That excuse isn't going to work. If it's real and it doesn't have good games, it will 100% be a failure. But again, I don't even believe it's a real product, I think everyone has been massively trolled by unsubstantiated rumors.
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.
Apple Silicon Mac Pro Reportedly Not Coming at WWDC, Mac Studio Refresh Likely Delayed Until M3
- Thread starter MacRumors
- Start date
- Sort by reaction score
And if this is true, then they shouldn't bother with a Mac Pro... they should just work on improving the Mac Studio as the highest end Mac.
Any Apple SoC that has Thunderbolt by necessity has PCIe. The only difference for a Mac Pro would be internal PCIe expansion slots. Those don't require any change on the SoC.
What would require a change is a heterogenous memory architecture. For example:
- if Apple were to provide DDR memory slots like the 2019 Mac Pro has, they could leave out the on-package memory altogether. This would be more expandable, but have worse latency.
- or, they could create a complicated architecture where the SoC takes its memory from both the package and the slots. Presumably, if they were to do that, you'd have to define at application startup whether a process gets the package memory (fast, but small) or the slots memory (potentially larger, but slightly slower).
- moreover, even leaving aside drivers and all, if they allow you to add GPUs via PCIe slots, those would need their own graphics memory. This would by necessity be slower: right now, an Apple SoC's CPU can write an entire region of memory, and then the Apple SoC's GPU can read that same region, with zero moving/copying involved. With a separate GPU, you'd lose that advantage.
But PCIe itself isn't a problem at all. There are no slots, but the bus is already there.
Exactly. That's… what the Mac Studio already is. Unless they provide some kind of internal expansion (RAM, PCI, storage, or all three), it… would simply be a Mac Studio but possibly with more thermal headroom, for a quad setup. Why would that need a separate product moniker at all?
Because, Apple thinks enough people want PCIe card slots, beyond what external chassis can do?
*nods in agreement*
Considering Apple went external expansion in the 2013 Mac Pro and was effectively forced to return to internal expansion in the 2019 Mac Pro, I could easily see external expansion (via Mac Studio) being viewed internally (to say nothing of externally) as non-starters for the 2023 Mac Pro.
PCIe slots almost certainly do require some new silicon... Having PCIe buses (or having Thunderbolt which can tunnel PCIe) is not the same thing as supporting a large number of PCIe slot lanes.
The Mac Pro could have about 90 lanes worth of PCIe slots... which is probably WAY more than even multiple M* Max silicon can offer.
Your point of power efficiency is completely irrelevant in a desktop workstation. We are talking about a desktop machine, not a laptop. It doesn't matter, as long as radiator noise is not an issue of the design.
Who said "large number"?
It could have 90, sure. It could also have 2,000. But the 2019 Mac Pro has 64, and I'm not sure why you would assume the new one has 50% more.
I was responding to "High performance, but not via hot inefficient bundles of PCI GPU and huge amounts of off-chip RAM." If this Mac doesn't have "inefficient bundles of PCI GPU and huge amounts of off-chip RAM", that doesn't leave much. Yes, you could theoretically do a Mac Studio but with PCIe card slots except for GPUs, but how much of a market segment is that? And at that point, wouldn't it be easier to just tell people to buy a Thunderbolt chassis with five PCIe slots?
No, I'm sure they don't think that, unless they are including GPU support in that thought. There are definitely not enough people that want PCIe slots that can't utilize GPUs to justify making a new crippled Mac Pro.
I rounded? I read the Mac Pro as currently having 92 lanes worth of slots, even if the Xeon is under provisioned for it. Apple probably would under-provision the slots again, but I saw no reason to have the upper-bound be the physical slots.
The difference between 64 and 92 is moot here, of course.
I think we just disagree and I think there are lots of non-GPU use cases out there (though I buy that Apple will support non-video-out GPUs of some kind), and that five PCIe slot Thunderbolt Chassis would be horribly bottlenecked.
They could make an expansion box for the Studio to do the same thing. I'm not convinced about the PCIe slots being fast enough though. The TB lanes are just too painfully slow when when hooked up to external SSD's. Maybe I just haven't seen the right peripherals yet and I'll be pleasantly surprised, but....
