INC Lab’s work with TSMC was published today, led by Sam Liu with Xinyu Bao, William Rogers, Mingyuan Song, and Chen-Yu Hu, under Professor Jean Anne Incorvia. They tested wafer-scale SOT-MRAM and modeled its use in both neural network inference/training and probabilistic graph modeling. The work shows that SOT-MRAM can remain binary while still being highly advantageous for analog crossbar array-based edge AI. While multi-weight is better for accuracy, edge AI is increasingly moving toward quantization for efficiency gains anyways, making SOT-MRAM, as it is today, a great in-memory compute option. Its unique stochastic but binary switching behavior is ideal for both noise resistance during inference and efficient training.
