Quartet: A 22nm 0.09mJ/inference digital compute-in-memory versatile AI accelerator with heterogeneous tensor engines and off-chip-less dataflow