When Apple announced their switch to Apple Silicon. My first thought was that it will work well except they will not be able to use it for Mac Pro. But now I think there is a way...
What if they place the M2/M3 max chips on a card. We call this "compute modules" because they have the processor, RAM and suport chips on the card. These cards plug into a fast bus. The user can buy any number of these cards and can replce them with newer cards. Each card has something like an "M3 Max" and you can buy them with some amount of RAM for different places. The way you upgrade the memory is to replace some or all of the compute modules.
Once you place the compute module on a bus then you can use the same bus to connect others kinds of modules like "GPU modules" or "storage modules" or "I/O modules" that have thunderbolt ports. The chassis is just a rack for holding these modules. But you need a REALLY fast bus to make this work and I don't think such a thing exists.
So the RAM could be upgraded when you replace a compute module with a more powerful compute module.
What if they place the M2/M3 max chips on a card. We call this "compute modules" because they have the processor, RAM and suport chips on the card. These cards plug into a fast bus. The user can buy any number of these cards and can replce them with newer cards. Each card has something like an "M3 Max" and you can buy them with some amount of RAM for different places. The way you upgrade the memory is to replace some or all of the compute modules.
Once you place the compute module on a bus then you can use the same bus to connect others kinds of modules like "GPU modules" or "storage modules" or "I/O modules" that have thunderbolt ports. The chassis is just a rack for holding these modules. But you need a REALLY fast bus to make this work and I don't think such a thing exists.
So the RAM could be upgraded when you replace a compute module with a more powerful compute module.
That's a Thunderbolt limitation. Thunderbolt 3 and 4 are PCIe 3. Thunderbolt 5 will double the bandwidth (which IIRC still isn't quite PCIe 5). Either way: internal slots can be a lot faster than that, if Apple gives them enough lanes.
Fair enough. My point is: adding more lanes isn't quite the same level of work as adding PCIe from scratch. Any Mx chip already has PCIe, but yes, to your point, it may not be enough lanes to support meaningful internal expansion.
This is much trickier in practice than it sounds. If you have two such modules, which of them is the authoritative scheduler? What about RAM — does each process get one module's RAM? Do you merge the RAM into a single pool (this creates huge synchronization bottlenecks)? Etc.
The beauty of the current design is that you have a smooth cache hierarchy. The CPU cores have L1 and L2 cache, the SoC has L3 cache, and then the package has "L4 cache", i.e. RAM. Having two packages or more complicates that a ton for relatively little gain.
(Such a "fast bus" doesn't really exist, for the above reason. The CPU itself has to be the fastest bus.)
Then Apple should keep its mouth shut and just focus on delivering products when they're ready.
Who implied Apple was misrepresenting themselves? So you have a complicated small workstation that has taking a lot of time to redo based on a different processor platform. Everyone here it seems likes to say why does it take so long, but they complain about various things they would like to see in this released. So who needs to really keep their mouths shut, the ones wanting everything, or the ones designing the workstation? Or is it all those clickbait people that keep telling you its coming sooner or later because they like being in the spotlight.
They normally do, but I think they didn't want people to speculate that they were going to discontinue the Mac Pro... except they obviously ran into trouble trying to adapt the M series to a Mac Pro replacement...... which means what they promised might have been meaningless and they actually will end up discontinuing the Mac Pro... lol.
I think we both suspect they want keep two products going with updates until there is enough of a technology leap to no longer need that both platforms. But they need to likely replace the 2019 model with something far cheaper to kit with. As I commented the 2003 to 2012 models were half the price and people regularly bought them to use for a variety of setups.
Thunderbolt in Apple Silicon does not support eGPU. Probably not possible to use Thunderbolt/PCIe for graphic card due to ARM CPUs limitations. Intel introduce Thunderbolt 5 and will be oficially released this year . Native PCIe 5.0 are available only for Intel/AMD CPUs. So it is impossible for silicon macpro to have these expansion slots.
I'm a bit confused by the above...
Apple could support eGPU for video-out on ARM Macs if they wanted to. They have seemed to just build the OS without support for it. That's a choice Apple has made. It's not the end of the world as Apple's iGPUs are "fine" for the vast majority of users.
Apple will certainly support Next Gen Thunderbolt (if called Thunderbolt 5 or something else) eventually.
