| Title |
A Comprehensive Analysis of STT-MRAM-based Bayesian Neural Network Implementation |
| Authors |
안다솜(Dasom Ahn) ; 이시열(Siyeal Lee) ; 나태희(Taehui Na) |
| DOI |
https://doi.org/10.5573/ieie.2025.62.12.46 |
| Keywords |
Bayesian neural networks; STT-MRAM |
| Abstract |
Bayesian Neural Networks (BayNNs) are probabilistic models capable of quantifying predictive uncertainty, making them a critical technology for building highly reliable artificial intelligence systems. Recently, hardware implementations of BayNNs have gained significant attention, with Spin-Transfer-Torque Magnetic Random Access Memory (STT-MRAM)-based Near-Memory Computing (NMC) architectures emerging as promising candidates due to their high energy efficiency, low latency, and suitability for repeated inference. This paper analyzes how uncertainty metrics in Monte-Carlo dropout-based BayNNs vary with respect to the dropout probability and number of inference iterations. Furthermore, it explores key implementation challenges in STT-MRAM-based NMC architectures?such as repeated weight access and Random Number Generator (RNG) instability caused by Magnetic Tunnel Junction (MTJ) variation?and proposes circuit- and architecture-level solutions to mitigate these issues. |