Apple Silicon Mac Pro Use Cases

[aParkerMusic]

Hi everyone,

I’ve been wondering what sorts of uses the current Mac Pro exists to fulfill. I know it’s not as customizable/expandable as the Intel model, but that it still allows some measure of expansion (PCIe ports?).

How can it be expanded? What sorts of modifications would people make, and for what uses?

Does any here currently use one of these Apple Silicon Mac Pros?

Appreciate anyone’s input!

 

[mfram]

I don't have a Mac Pro but I can guess what the external PCIe ports might be for.

First would be some kind of high-end network or other low-latency computer interconnect that can tie a bunch of computers together. Used to implement clusters of computers working together on various computational tasks.

Another use would be to connect to enterprise-grade fiber channel storage area networks. Think of them like external RAID storage controllers, except much more expensive with much more functionality. They can provide redundant access to hundreds of TB or maybe now PB amounts of storage. That would be needed if you're working on large video/media files or maybe AI. If a company already invested in high-end enterprise storage you'd need external PCIe ports to join an existing setup. I worked with these in the past on non-Mac servers.

 

[mcnallym]

Essentially non GPU PCI-E cards. So network, storage, video capture.
For people that for some reason unable to use a TB enclosure to put the cards in.

Is a very niche market now as far as Apple concerned.

 

[JouniS]

Essentially non GPU PCI-E cards. So network, storage, video capture.
For people that for some reason unable to use a TB enclosure to put the cards in.

Thunderbolt is pretty slow for some applications. It's usually one generation behind PCIe. You would need two Thunderbolt 5 ports for a single PCIe 5.0 SSD, or for a 100-gigabit network adapter.

 

[maflynn]

but that it still allows some measure of expansion (PCIe ports?).

Yes, there's some expansion cards available, but much of the functionality of those internal expansion cards have been replaced by external thunderbolt peripherals.

How can it be expanded? What sorts of modifications would people make, and for what uses?

It can't, the architecture of Apple Silicon means that ram, gpu, and internal storage is all locked in place at the time of purchase.

Does any here currently use one of these Apple Silicon Mac Pros?

I'm sure there's people using it, just pop over to the mac pro forum and ask. I think the majority of people who need a powerful desktop have opted for the studio. The best you can get with the Mac Pro is a M2 ultra, where at the Studio offers a M3 Ultra and its been rumored we'll be seeing a M5 Ultra studio, where those same rumors also indicate that Apple is shelving any Mac Pro updates.

Food for thought.
M2 Ulta Mac Pro: 24 core/60 gpu/32 neural cores. 64GB of ram, 1TB of storage: $7,000
M3 Ultra Studio: 28 Cores/60 GPU/32 neural cores. 96GB of ram, 1TB of storage: $4,000

Why would spend 75% more money on a slower desktop that has less ram?

 

[StuAff]

Best summaries I've seen (though the current CPU imbalance & addition of TB5 firmly tips the equation in favour of the Studio), from @rondocap https://www.macworld.com/article/24...licon-chip-pcie-expansion-storage-cables.html and Dave Plummer.

 

[dmccloud]

Something that I don't think has been released for the Studio yet, but would be a killer addition for the networking side would be plug-in cards that support ConnectX. Just as a comparison, the DGX Spark has both 10 Gbps Ethernet and 200 Gbps ConnectX. The latter format is already supported by companies such as Ubiquiti, and can go up to 400 Gbps. The ability to connect multiple Mac Pros in a cluster akin to what can be done with the DGX Spark via ConnectX would be a gamechanger.

 

[maflynn]

The ability to connect multiple Mac Pros in a cluster akin to what can be done with the DGX Spark via ConnectX would be a gamechanger.

Tahoe 26.2 allows you to connect multiple Studios via the TB5 interface

You can turn a cluster of Macs into an AI supercomputer in macOS Tahoe 26.2

Who needs a revamped Mac Pro when you can just turn several Mac Studios into a unified computing system? With the upcoming macOS Tahoe 26.2 release, Apple is introducing a new low-latency feature that lets you connect several Macs together using Thunderbolt 5.

 

[tenthousandthings]

Something that I don't think has been released for the Studio yet, but would be a killer addition for the networking side would be plug-in cards that support ConnectX. Just as a comparison, the DGX Spark has both 10 Gbps Ethernet and 200 Gbps ConnectX. The latter format is already supported by companies such as Ubiquiti, and can go up to 400 Gbps. The ability to connect multiple Mac Pros in a cluster akin to what can be done with the DGX Spark via ConnectX would be a gamechanger.

Nvidia acquired Mellanox for $6.9 billion in 2020, so ConnectX-7+ cards with drivers for macOS seem unlikely, to say the least. They support the InfiniBand protocol, and as noted above Apple is introducing InfiniBand support in macOS Tahoe 26.2, so I could be wrong but I guess maybe a third party (or Apple itself) could support those cards in a hypothetical 2026 Mac Pro.

 

[Alameda]

The takeaway I’m reading here is that a Mac Pro could make a beastly server.

