Abstract: Performing AI inference ubiquitously requires energy-efficient, small footprint and highly reliable processing devices. Heterogeneous processing architectures combining customized CPUs with ...