You may be confusing that with requirements for nested virtualization. That does indeed require an M3 (not M4, no macs have that yet :-( ), even though the hardware feature is in M2. Most likely, it's slightly broken in M2.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
M3 vs M4
- Thread starter Dulcimer
- Start date
-
- Tags
- m3 chip m4 chip neural engine
- Sort by reaction score
Rosetta 2 did not use to support AVX 2. [EDIT 3: Ah, I think I misunderstood your post to mean it always had supported it - i.e. prior to GPTK2. It does support it now and it does seem to be across every M* Mac.
About the Rosetta Translation Environment | Apple Developer Documentation
Learn how Rosetta translates executables, and understand what Rosetta can’t translate.
The above webpage as of writing for instance has not been updated to include AVX/2 support. Sorry for the confusion. ]
AVX 2 is new to GPTK 2.0 and presumably works through Rosetta 2 and indeed it doesn't mention a requirement for M4. It should be easily testable by developers with Sonoma beta.
https://developer.apple.com/games/game-porting-toolkit/
EDIT: While the above webpage talks about bringing your game to both Mac and iOS/iPadOS I don't think iPadOS actually supports running Rosetta 2. So unless that's changed, it would be extra strange to have an M4 requirement as no one would be able to use it at the moment. The WWDC24 developer video shows the emulation environment running on a Mac with no mention of iPad:
Port advanced games to Apple platforms - WWDC24 - Videos - Apple Developer
Discover how simple it can be to reach players on Apple platforms worldwide. We'll show you how to evaluate your Windows executable on...
So no I think this rumor about M4 only AVX2 support is wrong. Again, should be easily testable. @Xiao_Xi where did this rumor come from?
EDIT2: Ah it seems the beta AVX2 support is not fully working yet however CPUID claims it will work on M1s so it should be across the entire lineup once it is fully enabled.
Longhorn (@never_released@mastodon.social)
Kernel not aware yet that Rosetta supports AVX2 now, so no x86_64h binaries running or it being exposed through sysctl yet. You can use cpuid for probing its support rn.
Longhorn (@never_released@mastodon.social)
@chris@social.losno.co all, AVX2 works fine in Rosetta 2 on M1.
I think Hector got nested virtualization working in the M2 on Linux. This may be more like how Apple released clamshell support for more monitors for the M3 Air but the M2 could've done so as well (my fuzzy memory is that I think they changed the display controller since the M1 so it may not be able to).
EDIT: they [Asahi Linux team] don't quite have nested virtualization out yet [for Linux], but it was indeed developed and tested on the M2:
Hector Martin (@marcan@treehouse.systems)
@cinebox@hackers.town No, they beat us at that but barely... but only on M3 apparently? Linux NV support isn't fully upstream yet but it's been *developed and tested* on M2 so I have no idea what's up with that.
Last edited:
Doubt it due to die space, being able to have dual displays natively on the Air is already a good step in the right direction.
But I don’t see them pushing this any forward as the Pro/Max can do that.
Only reason they changed on the Air was due to the enterprise pressure
It appears that Apple has enabled nested virtualization for M3 and later.
isNestedVirtualizationSupported | Apple Developer Documentation
A Boolean value that describes whether the platform configuration supports nested virtualization.
As usual Vadim appears to be wrong about technical matters. Not sure where he got that from. Also ARMv9 AVX2 support what's he on about? Is he confusing AVX2 with SVE2? Either way it doesn't matter since the M4 doesn't have SVE2 either. Its vector processors are still NEONs. Maybe he thinks SSVE2/SME2 is how Apple is emulating AVX2? That's probably giving him too much credit and while I don't know for sure, I also kinda doubt that's how Apple is emulating AVX2, but maybe? I would've thought it was being done through the NEONs even though they are 128b. But regardless, AVX2 support seems to be functional for the M1, M2, and M3 (which even if it is the matrix accelerator being used to emulated AVX2, again seems unlikely to me, the M1-3 have the matrix accelerator as well - they're just using Apple's private extensions rather than ARM's official matrix extensions).
Bottom line: there's no way to make sense of what he's saying here. Which, as I opened with, is pretty much expected for MaxTech on anything technical.