I don't understand... why can't Apple just support PCIe gen5? I'm not aware of anything stopping them from supporting it.
No that's wrong, they built the CPU without support for it. There is nothing they can do with software with M1 or M2. Maybe they will fix this with M3, but it doesn't seem like something they are interested in.
macOS tracks active threads with a 64-bit integer. Windows/Linux all did this with the nominal move to basic 64 bit support weaved in.
Window has two somewhat forked kernels. "regular windows" and Server / "high core count" Windows. You'll see when ordering very high workstations in many "BTO" systems that the Windows version is out of whack if move from 6-8 core version to a 64 or more core version.
There has been lots of hand waving in several threads that there is 'trivial' fix for those of just do a typdef redefine of int64 to int128 and do a simple recompile and everything will work spectacularly well. The history of the upgrades of various Unix versions , Windows , and others is a different story. This has a secondary impact of impacting the schedule, layouts of some critical kernel structures , etc. The propagation ripple if do it robustly is more than just a recompile or a minimal impact 32-bit to 64-bit app conversion.
In single user , mostly single application space (where treads largely sharing same address space) ? Not really.
In flogging 10+ year old , unmodified legacy apps into going faster with "just use a bigger hammer" approach? Probably.
There is lots of work that is very amenable to "embarrassingly parallel" processing that is far more perf/watt effective on non general purpose cores than it is on the cores that the operating system has to track. Apple AMX can be kicked out into a specialized unit that the OS doesn't track in terms of 'pthreads' . NPU and GPU ... not pthreads.
Would they fall behind if didn't have other specialized core units weaved in ? Maybe. But that is an alternative universe.
Since they don't have simultaneously multithreading , SMT (Intel's marketing HyperThreading), they can gradually creep up to 64 cores over a long extended amount of time. If went 40 , 50 , 60 on every 2-3 generation gaps that would be more than several years. [ And no SMT closes some security loopholes they don't have to work around. ] The other issue have is that have to 'share' with the other core types. At some point have to rob-peter-pay-paul to get more die space (even with chiplets. The package can only get so big before substantive NUMA issues kick in. )
There are far more folks moaning and groaning about not being a "4090 killer" than about chasing maximal core counts with minimal clock speeds. CPU cores is not the one that they have a looming deficit over.
Unless, Intel/AMD fix the base clock decline as cores go up issue , Apple really isn't going to struggle much here in the single user , single app space. Especially where those apps have 30/70 , 40/60 , 50/50 single thread : multiple thread critical code sections.
Another large issues is that cache SRAM has stopped scaling. To get bigger on-die caches going to need to increasing scrifice on cranking up core counts. Or take the cache hit. ( can play 3D cache tricks , but that is going to cost more . )
Sierra Forest for workstation in the space that Apple Mac Pro has typically had traction in is likely very , very low. It will probably be a product in the broader and much larger general market , but primarily because it is a small niche. ( so much bigger pond makes it viable). There are folks like "Moorse's Law is Dead" who keeps chatting up SF as some cool deal for the gamer market. I think his analysis is pretty far off. he is just talking crazy core counts to draw ad clicks.
the Ampere Computing ARM max core count , Sierra Forest , Bergamo processor package stuff is way , way , way , way off in the swamp from where the mainstream workstation market is going in terms of substantively large sales. That stuff is far more so to shrink the 2-4+ socket server market into single sockets than single user workstations.
Nvidia isn't necessary for AI. It is the hype train... but crypto mining with generic GPU was a hype train 2-4 years ago too. AI processors are likely going to get more AI-ish and less generic GPU-ish over the next 4 years. It has already started and will only pick up speed as AI/ML gets more traction. Perf/Watt wise it doesn't make much sense to do otherwise long term.
Apple may retire the generic box with slots which is mainly bought as a container to fill with 3rd party stuff. But I don't think likely going to kill off the Mac Pro. I think there is disconnect of a user subbase who primarily only view it as the first. Apple does not . Didn't really before and probably even less so now. Apple views it as a system and a Mac first and optional additional as a secondary feature.
the 'display GPUs' I won't be surprised to see go. But zero to very few PCI-e slots would be. It will still likely have some "augment the contain" abilities. Just not catered to Nvidia gaming fan boys.