Even the M1 covers what 99% of people need, if you are a power user you are better off with something that is upgradable anyway. Having the option to add an nvidia card for AI is a big deal, and to upgrade to a newer generation when desired without having to buy a whole new pc is what power users need.
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.
ARM Mac Pro - WWDC 2023
- Thread starter Zest28
- Start date
- Sort by reaction score
I agree with your premise, but disagree with the last statement. The thing is, Apple is very serious about the GPU compute — on their own hardware. They don't care about the users using compute solutions from other vendors, because they want them to use their own Metal.
The way I see it is that suggesting that Apple supports third-party compute stack is like suggesting that Nvidia supports AMD or Intel GPUs in CUDA. They can do it, but it would be detrimental to their business. CUDA offers a streamlined programming model with guaranteed feature support and consistent behaviour across GPU models. It's not something that you can support across multiple vendors, not at this level of abstraction. With Apple it's similar — they support certain streamlined programming model that offers strong guarantees (zero memory copy, large amounts of GPU-addressable fast RAM, fast tile bulk transfer, SIMD shuffle and fill operations etc.). They want the software to be developed and optimised for their hardware, not somebody else's.
There are two common counterarguments to this strategy, and I think both can be refuted.
1. PCs support multiple GPU vendors, so should Apple. I think this argument falls flat from the beginning since Appel is not the PC business at all. They sell opinionated systems, not flexibly configurable boxes. There is no doubt that Windows/Linux on a generic PC offers much more flexibility and choice to the users, but that is simply not a concern for Apple. This was the case even when they were using third-party components, and is even more the case now, when they moved to their own hardware.
2. Apple will lose pro users if they don't give them access to fast Nvidia GPUs. Yes, they will, and yes, they already have. I don't think this is the problem for them, because they are looking at a long-term business vision, not short-term. Again, we have a good precedent in the industry: Nvidia. They were slowly building their hardware and software (CUDA) stack, "seducing" users by the advanced functionality. There is no reason why Apple can't do the same. We know they have the technology, talent, and of course money. The latter is a big factor — they can afford to spend much more money on R&D and manufacture without it eating into their profit margins.
To sum it up, I believe the reason why Apple doesn't allow third-party GPUs is similar to why they don't support Vulkan. While this strategy might be detrimental in the short term, the potential payout in the long run is much higher. If they can offer compelling hardware, good tools, and a consistent, comfortable programming model, they will end up with a rich software ecosystem in the future. It's a calculated risk.
I agree, PCI-e passthrough could be an interesting feature for a Mac Pro. Put an Nvidia GPU in, fire up a Linux VM, use CUDA. Of course, the GPU would be unusable under macOS. At the same time, there are also potential challenges with this, e.g. will Nvidia drivers even run on ARM Linux?
And Apple can make the chip however they want. Just because the MacBook Air doesn't have a chip that supports it, doesn't mean they can't add it to the Mac Pro.
That is precisely why I find it so frustrating hearing non stop discussions about the new MacBook Air instead of even ONE SENTENCE about a Mac Pro or if the Mac Studio is STILL one and done as some rumors suggest.
The question isn't "can they do it?", it's "Will they do it.? We'll just have to wait and see. Note I never said they wont, just that I'm not sure they will.
The poll should have an option for "how should I know?"
Probably so.
I want to see it though.
Okay, a real prediction if they announce it.
It will be an "Extreme" chip, though it will have more PCIe lanes and other expandability ideals other than just more cores. They might call it something besides M? to differentiate it. Maybe MX?
They will decouple the RAM from the SoC so it can be single tier. I really hate the idea of 2 tier RAM from an OS perspective, ugly doesn't cover what I see as the possible problems. At the least, everything will have to be rewritten to take *full* advantage of such a thing, and since nothing else will have 2 tiered RAM, it wont be worth it to change the OS and apps.
They will also have a real disk RAID capable controller in the box on a PCIe lane. (for expansion) They'll still have the built in disk controller, but it will only control the base storage and it'll be just like current machines.
More iGPU cores, but I think they'll also have to possibility of an Apple expansion card with more cores as well.
And it'll, of course, be out of my price range for the tasks I would use it for. I don't want much, do I...
There is no sane way Apple can confine decent PCI-e lane provisioning just to the "Extreme" SoC. That SoC is likely to cost over $4K just all by itself. Just the chip ( the more you include the new 'floor' for the RAM 128GB and rockets even higher. ) .
If Apple was adding decent PCI-e provisioning to the "Extreme" they pragmatically would have to add it some "Ultra"-class like SoC also in order to be able to still hit the $6K starting entry price they currently have. Rocketing the Mac Pro's entry price even higher is likely a death sentence for the product.
Apple has already set the price for the Pro , Max and Ultra ... doubling the Ultra's price gets something that is quite expensive. The Mac Pro can't 'start' there and build any kind of decent sustainable user base. Apple has already jacked up the MP entry price point 100% . It is already tipping on death spiral status. Even higher will make death just all that more likely .
