There is more substantial overhead translating to a different instruction set than in converting API calls to another API. Looking at basic benchmarks should be enough to demonstrate this. IIRC it's like 80% of the performance or something when using Rosetta as compared to native.

Also, Rosetta is more like a transpiler, since it basically recompiles the binary, whereas the others are literally layers that basically take calls in one API and translate them to another. They're pretty much the same thing as ANGLE.

> There is more substantial overhead translating to a different instruction set than in converting API calls to another API.

It's a one time only cost, since Rosetta only runs the first time you launch an app and the translation is written to disk to be used in the future.

That means there is no ongoing cost by your logic, and Proton translating an API call every time it is used is worse.