Last time I dug into this, the conclusion seemed to be that Windows' translation layer, built in cooperation with Qualcom themselves, is actually quite competitive. Performance numbers weren't that far off from Rosetta for desktop applications, with an estimated performance 0-15% lower than Rosetta, which is quite low for a platform without x64 memory model optimisations.

The first release could only do one architecture (32 bit or 64 bit, I'm not sure) but its capabilities have been expanded. A lot of benchmarks are outdated and it seems like nobody has been paying much attention to it in a while now.

The biggest problem with Windows on ARM doesn't seem to he the binary translation, but the slow SoCs the translation is running on.

The most important tools (browsers, Office) work fine on ARM without translation. The translation layer is necessary, but I'm not sure if common use cases will see all that much use for most people.