Yes it would have to specifically be designed to enable 2-4TB VRAM & high memory bandwidth, as workstations tend to do.
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.
Where Does Mac Pro Go Next After M4 Ultra?
- Thread starter MacRumors
- Start date
- Sort by reaction score
As I mentioned earlier, as far as I've been able to tell, workstations don't do this. I've not been able to find any commercially-produced workstations (as opposed to servers) that offer that much VRAM. I've twice asked if you can cite any, and have yet to receive a response on that.
I doubt Apple would want to redesign the entire MP enclosure to accommodate the power and thermal dissipation requirements of the numerous GPU cores needed to work on 2-4 TB LLMs efficiently. That would make the machine cost-prohibitive for nearly all of its current customers. Which is why I continue to maintain that, if Apple were to produce such a beast, it would be an entirely different machine from the MP.
Hence I continue to respectfully disagree with your position that the MP should offer the capability to run that class of LLMs.
Last edited:
What the heck is the point of the Mac Pro these days? There used to be a point to getting a Mac Pro back before Clueless Cook decided to remove upgradable memory and upgradable storage out of the greedy hope that customers will buy a new Mac Pro when they eventually need more memory and/or more storage.
Apple is on an incremental path moving to composable chiplet designs, and as so I have money on the M4 Ultra not happening. Newer processes have reduced the die size / transistor count, which is part of the reason the Pro chips are no longer simply laid out as two of the baseline chips, but a separate design to fit into the size constraints.
The Mac Pro has trended toward a ten year design lifecycle, e.g. the power Mac design was revised for Intel but wasn't replaced until the trash can in 2013, which was replaced sooner due to its failed design in 2019.
Ignoring that they optimized the trash can for high core count/high GPU count computational load (which has still not manifested for generalized computing, and component upgrades were capped by thermal cooling capacity and limitations), a common complaint was that they optimized the design down to what the machine itself needed for cooling and power at the expense of any internal expansion - everything had to connect via USB or Thunderbolt.
The Mac Studio is the spiritual successor of the trash can, while the Mac Pro is the device that allows for internal expansion. The Mac Pro's viability hinges on customers needing and being able to use that internal expansion.
The Mac Pro's ability to be shrunk is dependent on whether it could still meet the needs of customers who are buying it for its internal expansion. That means it needs to have sufficient pcie slots, along with supporting power and cooling. I don't think you could reduce the size by half without also cutting the design from 6 to 3/4 slots.
It has been in weird position for 10+ years. The current form factor made sense for heat management purposes with Intel-based chips and PCI extension cards. It should have been totally redesigned for the M-family chips. Apple has been struggling for decades to come with an adequate product for the most demanding Pro users. At this stage, I don't believe Apple is selling even a 100 units per month worldwide. So it's even below the "niche-category", it's just non-selling. It is somehow shameful that the top of the line, most powerful Mac, is actually totally obsolete and lame product. It's hurting the brand. The best Apple can do is just discontinue this product and focus their efforts on Mac Studio.
Science fiction theory: Apple wants to invest heavily in AI farms using M5 chips. If Mac Pro is going to be relevant going forward, it needs to be relevant in AI terms. Maybe Apple’s own server farms will create enough volume to have a special high-end AI-focused version of M5, that goes into Mac Pro but not Studio. That might revive the interest in the Mac Pro.
All I hear is a bunch of people who don't do 3D saying "why do we need this, what's with Nvidia", try doing 3D and you'l find out pretty quick
In theory yes, but… I work in the audio industry, and I don’t remember the last time I saw anyone doing anything audio related on a Windows computer. Maybe it’s just habit, maybe it’s just that the ecosystem in audio software is built around Apple, but the cited reason is usually that they simply don’t trust Windows. The one important factor in audio recording is stability. Which is also why in audio recording you would never, ever, run a beta, and usually not even run the latest public OS version.
The fact that you can buy more processing power for less money in a Windows setup is completely irrelevant in that world.
To get ‘high end’ people interested it would have to have some access to multiple Nvidia gpu’s. Or maybe an Apple separate Gpu card that has more than the Mx has. But at a lower price point.
The problem is the cost to do that - if it could lure game players in then maybe the cost would go down but can’t see that happening.
The problem is the cost to do that - if it could lure game players in then maybe the cost would go down but can’t see that happening.
