hello LoveNvidia,
besides TRM, you may also check NX SOM Data Sheet for some information.
you might have a try to execute TensorRT to use the DLA engines, by adding useDLACore options into the pipeline; please also check Topic 67031 for reference,
thanks