Keep on dreaming. There is not a single GPU in this performance class that as detachable RAM. There is a technical reason for that. Go looking your workstation GPU cards in the Windows/Linux market. See any deatachable RAM? Nope. Apple has the same constraints.
And the notion that Apple is going to 100% detach ALL the GPU cores. More than highly unlikely. The primary point of the "Extreme" is to crank the GPU core count sky high. If the GPU cores are 100% gone it isn't an "Extreme".
The vastly more straight forward thing to do is construct more deliberately designed chiplets and then can can a "less than an extreme" at a far more reasonable price and an "extreme" that dramatically fewer folks can afford.
It doesn't have to get 'ugly' if the Non-Uniform Memory Access (NUMA) RAM is mainly out of the mainstream App usage (apps which presume that it s all highly UMA . MP 2009-2012 models has RAM hanging off two different CPU packages. The VRAM in the Intel era was mainly NUMA. MacOS isn't extremely brittle to NUMA access.
The bigger issue is that cannot squeeze max performance out of the M-series SoCs if don't leverage the UMA characteristics explicitly. For the whole ecosystem it is far better outcome long term to have software which gets the most out of the design decisions that Apple has made about UMA.
If Apple took expandable RAM and totally assigned it to the File System disk caching subsystem than that would be extremely opaque to most apps and lead to faster storage I/O (in most cases ). Lots of very large memory footprint apps are pragmatically building their own RAM SSD by mmap-ing large chunks of files into RAM. Paging mmap file blocks into a swap file on a RAM SSD wouldn't be all that 'ugly' to do. macOS compresses RAM pages not being used already. That new task would be just a variation on roughly the same theme.
Apple barely supports 3rd party single SSDs ( trimforce "on" isn't the default ). Apple leveraged the "for free" SATA controller in the MP 2019's PCH I/O controller chip. ( Apple never touched any if Intel's PCH RAID features in any model. ) It is just going to be a hurdle to get them to put a simple discrete SATA controller on the board, let alone some highalutin 'real RAID' controller. Likely, if there is a x1 PCI-e v4 SATA controller out there for a couple of "JBOD" SATA channels that is all you'll get.
IF some users wants a "real RAID" card ... that is what the PCI-e slots are there for. Apple's isn't going to install it. (or put it in the BTO configuration). Also, kind of funny how folks grumble at soldered on RAM , but don't grumble about soldered on 'RAID card".
The Mac Pro is likely going to be sold just like the rest of the Mac line up; one Internal Apple SSD. That's it. The major difference with the Mac Pro is that the user can optionally add another 3rd party disk(s). Apple doesn't want to mandate 'what' those are though.
I/O expansion is also a big thing. We will find out eventually, but card slots for I/O and SSD expansion I think are the bare minimum expectations. I think MacOS also supports tiered RAM so you could have non-unified RAM (CPU Only or somesuch) as another easy offering.
-d
Standing by to eat crow along with everyone else who has speculated.....🤣
Although I would love for the announcement to have Gurman eating massive crow.
Although I would love for the announcement to have Gurman eating massive crow.
It's a high cost machine for a reason. But more PCIe lanes would be nice, but a waste on anything less than the Pro. I agree it's different than what we have now, but doable.
That doesn't mean it can't be done. 2 tier is going to be much more of a pain, so you'd be limited to the amount of RAM you could have on chip, and that wont be enough for a Pro machine. The timing differences are enough to make the whole thing go crappy, and that's not the only major problem.
Perhaps they only sell GPU's with build in RAM for the same reason Apple does for SoC, they can make more money at it...
I never said they'd detach the iGPU cores, the SoC would still have the same cores the extreme chip should have. just that there would be a possibility of expanding even that. Maybe not for general purpose stuff, but specialized.
I'm not wild about NUMA for something like this either, it's ugly. They only place I see NUMA being handy is for virtualization.
But as I've said before, I'm not a hardware guy, just software and more specifically OS guy. What I want may not be possible hardware-wise, but if it isn't I'd just throw in the towel as it wouldn't be good enough. A stack of compromises is just a house made of cards. Maybe that's why Apple has taken so long and we still don't have a machine.
As for RAID, I'm also an IT guy and just love redundancy, and would buy a machine with it built in over one that's a separate card, and JBOD isn't even something I've EVER done. If it can't do RAID 5 at the very least, it's not reliable. I'm actually quite surprised that Apple never used it, I didn't know that. (even when they tried to do a server version of the OS?)
Anyway, what I said is all wishes. I can get what I describe quite easily and a lot cheaper in Intel/AMD land...
My guess is they will have something saying 'more on the mac pro later this year'. Maybe time it with some M3 stuff.
LOL epic fail. A Mac Pro with only M2 Ultra? Seriously? Way worse than Mac Pro 2019. They just proved that they can't make workstations.
Not garbage, but RAM is not so good. I like that you can put "video i/o card" I wonder if it's a full GPU, but I kind of doubt it.
Not much surprises, not what i wanted, but not nothing either. Available next week!
well I think the idea is maybe in a few years you pull out a black expansion box from the pro guts and put in a new one containing an M3 Ultra (for example).