Not exactly useless. Mac Pro has SIX (6) PCIe 4.0 slots, and we can expect Gen 5.0 upon the next refresh, plus Thunderbolt 5. Those slots allow for truly massive internal SSD storage, and interfaces that move data much faster than even 10Gb Ethernet. Just because you don't use that does not mean there is nobody who does.
M4 Ultra's GPU cores are predicted to match the performance of NVIDIA's RTX 5090 GPU (which NVIDIA's RTX 4090 can also claim as well), and Apple has a lot of headroom for performance uplift in future generations. Plus, six PCIe 5.0 slots offer the tantalizing possibility of Apple producing M4 Ultra (or even successive generation) systems-on-a-card, using the Mac Pro case as an enclosed cluster for applications, such as artificial intelligence, that lend themselves to distributed processing. The Afterburner card was a preview of that possibility. Mac Pro doesn't have a 1300W power supply for no reason.
If M4 Ultra does match RTX 5090 performance, NVIDIA has a serious competitor. I think NVIDIA may be at apogee now.
d
I agree, except I think (without much actual insight to back it up) that the point where Apple will challenge NVIDIA will be M5 chips, not M4 Ultra. If nothing else, then because that is what Apple themselves seem to be banking on.
You may both be right. It may sound a little crazy, but there’s an Occam’s Razor argument (involving the Sequoia model identifiers Mac17,1 and Mac17,2) for M5 Ultra being announced at WWDC, to launch in late summer or early fall.
The M4 Max Mac Studio (Mac16,9) would launch at WWDC, with the M5 Ultra Mac Studio and M5 Ultra Mac Pro following a few months later.
In addition, the real kicker would come in March 2026: introducing the M5 Super — two M5 Ultras packaged together via TSMC CoWoS-L, the same advanced packaging technology used for Nvidia’s Grace Blackwell Superchips.
As @ZipoBibrok5x10^8 stated, the need for the PCIe slots is there for some people. Apple can't completely ignore them. What I'm amazed at is the additional fluff Apple puts in to the design of the Pro. They are beautiful to look at. The wheels are phenomenal, it's crazy to think they cost more than wheels for the average Toyota Corolla. 😄
I guess they figure if you need it, you're willing to pay for it. I miss the simpler, more affordable designs of the pre-trash can days. I wonder how much they've lost in sales from people back in the day that really didn't need a pro, but bought one anyways. Today you have to REALLY need one to buy one.
It is mainly for the graphics workloads and AI workloads that are all optimised for nVidia at this point.
Either Apple needs to provide something that can match high end Quadro and H-series cards in terms of performance or they need to find a way for them to be integrated into systems running Apple Silicon.
Apple Silicon is excellent at certain tasks and the shared memory architecture, allowing the GPU cores to use (almost) all of the memory on the system can provide huge advantages in certain areas, but the overall performance of those integrated graphics cores pales in comparison to high-end nVidia products.
Many companies are buying traditional PCs with multiple nVidia graphics cards, purely because they need the graphics or AI horse power that Apple just can't provide at the current time.
One is Alec Lindsay from O9O Media, he regularly laments on MacBreak Weekly that they have had to buy Windows PCs for certain tasks, even though they are an Apple shop... He has bought several Mac minis, for example, to carry out certain rendering tasks, because they are faster than traditional PCs, but they have also bought PCs with nVidia cards, because they do other tasks that Macs can't or can't do effectively. (He did a lot of work for Lucas Arts on Star Wars, for example and they do a lot of live broadcasts, including pioneering 3D content and modelling, including Vision Pro content.)
Me? I'm struggling to push my new MacBook Pro M4 Pro, but different folks have different needs. There is a small subset of users that need huge amounts of power in areas that Apple Silicon isn't delivering, whilst in other areas the Apple Silicon is running rings around traditional Intel kit. At the end of the day, it comes back down to the right tool for the right job.
We already have nVidia's project digits aiming for that LLM development market at the low end. Powerful workstations are a thing.
Expecting Thunderbolt v5 is basically a given since the M4 class so far have it. However, PCI-e v5.0 not so much. There is nothing indicative of M4 family so far to suggest PCI-e v5.0. Thunderbolt v5 is in no way technically dependent upon PCI-e v5. Two different standards. The version numbers are not required to match up.
