Details
Presenter(s)
![Yandong Luo Headshot](https://confcats-catavault.s3.amazonaws.com/CATAVault/ieeecass/master/files/styles/cc_user_photo/s3/user-pictures/11871.jpg?h=2c4e73f8&itok=iVg4O5-b)
Display Name
Yandong Luo
- Affiliation
-
AffiliationGeorgia Institute of Technology
- Country
Abstract
Three techniques are proposed to mitigate inference accuracy degradation due to the non-ideal device property of STT-MRAM: 2T-2MTJ bit-cell, redundancy for MSB weights and hybrid-layer mapping scheme. Benchmarking results show that inference accuracy can be maintained at >87.5% for CIFAR10, with minimal 8% energy and 4% chip area overhead.