Efficient Two-Dimensional Accessible Memory

University of Pittsburgh researchers are developing a two-dimensional accessible memory architecture. Based on non-volatile devices, this novel system required no additional control devices or routing wires. This RC-NVM (Row-Column-Non-Volatile Memory) based technology, is a dual-addressable memory designed to support both row- and column- oriented access and can improve memory access performance over 14.5 times with only 15% area overhead. Full development and optimization of this novel memory architecture could revolutionize many industries including the fields of data analysis, finance, and artificial intelligence. 

Description

In-memory databases (IMDBs) are an ever-growing area of interest. IMDB, where data are stored in the computer RAM rather than on a disk, promises faster data access and processing speeds, allowing for real-time transactions and analysis on the same database. As process technology scales down, conventional memory solutions are no longer suitable. There is a need to develop better memory systems to allow faster access and analysis of databases without the dramatic increases in area overhead, semiconductor needs, or power consumption. This novel two-dimensional approach uses crossbar-based non-volatile memory (NVM) technology to improve the efficiency of IMDBs with only moderate increase in area overhead compared to conventional approaches.

Applications

• Artificial intelligence
• Financial trading
• Database management and analysis
• Security systems, including counter terrorism

Advantages

Current IMDBs operate using dynamic random-access memory (DRAM) and workloads are categorized into on-line transactional processing (OLTP) and on-line analytical processing (OLAP). OLTP workloads are row-based, while OLAP workloads are column-based, often requiring two copies of data to reside in the RAM, using up memory resources. While efforts have been made to combine workloads into a single database, this has led to poor DRAM row-buffer and cache utilization, and degraded memory efficiency. As databases grow the efficiency is decreased further.

RC-NVM overcomes these shortcomings by applying crossbar architecture. Word lines and bit lines are identical, and rows and columns can be accessed in a single database. Unlike existing semiconductors used in DRAM which are capacitor-switch models, NVM structures are modelled as a resistor switch providing fast access speed to databases, improving performance.

Invention Readiness

A row-column-accessible RAM using NVM technology has been produced and made applicable for IMDBs. Experiments confirmed this approach could improve the memory access by up to 14.5 times with only 15% overhead. Work is now required to optimize this approach to improve efficiencies in IMDBs.

IP Status

https://patents.google.com/patent/US10762957B2

Related Publication(s)

Wang, P., Li, S., Sun, G., Wang, X., Chen, Y., Li, H., Cong, J., Xiao, N., & Zhang, T. (2018). RC-NVM: Enabling Symmetric Row and Column Memory Accesses for In-memory Databases. In 2018 IEEE International Symposium on High Performance Computer Architecture (HPCA) (pp. 518–530). 2018 IEEE International Symposium on High Performance Computer Architecture (HPCA). IEEE. https://doi.org/10.1109/hpca.2018.00051