Session

DescriptionMultimodal transformers get real-time performance boost using PIM-GPU collaboration. What's more, DNN inference tasks can be modeled as sporadic non-preemptive gang tasks. For parallel real-time applications on SoCs, an extra cache layer shared by a cluster improves latency. To protect CNNs, algorithm-based comprehensive fault tolerance can be used to "maintain sanity". Analysis of NN layers reveals that faults transitioning from 0 to 1 significantly impact classification outcomes. From the networking side, a real-time TSN end station design enables ultra-low latency and nanosecond-level transmission accuracy.