Enabling Logic-Memory Synergy Using Integrated Non-Volatile Transistor Technologies for Energy-efficient Computing

Sandeep Krishna Thirumala, Purdue University

Abstract

Over the last decade, there has been an immense interest in the quest for emerging memory technologies which possess distinct advantages over the traditional silicon-based memories. The unique opportunities ushered by these technologies such as high integration density, near-zero leakage, non-volatility and, in some cases, excellent CMOS compatibility, has triggered the development of design techniques, enhancing the computation capabilities of various systems. Further, in the era of big data, the emerging memory technologies offer new design opportunities to address a pressing need of achieving close integration of logic and memory sub-systems with an objective to overcome the von-Neumann bottleneck associated with the humungous cost of data transfer between logic and memory. Such a logic-memory coupling not only enables low power computation in standard systems, but also promises high energy efficiency in unconventional compute platforms such as the brain-inspired deep neural networks (DNNs) which have transformed the field of machine learning (ML) in recent years. However, in order to exploit the unique properties of the emerging memory technologies for efficient logic-memory integration, there exists a strong need to explore cross-layer design solutions which can potentially enable efficient computation for current and future generation of systems. Motivated by this, in this dissertation, we harness the benefits offered by the emerging technologies and propose novel devices and circuits which exhibit an amalgamation of logic and memory functionalities. We propose two variants of memory devices: (a) Reconfigurable Ferroelectric transistors (R-FEFET) and (b) Valley-Coupled-Spin Hall (VSH) effect based magnetic random-access memory (VSHMRAM), which exhibit unique logic-memory unification. Exploiting the intriguing features of the proposed devices, we carry out a cross-layer exploration from device-to-circuits-to-systems for energy efficient computing. We investigate a wide spectrum of applications for the proposed devices including embedded memories, non-volatile logic, compute-in-memory circuits and artificial intelligence (AI) systems. The first technology of our focus is ferroelectric transistor (FEFET), which is being actively explored for logic and memory applications. Experimental studies have showcased volatile (logic) or non-volatile (memory) characteristics for FEFET by employing static/design time optimizations. However, if run-time tuning of non-volatile and volatile modes can be achieved, several new avenues for circuit design will open. Inspired by this, we propose Reconfigurable FEFET (RFEFET), which has the capability to dynamically modulate its operation between volatile and nonvolatile modes, enabling true logic-memory synergy at the device level. Utilizing these unique features of the R-FEFET, we propose an embedded non-volatile flip-flop design (R-NVFF) featuring a fully automatic backup operation (during power shut down) without the need of any external circuitry or signals. Compared to a previously proposed FEFET based NVFF, the proposed R-NVFF exhibits 69% lower check-pointing energy (which includes backup and restore operation). We also propose non-volatile memory (NVM) with highly energy-efficient read and write operations enabled by the dynamic reconfigurability in R-FEFETs. Our proposed NVM exhibits 55% lower write power, 37%-72% lower read power and 33% lower area compared to an FEFET-NVM. Finally, we implement the proposed NVM and R-NVFF in a state-of-the-art intermittently-powered platform and show up to 40% energy savings at the system-level. Another technology, which has sparked immense interest in spintronic applications, is the Valley-coupled-Spin Hall (VSH) effect in two-dimensional Transition Metal Dichalcogenides (2D TMDs). The unique generation of out-of-plane spin currents in monolayer TMDs can potentially enable efficient switching of nano-magnets. In this dissertation, we propose an emerging spin-based memory device featuring close logic-memory integration utilizing the VSH effect in 2D TMD transistors, where the information is stored in nano-magnets (which are unified with the transistor), for energy efficient computing.

Degree

Ph.D.

Advisors

Gupta, Purdue University.

Subject Area

Artificial intelligence|Energy

COinS