solid-state Circuits
[Institute of Electrical and Electronics Engineers]
日期:2023-01-01
卷期号:: 1-12
标识
DOI:10.1109/jssc.2023.3313519
摘要
Recent SRAM-based in-memory computing (IMC) hardware demonstrates high energy efficiency and throughput for matrix–vector multiplication (MVM), the dominant kernel for deep neural networks (DNNs). Earlier IMC macros have employed analog-mixed-signal (AMS) arithmetic hardware. However, those so-called AIMCs suffer from process, voltage, and temperature (PVT) variations. Digital IMC (DIMC) macros, on the other hand, exhibit better robustness against PVT variations, but they tend to require more silicon area. This article proposes novel DIMC hardware featuring approximate arithmetic () to improve area efficiency without hurting compute density (CD). We also propose an approximation-aware training model and a customized number format to compensate for the accuracy degradation caused by the approximation hardware. We prototyped the test chip in 28-nm CMOS. It contains two versions: the with single-approximate hardware () and with double-approximate hardware (). The measurement results show that supports a 4 b-activation and 1 b-weight (4 b/1 b) CNN model, achieving 327 kb/mm $^2$ , 458–990 TOPS/W (normalized to 1 b/1 b), 8.27–392 TOPS/mm $^2$ (normalized to 1 b/1 b), and 90.41% accuracy for CIFAR-10. supports a 1 b/1 b CNN model, achieving 485 kb/mm $^2$ , 932–2219 TOPS/W, 14.4–607 TOPS/mm $^2$ , and 86.96% accuracy for CIFAR-10.
科研通智能强力驱动
Strongly Powered by AbleSci AI