Emulation based Power and Performance Workloads on ML NPUs