None of the slots of the MP 2023 are feed directly from the M-series SoC package. All the slots are feed through a PCI-e switch. It would be extremely easy for Apple to change the backhaul between the SoC and the switch to v5 and the rest of the slots stay at exactly the same v4 peak throughput they have now. What would get is a more balance provisions overall max throughput. For example two x8 PCI-e v5 would be better overall backhaul bandwidth than x8 + x16 PCI-e v4 which have now . ( that upgrade would be basically equivalent to two x16 PCI-e v4 ).
Dual x8 would be more bandwidth at better Performance / watt for the overall package. That is pretty much the formula that Apple has be following along the M-series evolution. And the fact that the M2 had a asymmetric backhaul provisioning is indicative that there are constraints (balancing needs for other subsystems on the SoC chips. ).
If the rest of the M4-family doesn't have PCI-e v5 at all then easiest path for Apple would be just to keep the same set-up as before. Or a some kind of aggregate bandwidth uplift so get better balance across the slots.
There are bigger software/firmware problems there than hardware. Apple needs to address those first.
There isn't going to be six times PCI-e v5 bandwidth going to be there. The number of lanes coming out of the SoC package is going to be limited. ( at best 32 and very likely less than that. )
Afterburner was more so a wide distribution beta for putting video encoding onto the Apple Silicon itself. (throughly work out the bugs before putting it on a die that is harder to 'fix' in the field). Afterburner got subsumed over time. It is something that harbingers less slots ; not more slots.
Apple isn't particularly trying to put "AI" into slots either.
If wanted to do distributed processing across all of the slots in a Mac Pro then more backhaul bandwidth is the short term need. That doesn't necessarily mean PCI-e v5 to a single slot. If looking at aggregate across all of the slots then more to distribute to multiple PCI-e v4 slot would be a substantive improvement.
It is cheaper because it was already there in the MP 2019. That power distribution is not out to the slots though. The aggregate AUX power distribution went down relative to the MP 2019; not up. The design suggests that Apple put that power in reserve for themselves (Apple Silicon) and not the slots. Eventually if there was something bigger than an Ultra in power consumption they wouldn't have to change as much.
Not really a serious competitor even if got close ( which very likely will not outside of some 'cherry picked' select benchmarks). Nvidia is pragmatically banned from the Mac market. So there are extremely few (growing closer to zero over time) Nvidia cards in the Mac market. So the M4 will displace no Nvidia cards there.
The M4 Ultra will be soldered to a Macintosh. In the overall Windows PC market that natively boot Windows (or Linux) the M4 Ultra also doesn't even compete.
You have to directly compete to be a competitor. Unless Apple creates a M4 Ultra on a add-in-card there isn't going to be much competition in the general market.
In terms of the hype train. Perhaps. On top of the GPU market is more blurry. Depends upon if another hype bubble pops up and whether they are first one to catch the major tailwinds there.
Digits' specs:
128 GB RAM. ( unifeid system RAM)
That is off your "minimal TB" purposed requirement by an order of magnitude.
Don't have to put the largest piece of data possible into memory to compose a better algorithm. So yeah they will get used but this "every single developer seat needs 2-4TB or RAM" to be productive is a lot of hand waving.
more like Mac Pro 2013 ---> iMac Pro ---> Mac Studio
(somewhat of a deviation in the middle by also tossing in a screen but some non-screen internal elements of iMac Pro moved across. )
Mac Studio is a successor to the large screen iMac. ( lots of large screen iMac fans don't want that to be true, but it basically reality now. )
That ended on MacOS before the transition to Apple Silicon. That is basically over. Nvidia helped blow up and burn that 'bridge' . Nvidia got mega richer by paying attention to other markets. They aren't coming back to beg for absolution. Same thing is largely true for Apple. Across the whole mac line there is a general large uplift by Apple focusing solely on their own Apple GPU graphics driver evolution.
Nvidia moving in direction of selling their own branded hardware systems is just even more gas on the bridge fire.
Maybe a bit simplistic - rework the paradigm so that yes, the M4 or M5 can offload specific tasks to "modules" that are dedicated functions - items likely to see benefit - various video work, AI, 3D rendering and more. This would be what sets it apart from the Studio - ability to add modules or at least order them at purchase. Any activity that will take more than a few moments should be offloaded.