It appears that Apple has enabled nested virtualization for M3 and later.
isNestedVirtualizationSupported | Apple Developer Documentation
A Boolean value that describes whether the platform configuration supports nested virtualization.developer.apple.com
Sorry my "they" was unclear. I was referring to the Asahi Linux team, not Apple. Asahi Linux don't quite have nested virtualization upstreamed to Linux yet but when they [Asahi] do, it'll be for the M2. So it is in theory possible for Apple to support nested virtualization on the M2 too.
I've re-re-re-edited my previous post for clarity.
Last edited:
Ratchet & Clank: Rift Apart is running on an M1 chip here 🤷🏼♂️
Now that is just frustrating. Vadim has a large audience, why post obviously nonsensical things?
@vadimyuryev Please, correct this. You are doing the community a disservice. AVX2 emulation is present on every ARM Mac, Armv9 does not have any "AVX2 support". What's more, M4 does not have SVE (scalable vector extension), which is expected to be included with a typical ARMv9 implementation.
I forgot he had an account here. Hopefully he sees this and corrects the record.
A more interesting conversation: is there any info yet on how Apple is doing AVX2 support? Presumably through the NEONs, right? Its performance should be interesting to test.
About the AVX2, I believe this is a step in the right direction , right now we just need to wait and see how developers adapt their GPKT2 pipelines as well in future releases.
But one thing is clear, there’s more and more games available on the Mac and publishers like Ubisoft are releasing some of their games on day 1
Most likely it’s related to have the videos posted on time and there isn’t a lot of info until after the panels and Apple documentation is fully updated during WWDC
Regardless their videos are really good for the general audience who wants to learn more, their real world tests are good. The only complaint I have is that they should have more work general Microsoft workloads opened.
But one thing is clear, there’s more and more games available on the Mac and publishers like Ubisoft are releasing some of their games on day 1
I think the guys at MaxTech got very confused over the latest couple of news. I love their channel but there-s some things that for sure need improvement.As usual Vadim appears to be wrong about technical matters. Not sure where he got that from. Also ARMv9 AVX2 support what's he on about? Is he confusing AVX2 with SVE2? Either way it doesn't matter since the M4 doesn't have SVE2 either. Its vector processors are still NEONs. Maybe he thinks SSVE2/SME2 is how Apple is emulating AVX2? That's probably giving him too much credit and while I don't know for sure, I also kinda doubt that's how Apple is emulating AVX2, but maybe? I would've thought it was being done through the NEONs even though they are 128b. But regardless, AVX2 support seems to be functional for the M1, M2, and M3 (which even if it is the matrix accelerator being used to emulated AVX2, again seems unlikely to me, the M1-3 have the matrix accelerator as well - they're just using Apple's private extensions rather than ARM's official matrix extensions).
Bottom line: there's no way to make sense of what he's saying here. Which, as I opened with, is pretty much expected for MaxTech on anything technical.
Sorry my "they" was unclear. I was referring to the Asahi Linux team, not Apple. Asahi Linux don't quite have nested virtualization upstreamed to Linux yet but when they [Asahi] do, it'll be for the M2. So it is in theory possible for Apple to support nested virtualization on the M2 too.
I've re-re-re-edited my previous post for clarity.
Most likely it’s related to have the videos posted on time and there isn’t a lot of info until after the panels and Apple documentation is fully updated during WWDC
Regardless their videos are really good for the general audience who wants to learn more, their real world tests are good. The only complaint I have is that they should have more work general Microsoft workloads opened.
This isn't a recent phenomenon with them. It's ongoing. Unfortunately for a general audience whenever they try to explain something even just a little bit technical they are typically very muddled and often flat out wrong like the above. This is far from the first time that they've essentially put out misinformation. I don't think it's malicious, but while they can run the tests and present them, they simply don't have the necessary expertise to expand on technical details and interpret the results beyond the surface level. And worse still they don't seem to pay attention to the people who do have the technical background to tell them differently. Which is why @thenewperson is rather cynical that a correction will be issued though in this case given that it is so glaringly obviously, provably wrong, that I hope it will be.
Look I'd love for them to become the channel they clearly want to be: a popular YouTuber with lots of Mac content (I know they do other things too) able to break down technical content for a general audience, but at this point even I have to admit that MaxTech simply doesn't have the right people for the last part. At first, I was optimistic that they'd learn, connect with engineers and technical experts, and, overtime, do better, but honestly the above shows that they're just getting worse.
