Loading...

Device Modeling and Design of Bio-Inspired Computing Circuits Based on Spintronics

Alibeigi, Iman | 2024

0 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 57240 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Tabandeh, Mahmoud; Bagheri Shouraki, Saeed; Rajaei, Ramin
  7. Abstract:
  8. Artificial intelligence, as the most important topic of the current decade include various aspects such as artificial neural networks, pattern recognition, machine learning, audio and image processing, and random number generation. To implement hardware hosting artificial intelligence algorithms, a system must be designed with very high processing power, utilizing parallelism and consuming low energy. The best example of such a system is the human brain. For this reason, different structures of bio-inspired or neuromorphic computational systems have been designed taking inspiration from the human brain and are called bio-inspired or neuromorphic computation systems. The limitations of CMOS technology at the nanoscale have made it an unsuitable option for this purpose and a new nanoscale technology called spintronics has been introduced as an alternative solution. In this dissertation, we first review previous hardware structures proposed for implementing bio-inspired systems using spintronics. A new structure for generating random numbers using spintronics technology is introduced, which has lower power consumption than previous designs, and the generated random bit streams are statistically corrected using a post-processing network. We then introduce a structure for implementing synaptic weights with low power by connecting series of spintronic devices, which only reduces the accuracy of hardware implemented networks by 1% to 2% in comparison to software implemented networks. Furthermore, a new spintronic element is introduced and modeled. The implemented memory cell using this element exhibits an error rate of approximately 1.5% against process variations and demonstrates a 36% power reduction in comparison with other similar samples. Finally, we suggest utilizing this element to implement binary neural networks and implement an XNOR network employing the introduced element
  9. Keywords:
  10. Artificial Intelligence ; Artificial Neural Network ; Bioinspired Design ; Neuromorphic Enginnering ; Spintronic Devices ; Parallel Processing ; Binarized Neural Networks

 Digital Object List

 Bookmark