I’m sure that Apple could hire a team to make all of the hardware and software stuff that server people love. But then they’d have to lure IT people away from Windows or Unix to make the sale. Apple has never been an enterprise-focused company.

Apple could offer a big power/cooling advantage with the M-series processors, but I think it would take a lot of effort and time to obtain even 10% market share, because their customers would need to port so much software over to MacOS or a variant. Yes, they could leverage their UNIX base, but customers would still need to test all of their software on this new platform, and Apple hasn’t been a trustworthy long-term supplier of IT equipment.

On the workstation side, it seems that the Studio platform and architecture will be their direction for some time to come.

 

[deconstruct60]

Nvidia acquired Mellanox for $6.9 billion in 2020, so ConnectX-7+ cards with drivers for macOS seem unlikely, to say the least. They support the InfiniBand protocol, and as noted above Apple is introducing InfiniBand support in macOS Tahoe 26.2, so I could be wrong but I guess maybe a third party (or Apple itself) could support those cards in a hypothetical 2026 Mac Pro.

Apple isn't introducing Infiniband protocol support. All Apple is doing is "Infiniband verbs". Really software API calls to networking rather than the actual networking protocol. There is software libraries that layer on top of RDMA to get work done. Infniband is not == to RDMA , but at a more higher software abstraction layer folks have used the 'verbs' associated with the protocol as common ground. ( RDMA on converge Ethernet. RoCE). On Linux there is a 'infiniband verb' library that is a 'standard' library to bind/write to.

Modern ConnectX connection will 'smoke' Apple's solution.

There already RDMA Ethernet cards for macOS.

https://www.atto.com/introducing-atto-technology-ethernet-suite/

What Apple is doing is more of a more affordable ( 4 Studios with point-to-point links to the other 3 nodes ) by avoiding switching and additional latency. But not fastest; not going to hit 100GbE speed let alone 200 or 400.

 

[deconstruct60]

Tahoe 26.2 allows you to connect multiple Studios via the TB5 interface

You can turn a cluster of Macs into an AI supercomputer in macOS Tahoe 26.2

An AI supercomputer from 5-10 years ago. "Supercomputer" is a bit over the top in respect to a modern context. TBv5 can't handle a single 100GbE let alone multiple sockets per adapter. Scaling past 4 nodes is way under what a modern supercomputer minimum standards in the modern era.

Apple has more of a "home lab" solution where the focus is likely on new, novel algoritms and development than on contemporary Supercomputer sized problems.

 

[theorist9]

I’ve been wondering what sorts of uses the current Mac Pro exists to fulfill.

The M2 Mac Pro offered about twice the I/O bandwidth of the M2 Ultra Studio. I'm not sure how it compares for I/O with the M3 Ultra Studio.

In addition, the following statement by maflynn is incorrect, since the Mac Pro allows you to upgrade the internal SSD post-purchase (it's the only AS device that offers this), with kits available for 2 TB, 4 TB and 8 TB, e.g.:
Apple 4TB SSD Upgrade Kit for Mac Pro [The RAM and GPU, however, are not upgradeable.]

It can't, the architecture of Apple Silicon means that ram, gpu, and internal storage is all locked in place at the time of purchase.

 

[maflynn]

In addition, the following statement by maflynn is incorrect,

Ok
But that still doesn't really change the fact that the Mac Pro, like the rest of Apple's Mac Line up cannot be upgraded (ssd aside). Technically you can upgrade the mini and Studio's storage as well, its just not sanctioned by apple. So again the use cases for a 6,000+ mac pro can largely be handled by the $2,000 Studio.

 

[tenthousandthings]

Apple isn't introducing Infiniband protocol support. All Apple is doing is "Infiniband verbs". Really software API calls to networking rather than the actual networking protocol. There is software libraries that layer on top of RDMA to get work done. Infniband is not == to RDMA , but at a more higher software abstraction layer folks have used the 'verbs' associated with the protocol as common ground. ( RDMA on converge Ethernet. RoCE). On Linux there is a 'infiniband verb' library that is a 'standard' library to bind/write to.

Modern ConnectX connection will 'smoke' Apple's solution.

There already RDMA Ethernet cards for macOS.

https://www.atto.com/introducing-atto-technology-ethernet-suite/

What Apple is doing is more of a more affordable ( 4 Studios with point-to-point links to the other 3 nodes ) by avoiding switching and additional latency. But not fastest; not going to hit 100GbE speed let alone 200 or 400.

What do you make of the Ultra Ethernet consortium—that seems like the obvious way to bring those higher speeds, *if* Apple wanted to provide that as an option for the Studio?

 

[deconstruct60]

What do you make of the Ultra Ethernet consortium—that seems like the obvious way to bring those higher speeds, *if* Apple wanted to provide that as an option for the Studio?

The consortium ... I don't think it is about 'higher speeds' as much as getting useful, full throughput from Ethernet. The legacy congestion control , the switching issues at ever faster speeds, and interoperable ( more vendor competition). The consortium is trying to 'tap dance' between following the Ethernet standards and trying to solve some deep seated disconnects with very high speed , 'scale out' with some data traffic patterns that get entangled with the legacy protocols. I doubt Apple is interested at all. The major Ethernet switch players all want some advantage, but need to be compatible for this consortium. That could be an unstable oligopoly.


