Accurate and Energy-Efficient Bit-Slicing for RRAM-Based Neural Networks

作     者：Diware, Sumit Singh, Abhairaj Gebregiorgis, Anteneh Joshi, Rajiv V. Hamdioui, Said Bishnoi, Rajendra 

作者机构：Delft Univ Technol Dept Comp Engn NL-2628 CD Delft Netherlands IBM Thomas J Watson Res Ctr Yorktown Hts NY 10598 USA 

出 版 物：《IEEE TRANSACTIONS ON EMERGING TOPICS IN COMPUTATIONAL INTELLIGENCE》 (IEEE Trans. Emerging Topics Comp. Intell.)

年 卷 期：2023年第7卷第1期

页      面：164-177页

核心收录：

基　　金：ECSEL Joint Undertaking (JU) - EU H2020 

主　　题：Neural networks Sensors Arithmetic Voltage Virtual machine monitors Degradation Common Information Model (computing) Computation-in-memory bit-slicing neural networks non-zero G(min) error conductance variation non-idealities 

摘      要：Computation-in-memory (CIM) paradigm leverages emerging memory technologies such as resistive random access memories (RRAMs) to process the data within the memory itself. This alleviates the memory-processor bottleneck resulting in much higher hardware efficiency compared to von-Neumann architecture-based conventional hardware. Hence, CIM becomes an attractive alternative for applications like neural networks which require a huge number of data transfer operations in conventional hardware. CIM-based neural networks typically employ bit-slicing scheme which represents a single neural weight using multiple RRAM devices (called slices) to meet the high bit-precision demand. However, such neural networks suffer from significant accuracy degradation due to non-zero G(min) error where a zero weight in the neural network is represented by an RRAM device with a non-zero conductance. This paper proposes an unbalanced bit-slicing scheme to mitigate the impact of non-zero G(min) error. It achieves this by allocating appropriate sensing margins for different slices based on their binary positions. It also tunes the sensing margins to meet the demands of either high accuracy or energy-efficiency. The sensing margin allocation is supported by 2 s complement arithmetic which further reduces the influence of non-zero G(min) error. Simulation results show that our proposed scheme achieves up to 7.3x accuracy and up to 7.8x correct operations per unit energy consumption compared to state-of-the-art.