SHF: Small: Reliable Storage and Computation in Memory Technologies

University Of Texas At Austin, Austin TX

Investigators

Abstract

Memory circuits are a core building blocks for electronic systems and often are a bottleneck limiting system performance. A number of promising new memory technologies have emerged, including phase-chase memory, spin transfer torque memory, and resistive random access memory (RRAM), which offer significant improvements over conventional memory technologies in terms of speed, density, power consumption, and scalability. An exciting application for RRAM is its ability to perform matrix multiplication through current summation which significantly speeds up matrix-vector multiplication and thus making it ideal for implementing neural networks. A major challenge preventing widespread adoption of these emerging memory technologies is that they have a number of reliability issues that can cause errors during operation and shorten their useable lifetime. The goal of this project is to develop transformative approaches for improving the reliability of these emerging memory technologies to address these challenges. The educational impact of this project will include training of researchers in high impact areas, dissemination of new educational materials, and providing research opportunities for undergraduates and underrepresented groups. Error correcting codes (ECC) are widely using for conventional memories including in caches and main memory as well as secondary storage to protect against transient and permanent errors. However, direct application of the conventionally used ECC codes for emerging memory technologies in these applications is not feasible for several reasons, including significantly higher error rates, the need for faster decoding, and efficient application in multilevel memories. This project will investigate new ECC codes that can handle high error rates and are efficient for multilevel memories while providing high-speed decoding. RRAM is prone to a number of failure mechanisms that can result in either hard or soft errors that can affect the accuracy of matrix multiply computations. New methodologies to ensure reliable computation in RRAM will be investigated including both application-independent as well as application-dependent schemes that can use functional properties to reduce overheating. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.