Symmetric 2-D-Memory Access to Multidimensional Data

• Published in: IEEE transactions on very large scale integration (VLSI) systems 2018-06, Vol.26 (6), p.1040-1050
• Main Authors: George, Sumitha, Li, Xueqing, Liao, Minli Julie, Ma, Kaisheng, Srinivasa, Srivatsa, Mohan, Karthik, Aziz, Ahmedullah, Sampson, John, Gupta, Sumeet Kumar, Narayanan, Vijaykrishnan
• Format: Article
• Language: English
• Subjects: 2-D memories; Arrays; big data analytics; Data transfer; ferroelectric FET (FEFET); Nonvolatile memory; nonvolatile memory (NVM); nonvolatility; Organizations; Random access memory; Transistors
• ISSN: 1063-8210; 1557-9999
• DOI: 10.1109/TVLSI.2018.2801302

In this paper, we propose a novel memory architecture with the capability of single-cycle row-wise/column-wise accesses. Such an architecture is highly suitable for workloads featuring spatial locality in multiple dimensions, which is a characteristic of many matrix and array operations. We describe in detail the circuit design techniques enabling the proposed architectures, as well as the viability of emerging memory technologies based on ferroelectric transistors (FEFETs) for our design. Compared to FEFET memory with standard 1-D access, we achieve 5% energy savings for the proposed memory featuring 2-D read and 93% energy savings for memory with 2-D read and write, for 32 bit column read and write. In addition, we get around 11% and 95% delay savings for 2-D read-enabled memory and 2-D read-write memory, respectively. The application analysis shows that 2-D read-enabled memory achieves around 86% average decrease in row-buffer transactions in 256\times 256 size matrix operations without any array area increase. The 2-D read write memory offers 87% decrease in row-buffer transactions with 28.5% increase in array area compared to the 1-D FEFET memory.