CrossOver ARM64 and FEX: Mac Gaming After Rosetta
August 2026 · Mac Gaming Benchmarks Subscribe on YouTube — full runs, every machine tested.The Short Version
Apple has said Rosetta 2 will be largely gone with macOS 28, kept only for “certain older, unmaintained games that rely on Intel-based frameworks.” Rosetta is what lets Intel code run on Apple silicon, and it is underneath almost every way you currently play Windows games on a Mac.
This week CodeWeavers shipped their answer: a CrossOver Preview build compiled natively for ARM64, using their own open-source x86 emulator, FEX, instead of Apple's translator. I installed it on an M5 MacBook Air and measured what it costs, with power and temperature logged once a second on every run. CrossOver has a free trial, and our CrossOver install guide covers first-time setup if you're starting from scratch.
The headline: FEX matched or beat the Rosetta-based path in everything I tested, and in Dota 2 it also beat the native Mac build. Windows games in a bottle survive Rosetta's removal. The thing actually at risk is your “native” Mac library.
Most “Native” Mac Games Are Not Native
One command settles it. lipo -archs reports which CPU architectures a binary was built for. Run it against Mac games you already own:
lipo -archs — installed Mac game binaries
“Native Mac version” and “built for Apple silicon” are two different things, and the search results conflate them constantly. Valve's own Mac build of Dota 2 is Intel code — it runs on Rosetta today, exactly like the Windows version does in a bottle. Cyberpunk 2077 is the control: arm64 only, a real Apple silicon port.
So the macOS 28 deadline is not a CrossOver story. Unless a studio recompiles, a chunk of the native Mac catalogue is on the same countdown.
Dota 2: Three Ways to Run It
Dota 2 is the only game that reaches all three paths, which makes it the cleanest test available. The native Mac build runs through Rosetta and MoltenVK. The Intel bottle runs Wine plus Rosetta plus DXMT. The ARM64 bottle runs Wine plus FEX plus DXMT. Every run used the same replay over the same fixed tick range, so all three rendered identical work.
Dota 2 — same replay, ticks 29000–31801
Average FPS. The ARM64 run started 16°C hotter than the native run and still won, so that comparison is biased against FEX and survives anyway.
The averages undersell it. The interesting number is the tail:
Dota 2 — 1% lows, and frames over 100ms
Across 1,905 frames each: the ARM64 bottle dropped 5 frames over 100ms, the native Mac build dropped 20, and the Intel bottle dropped 21 with a worst frame of 967ms. The emulated Windows game runs smoother than Valve's own Mac build.
The honest caveat: the gap between the two bottles is not a clean measurement. The Rosetta run started 9°C hotter than the ARM one, and as the next section shows, that matters enormously on this machine. Treat FEX as at least as fast, not twice as fast, until I re-run Intel from cold.
What I Got Wrong About Throttling
I have said on this channel that the MacBook Air doesn't really throttle — that Apple handles heat better than people give it credit for. That was wrong, and the data is unambiguous.
Same benchmark. Same settings, verified line by line in the config file. Same laptop, twelve minutes apart. The only difference was how warm it was when I pressed start.
FFXIV Dawntrail — identical settings, twelve minutes apart
A 73% swing from a 12.6°C difference in starting temperature. The hot run also dipped below 30 FPS twelve times and bottomed out at 6; the cold run never went below 34. Total system power fell from about 34W to 19W across five minutes — the M5 Air is fanless, so this is physics, not a defect.
Why I got it wrong: I was running short tests. Twenty seconds, sixty seconds — all of that lives inside the boost window, before the machine has to be honest. Every number in this article is a cold start with temperature logged.
Cyberpunk 2077: What Translation Actually Costs
Cyberpunk cannot test FEX at all — it is Direct3D 12, and the ARM64 build has no D3D12 path. But it is the only game here with a genuine Apple silicon port, so it answers a different question: what does the Windows layer cost?
Cyberpunk 2077 — 2560×1600, High preset, mains power
Effectively a tie — and the bottle started 21°C hotter. Median frame time was identical at 49.96ms versus 49.88ms. But the bottle spent 2.2× the CPU (36.3% vs 16.6%) and 16% more power (18.4W vs 15.8W) to land on the same number, and it stuttered where native didn't.
That is the translation tax, and it doesn't show up in frames per second at all. It shows up in watts — and on a fanless laptop, watts become heat, and heat becomes the throttling chart above.
