Name of the Speaker: Sourodeep Roy (EE19D414)
Name of the Guide: Dr. Enakshi Bhattacharya and Dr. Bhaswar Chakrabartoi
Date/Time: July 27th, 12:00 to 1:00 PM
Artificial neural networks (Deep Learning) have shown excellent accuracy in image classification and recognition tasks. For performing these tasks accurately, efficient training of the network is essential. CMOS-based architectures have been commonly used as accelerators for training the networks. But in this case, the computing unit requires data from the off-chip memory, which leads to time delay and energy consumption. In-memory computing with non-volatile memory can significantly improve energy efficiency and latency due to the ability for one-step vector-matrix multiplication and local storage of network parameters. Resistive Random Access Memory (RRAM) is a viable candidate among emerging non-volatile memory technologies as it provides high device density, multi-bit storage, and low latency in write/read operations. The ability of conductance modulation in RRAM devices is analogous to that of a biological synapse. The first part of this talk discusses the need for in-memory computing. We then discuss different emerging non-volatile memory technologies that are at our disposal. We then discuss the operating principles of RRAM, the advantages of RRAM technology, and the technological challenges for practical applications.
Various high-k oxides have been previously used as an insulator for the RRAM technology. However, Silicon Oxide (SiOx) can still emerge as a great candidate due to the advantages of availability, low cost, and excellent CMOS compatibility. In the final part of this talk, we discuss the development of a SiOx RRAM technology. We use Inductively Coupled- Plasma Enhanced Chemical Vapor Deposition (ICP-PECVD) to develop SiOx as the active switching layer. We design and fabricate Au/SiOx/Ti/Au RRAM cross-point devices using photolithography and electron-beam evaporation. The devices exhibit stable, bipolar operation after an initial electroforming step. Detailed electrical characterizations are performed to evaluate the performance in terms of operating voltages, dynamic range, endurance, retention, and synaptic weight updates. We vary the thicknesses of the SiOx switching layer as well as the thickness of the Ti top electrode to optimize the RRAM performance. Detailed investigations have also been carried out to understand the resistance switching mechanism.