TOC
Krishnendu Chakrabarty
生駒 : 奈良先端科学技術大学院大学, 2018.12
Lecture ArchiveNo. | Printing year | Location | Call Number | Material ID | Circulation class | Status | Waiting |
---|---|---|---|---|---|---|---|
1 |
|
|
M016011 |
|
|
|
The performance of today’s computing infrastructure is limited by the energy consumption involved in processing, storing, and moving data. In addition, the exponential increase in the volume of data that must be handled by our computational infrastructure is driven in large part by machine-learning applications such as deep neural networks. Conventional computing architectures are unable to deal efficiently with this data volume or with the requirement to transform data into actionable information. Moreover, a major showstopper towards energy efficient computing is the high error rate associated with traditional techniques. There is a need to incorporate fault tolerance in emerging computing architectures and circuit designs. This talk will first introduce the audience to the exciting and emerging area of brain-inspired neuromorphic computing systems for machine-learning hardware. First, the presenter will describe RRAM-based crossbars and their role in neuromorphic computing systems. Following this, the need for testing and fault tolerance will be motivated in light of imperfect fabrication technologies, as well as technology limitations such as write endurance in RRAM cells. The speaker will present a physics-based classification and analysis of memristor fault origins. These faults origins will be systematically attributed to process variations and manufacturing defects. This study of memristor fault origins and the resulting conclusions provides valuable feedback for the fabrication and the design of memristor-based circuits and systems. Fault models and test solutions will be presented. Subsequently, techniques for online testing and fault-tolerant training will be described. Finally, time permitting, the speaker will describe an efficient fault tolerance method inspired by algorithm-based fault tolerance (ABFT), and referred to as extended-ABFT (X-ABFT), for fault detection and error correction.
2018
電子化映像資料(1時間29分24秒)
情報科学領域・コロキアム ; 平成30年度
講演日: 平成30年12月4日
講演場所: 情報科学棟大講義室(L1)
Japan
English (eng)
English (eng)