The Benchmark You Can Run Yourself
The FINAL FANTASY XIV: Dawntrail benchmark is a free 4GB download — no game, no subscription. Because it is a portable folder, I dropped the identical copy into both bottles, so the workload is the same bytes on both sides. It also writes its own score file, which means no screen recording and no transcription errors.
FFXIV Dawntrail — 1920×1080, High, DX11, both from a cold start
Call it a tie — 5.5% is inside run-to-run variance, and the Rosetta run started 6°C cooler, so that reading is the friendly one for Apple. Which is the real headline: something CodeWeavers shipped this week matches something Apple has shipped for five years. Both runs used the benchmark's own defaults, including FSR upscaling, with config files verified identical.
If you have an Apple silicon Mac, run it cold, play something for ten minutes, then run it again and compare. I would genuinely like to see how big the drop is on the machines that have a fan.
What Still Doesn't Work
CodeWeavers are blunt about this themselves: the build is “not expected to be useful for anything beyond testing at this point.” What I hit:
- No Direct3D 12 in the ARM64 build — so no Cyberpunk, and no modern D3D12 titles.
- No D3DMetal in the ARM64 build either. It is DXMT only, so Apple's Game Porting Toolkit does nothing for it.
- Tomb Raider crashed in every bottle after I switched the Intel bottle's backend to D3DMetal. The telemetry shows the GPU never engaged — 19.8% active, 470MHz — while the CPU loaded normally.
- But Steam launches, which CodeWeavers warned often doesn't happen with game launchers. That is the exception, not the rule — a commenter reports Battle.net launching while Diablo IV doesn't.
How These Numbers Were Measured
macOS has no HWiNFO equivalent, so every run was logged once a second by a script wrapping macmon: total system power, CPU and GPU temperature, P-core and E-core clocks, GPU utilisation, RAM and swap. Scores come from the games themselves rather than a counter on screen — Cyberpunk and FFXIV write their own result files, and Dota 2's timedemo was bounded to a fixed tick range so every stack rendered identical work.
Machine: M5 MacBook Air, 10-core CPU and 10-core GPU, 32GB, macOS 27.0 (26A5388g). CrossOver Preview ARM64, Windows 10 on ARM bottles, DXMT for Direct3D 11 in both. Telemetry was trimmed to the benchmark window rather than the whole capture, because averaging in menu time quietly moved several results before I caught it.
The limits of this test: one machine, one afternoon, on a first build. Start temperature moved results more than any software choice did, so any comparison where the hotter machine lost is flagged above as needing a re-run rather than presented as a finding.
CrossOver or Parallels? (And Where Whisky Fits)
This article is about one tool, but it is the wrong tool for some people, so here is the honest split. CrossOver is what everything above was measured on: no Windows licence, no virtual machine, one app per game. Parallels is a full Windows-on-ARM virtual machine — more overhead and a Windows licence to buy, but you get an actual Windows desktop, which matters if you need Windows for work as well as games. Whisky is the free option people ask about most, and it is no longer maintained, which is the whole reason it doesn't appear in any test here.
There's a neat irony worth noticing: Parallels runs Windows on ARM, so your x86 game is being emulated there too — just by Microsoft's translator instead of Apple's or CodeWeavers'. Every path on this page is emulation. The only question is whose.
I broke the three down properly, including what each one actually costs, in Whisky vs CrossOver vs Parallels. If you just want the shortest route to playing a Windows game on a Mac today, start with the CrossOver free trial and confirm your specific game runs before paying anything — that advice goes double for the ARM64 Preview build, which is explicitly not finished.
So Is Mac Gaming Over?
No. But the thing that saves it isn't Apple's — it is a bundled open-source emulator from a company in Minnesota, and today it is rough. It has no DirectX 12, no D3DMetal, and plenty of things simply don't launch.
It also already matches Apple's own translator, and in one game it beat it outright, while running smoother than that game's official Mac port. That is week one.
The bigger surprise is how much of the “native” Mac library is sitting on the same countdown. Run lipo -archs on your own games and see what comes back.
Related Reading
- Whisky vs CrossOver vs Parallels — which compatibility layer to actually buy, and why the free one is a dead end.
- How to install CrossOver on a Mac — first-time setup, end to end.
- Cyberpunk 2077 on Mac — the one game here with a real Apple silicon port, tested both ways.
- Dota 2 on Mac — more on the game that runs all three paths.
- MacBook gaming benchmarks by model — every machine tested, in one table.