EDIT: looks like his Twitter post has been deleted. I don't have Twitter. Did he issue any kind of correction? or just delete?
Last edited:
Rosetta 2 did not use to support AVX 2. [EDIT 3: Ah, I think I misunderstood your post to mean it always had supported it - i.e. prior to GPTK2. It does support it now and it does seem to be across every M* Mac.
About the Rosetta Translation Environment | Apple Developer Documentation
Learn how Rosetta translates executables, and understand what Rosetta can’t translate.developer.apple.com
The above webpage as of writing for instance has not been updated to include AVX/2 support. Sorry for the confusion. ]
AVX 2 is new to GPTK 2.0 and presumably works through Rosetta 2 and indeed it doesn't mention a requirement for M4. It should be easily testable by developers with Sonoma beta.
https://developer.apple.com/games/game-porting-toolkit/
EDIT: While the above webpage talks about bringing your game to both Mac and iOS/iPadOS I don't think iPadOS actually supports running Rosetta 2. So unless that's changed, it would be extra strange to have an M4 requirement as no one would be able to use it at the moment. The WWDC24 developer video shows the emulation environment running on a Mac with no mention of iPad:
Port advanced games to Apple platforms - WWDC24 - Videos - Apple Developer
Discover how simple it can be to reach players on Apple platforms worldwide. We'll show you how to evaluate your Windows executable on...developer.apple.com
So no I think this rumor about M4 only AVX2 support is wrong. Again, should be easily testable. @Xiao_Xi where did this rumor come from?
EDIT2: Ah it seems the beta AVX2 support is not fully working yet however CPUID claims it will work on M1s so it should be across the entire lineup once it is fully enabled.
Longhorn (@never_released@mastodon.social)
Kernel not aware yet that Rosetta supports AVX2 now, so no x86_64h binaries running or it being exposed through sysctl yet. You can use cpuid for probing its support rn.mastodon.social
Longhorn (@never_released@mastodon.social)
@chris@social.losno.co all, AVX2 works fine in Rosetta 2 on M1.mastodon.social
I think Hector got nested virtualization working in the M2 on Linux. This may be more like how Apple released clamshell support for more monitors for the M3 Air but the M2 could've done so as well (my fuzzy memory is that I think they changed the display controller since the M1 so it may not be able to).
EDIT: they [Asahi Linux team] don't quite have nested virtualization out yet [for Linux], but it was indeed developed and tested on the M2:
Hector Martin (@marcan@treehouse.systems)
@cinebox@hackers.town No, they beat us at that but barely... but only on M3 apparently? Linux NV support isn't fully upstream yet but it's been *developed and tested* on M2 so I have no idea what's up with that.social.treehouse.systems
Can you explain to me, like I'm a 6 yr old, why nested virtualization is interesting to me, to the average person, or even to the non-average person?
And will remain NEON.
SVE/2 AUGMENTS NEON, it doesn't replace it. Like AVX512 augments AVX2, it doesn't replace it.
Again, why is this interesting?
Is the point that it's a game based on AVX2 rather than on a GPU?
Not being snarky, trying to understand the technical implications.
Sure, I was just pointing out that the vector units don't support SVE2 like Vadim seemed to be suggesting and certainly has claimed in the past. I don't know if you saw the Twitter post that spurred this convo, but it was incredibly muddled talking about how ARM v9 was required for GPTK 2's AVX2 support.
Me? Personally, no, as a virtualization feature I would prefer PCIe passthrough. I'm probably niche so I'm not actually expecting Apple to support my use case anytime soon. But Apple saw fit to add hardware acceleration for nested virtualization and now support it in software. So enough users must want it that Apple did all that. My point was that the M2 had the hardware support, but Apple is, so far, only adding it to the M3. How many people that makes a difference to is unknown to me.
This again was based on the misunderstanding from Vadim that GPTK 2 AXV2 support would be limited to the M4, which it will not.
It runs on the GPU but it also required AVX2 emulation to work. This is apparently true for a decent number of recent games. A number of developers/porters have appeared over the moon ecstatic about AVX2 emulation support (more so than even ray tracing emulation support at least from what I've seen - one suspects because a game doesn't work at all without AVX2 support but the other just means the ray tracing can't be emulated). So obviously a big deal for them. And again, the TechTuber in question had said this emulation wouldn't work on the M1 when it ... does.
