Muhammad Husnain Mubarik

From Bit-Position Sensitivity to Unequal Error Protection for DNN Inference Memory

Muhammad Husnain Mubarik, Karthik Mohan Kumar, Pedro Antonio Pena, Keshavan Varadarajan, Kunal Tyagi

We characterize per-bit-position fault sensitivity in ML inference across 16 workloads -- spanning transformer-based models and attention-free CNNs -- and across three floating-point formats. Our central empirical finding is a sharp bit-sensitivity transition: flipping any of the…Read more