Hector Martin

As for Rosetta 2, it's good, but I'm still *really* curious how it'll do in the audio domain. We're talking lots of floating point processing with some integer mixed in, written by lots of different teams, some scalar, some vector, *definitely* a lot of it not well optimized.

And with hard realtime constraints - if the JIT fires off anything substantial in the audio processing thread, you *will* get a dropout - and even if it's not substantial, you'll probably get a pile of priority inversion hazards that will cause inconsistent dropouts.

So it looks like for day-to-day stuff Mac users can probably be confident that they won't lose much vs. their older Intel Mac under Rosetta 2, and gain in many instances. But I wouldn't put my money on M1+R2 for all workloads yet.

It'll be interesting to see these performance details worked out in more detail; e.g. people have talked about M1 being way faster at ObjC object management, so presumably it has *way* faster atomics. That matters a lot for some kinds of software, and not at all for others.

But the question is how, and why - presumably their bus system is tighter than typical x86 ones? I'm looking forward to a deeper dive, and whether AMD/Intel care to improve this in the future.

Also, remember that Apple cheated with their control over the CPU for Rosetta 2. Getting R2 x86 performance on any other ARM is impossible, due to the memory model mismatch. You have to massively slow down all loads and stores.

So Apple straight up implemented the x86 consistency model on their cores. That's the kind of high-impact detail that makes or breaks emulation performance for a different arch. Did they do this for any other x86-isms? Nobody knows so far.

n.b. the x86 consistency model is a part of the architecture since... one of the 8.3 extensions, that's implemented in A55/75 and newer. Of course, only with specific instructions, but they're not inherently slower except that you lose addressing mode flexibility