With Apple and Google doing custom chips, I don't think that matters as much as people think anymore. Surely, a lot of functionality that used to be done in software is now done on chip. A generic benchmark isn't going to capture that.
I want to be pragmatic here. Coprocessor were all the rage for the Amiga, then were out out of the loop, and were back for with the GPU. A bit like we're back to the mainframe experience with Kubernetes and all (don't throw tomatoes at me ;))