HR3AM: A Heat Resilient Design for RRAM-based Neuromorphic Computing

Xiao Liu, Minxuan Zhou, Tajana Rosing, Jishen Zhao

Abstract

RRAM based accelerators have been widely adopted in many neuromorphic designs. However, RRAM cells are sensitive to temperature, which changes RRAM’s conductance. Such heat-induced interference can significantly decrease the computational accuracy because values are functions of RRAM conductance. In this paper, we propose HR <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> AM, a heat resilience design, which improves accuracy and optimizes the thermal distribution of RRAM based neural network accelerators. HR <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> AM consists of two key mechanisms: bitwidth downgrading and tile pairing. Bitwidth downgrading re-represents weights by shifting the conductance to improve the network inference accuracy. Tile pairing matches hot crossbar units with pre-defined idle units to mitigate high-temperature issues. We evaluated HR <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> AM on four real world neural network models. Results show that HR <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> AM improves classification accuracy by up to 41.8% compared with current state-of-the-art designs. For thermal optimization, HR <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> AM effectively decreases the maximum temperature by 6.2K and average temperature by 6K.
BibTeX
@inproceedings{liu2019hr,
  author = {Xiao Liu and Minxuan Zhou and Tajana Rosing and Jishen Zhao},
  title = {{HR<sup>3</sup>AM: A Heat Resilient Design for RRAM-based Neuromorphic Computing}},
  booktitle = {ISLPED},
  year = {2019},
  doi = {10.1109/islped.2019.8824926},
}

Loading related content...

Figure