Last edited:
I have read two cases, but I don't know if they are popular. The first, running Desktop Docker in a VM instead of macOS and the second, running WSL in a Windows VM.
well, I get it that Apple now pushes customers to buy a Pro machine with at least Pro M chip ... to get a functionality that was available even on base Macbook Air since 2012 to the last Intel model in 2020
so not only they downgraded the Air ... but also a Pro with non-Pro chip ... it's a mess
The issue is that Apple display controllers are huge with massive SRAM caches for storing display state. The likely reason people give for why Apple does this is power savings, but there may be other uses for it as well. The downside is that the display controllers take up 25% of the base M-SOC die area. Increasing the number of them would balloon the cost of the die unreasonably. Now there is a discussion to be had (and people have had it on here) about why the display controllers have to be so big with questions like if it is just for power savings (and we don't know that) why can't they use more, smaller controllers that might burn somewhat more power but ultimately if you are connected to multiple displays you are almost certainly plugged in anyway? But I'm not fully equipped to answer that question and from what I remember of the discussion I don't think anyone has really any hard info, just speculation.
Even on the Pro and Max which have multiple external display controllers, they are all the same big size and only take up less die space proportionally due to how big the rest of the units, especially GPU, get on those SOCs. So Apple likely has a reason to spend that die area on large controllers even for those multi-external display controller chips, it's just not totally clear what that is beyond power savings which doesn't seem to quite fit to me or at least doesn't tell the whole story.
If I was to participate in unfounded wild speculation I might suggest that Apple's fast resolution switching and other display techniques is based on having such large caches but, again, I don't know that. And "knowing that" would require detailed comparisons between AS Macs and Intel Macs.
Last edited:
That person on X and YouTube claimed that AVX2 was only going to work on M4 but here is a game that requires AVX2 that runs on M1.
You see this everywhere in life. A person is a subject matter in X, and therefore everyone in the room implicitly thinks that person is informed in Y. Except for the subject matter expert on Y who is shocked that anyone is listening to this patently wrong idiot.
Thing is I feel this way and I can't even claim to be a subject matter expert myself, not in this topic anyway. And on the flip side, knowing enough to be dangerous, I'd hate to think how many wrong things I'd say with complete confidence if I were to try to run a TechTube channel. At least being a small voice participating on forums my wrongness is fairly well contained to a small audience and, since it is a discussion, can be more easily corrected by someone who knows of what they speak.
I contribute where I can. My job has me involved in tons of different fields of technology and troubleshooting. I’m the first to ask for clarification here on things I’m out of my depth on.
That said, the smartest people are always the ones that both A) Understand how little they know given the depths of knowledge on a given topic ,and B) Can determine how much another person actually understands about a topic by actually listening to them.
This MaxTech channel has always given me negative feelings in regards to B). As soon as a topic isn’t directly related to gaming specs they seem to have an above average (but nowhere near adequate to *explain*) understanding of computers. Lots of purely wrong extrapolations and throwaway comments that’s underline they don’t get how X or Y works under the hood.
Yes (so far). Rosetta 2 support for AVX2 is part of GPTK 2.0 announced as a Sequoia upgrade and is likewise part of the Sequoia Beta. Apple might choose to expand official support to older OSes. They may not. For those unconcerned with official support, this technique looks like it will work for those who don't want to upgrade from Sonoma and want Rosetta-AVX2, but once Sequoia leaves Beta, I don't think that'll be a major hangup.
The performance benchmarks are interesting. Difficult to know how expected that is, emulating 256b AVX2 with presumably 128b NEON, or if we can expect further improvements. Also not clear, beyond the emulation, how efficiently this particular benchmark is testing the two sets of vector units but I suppose the point is that unless you're the developer you aren't in a position to optimize the code for emulation in NEON. Finally of course as one person in that thread points out, if the AVX2 emulation is "good enough" to not be a (severe) bottleneck for these games, absolute emulation performance is less important.
Yep, that's the story for the last 3 consumer grade releases outside of the Pro.
And the thing is, when you start adding storage and memory on the Air model (Especially IN EUROPE) there's no point in getting any of the Air or desktop versions and just instead opt for the Pro versions.