tutorial to investigate the xnnpack to save initial time and memory consumptions.

To whom it may concern,

Xnnpack delegate are successfully created on my tflite model deployment and accelerate the inference. However, the initial time and the memory consumption increased (200ms-->700ms, 18MB --> 70MB) after turn on the xnnpack delegate.

Is such increase normal? Could you please provide the tutorial to investigate how to save the initial time and memory consumption?

I'd appreciate any help I can get.
Kind regards,

Li