Rosetta 2: This Time It's Personal (and busting an old Rosetta myth)

Apple's back in the RISC camp, though I still hate the name Apple Silicon, as if Apple has some special sauce for certain inorganic elements that makes it any better than any other kind of silicon. With the release of the M1 ("merely" an A14 on steroids by all accounts) a series of benchmarks have been turning up on Geekbench, which because I'm such a big conspiracy theorist I suspect are probably being astroturfed out of Infinite Loop itself. One that particularly attracted my attention, however, is this one which shows Rosetta 2 (the x86_64-on-AARM emulator analogous to the Rosetta PPC-on-Intel emulator in 10.4-10.6) exceeding the single-core performance of Apple's other Intel machines on Intel apps. The revenge-of-the-G5 Mac Pro is conspicuously absent (for the record a cursory search on the 2019 model yields scores from around 1024 to 1116 depending on configuration), but M1 still eclipses it and even edges past the i9 in the current 27" iMac.

That's pretty stupendous, so I'd like to take a moment to once again destroy my least favourite zombie performance myth, that the original Rosetta was faster at running PowerPC apps than PowerPC Macs. This gets endlessly repeated as justification for the 2005 Intel transition and it's false.

We even have some surviving benchmarks from the time. Bare Feats did a series of comparisons of the Mac Pro 2.66, 3.0 and the Quad G5 running various Adobe pro applications, which at the time were only available as PowerPC and had to run in Rosetta. The Mac Pros were clearly faster at Universal binaries with native Intel code, but not only did the Quad G5 consistently beat the 2.66GHz Mac Pro on the tested PowerPC-only apps, it even got by the 3.0GHz at least once, and another particular shootout was even more lopsided. The situation was only marginally better for the laptop side, where, despite a 20% faster clock speed, the MacBook Pro Core Duo 2.0GHz only beat the last and fastest DLSD G4/1.67GHz in one benchmark (and couldn't beat a 2.0GHz G5 at all). Clock-for-clock, the Power Macs were still overall faster on their own apps than the first Intel Macs and it wasn't until native Intel code was available that the new generation became the obvious winner. There may have been many good reasons for Apple making the jump but this particular reason wasn't one of them.

And this mirrors the situation with early Power Macs during the 68K-PPC transition where the first iterations of the built-in 68K emulator were somewhat underwhelming, especially on the 603 which didn't have enough cache for the task until the 603e. The new Power Macs really kicked butt on native code but it took the combination of beefier chips and a better recompiling 68K emulator to comfortably exceed the '040s in 68K app performance.

If the Rosetta 2 benchmarks for the M1 are to be believed, this would be the first time Apple's new architecture indisputably exceeded its old one even on the old architecture's own turf. I don't know if that's enough to make me buy one given Apple's continued lockdown (cough) trajectory, but it's enough to at least make me watch the M1's progress closely.