When here about hooking massive AI datacenters in separate geographic location together ( in part because each data center exhausted the local electrical supply) that is the sweet spot for Ultra Ethernet. Similar thing with these massive campus of 4-5 relatively large buildings. Again where UltraEhernet probably will get traction.

Infiniband has scale out limitations. Can try, but it gets very expensive (not that Ultra Ethernet will be dirt cheap; just relatively cheaper.). I think some folks hope that Ultra Ethernet could get more traction at the rack level scale to displace more Infiniband. For example:

Broadcom Tries To Kill InfiniBand And NVSwitch With One Ethernet Stone

InfiniBand was always supposed to be a mainstream fabric to be used across PCs, servers, storage, and networks, but the effort collapsed and the remains

www.nextplatform.com

It is pitched as a "Infinband killer" likely will run into problems.


I suspect that whatever Apple is doing with Broadcom is oriented toward , at least in part, very fast Ethernet (and perhaps other compute and i/o). For example, all the 'inbound' client requests are coming in through Ethernet so if aggregate tons of substantive data 'pulls' onto a single machine they'll need a fast 'highway'. Broadcom deals with the 'network drama' and it is just bit deliver to Apple chip/chiplet that they'll consume.

However, given the limited edge space on a die that PCI-e would get swapped out for UALink ( or UCI-e or whatever relatively very low energy per lane proprietary connector Apple chose). However, I doubt that will do anything for a Mac Studio. Apple is still shipping a product with 1GbE Ethernet connectors defined in the last century. Apple thinks the 10GbE on the studio is a 'big deal'. Apple is doing their own Wi-Fi chip so 'Wi-Fi 7' (and the next upgrade) will likely get labeled a 'big deal' in the Mac product line also. (Apple wifi is discrete so can just leave it off of a data center system) It probably will be another decade or two before they thought 100GbE was reasonable workstation speed. That's why a Mac Pro would be useful in the line up.

If Apple-Broadcom puts the datacenter I/O all off on another chip (or chiplet) then perhaps a Mac Pro could end up on a chip that has different I/O mix for Mac Pro versus studio. Mac Pro drops some thunderbolt and display out die space to subsume the PCI-e backhaul switch. ( Even back in the MP 2013 days 6 Thunderbolt ports seemed like 'overkill'. Still equally kind of goofy in 2026 Studio even with TBv5. Eight TB ports is generally goofy. There may be some extremely narrow corner cases where that might make sense, but not enough for an Apple minimal unit volume product. )

 

[drmeatball]

I have a 2023 Mac Pro, got it when Microcenter clearanced them last Christmas for about 50% off ($3599 - cheaper than an identical Mac Studio at the time). 64GB ram, but loaded with 24TB of NVMEs on PCIE cards (all 4TB WD SN850x model). I do content creation ranging from audio to video to game development (Unity).

I installed Tahoe to one of the 4TB drives, not even using the internal 1TB.

Love it!

 

[mr_roboto]

Apple isn't introducing Infiniband protocol support. All Apple is doing is "Infiniband verbs". Really software API calls to networking rather than the actual networking protocol. There is software libraries that layer on top of RDMA to get work done. Infniband is not == to RDMA , but at a more higher software abstraction layer folks have used the 'verbs' associated with the protocol as common ground. ( RDMA on converge Ethernet. RoCE). On Linux there is a 'infiniband verb' library that is a 'standard' library to bind/write to.

The killer IB app anyone wants IB for is RDMA, and IB Verbs is the API applications use to do IB RDMA. So... what exactly are you complaining about?

Yes, yes, you've found a technicality which gives you an excuse to rant about how Apple isn't doing real IB. But nobody actually cares about that. There's a deep history of doing IB-compatible communications over random interconnects which aren't one of the ones created by the Infiniband Trade Association (IBTA) specifically for IB. The Verbs API provides the abstraction required for this. Software that speaks IBV doesn't have to care whether it's running on top of IB, RoCE, or Thunderbolt.

Speaking of RoCE... it's a great example of how confused you are. If Apple's Thunderbolt RDMA isn't a form of IB, neither is RoCE. It's right there in the acronym - RoCE is RDMA over Converged Ethernet. ("Converged" being corporate IT buzzword speak for "we're running more protocols than just UDP/IP over our ethernet".)

Modern ConnectX connection will 'smoke' Apple's solution.

There already RDMA Ethernet cards for macOS.

https://www.atto.com/introducing-atto-technology-ethernet-suite/

What Apple is doing is more of a more affordable ( 4 Studios with point-to-point links to the other 3 nodes ) by avoiding switching and additional latency. But not fastest; not going to hit 100GbE speed let alone 200 or 400.

Nobody cares how bad you think something will 'smoke' something else while completely misunderstanding what this feature is for. There's demand out there for Mac Studio clusters, and the best IO a Studio has is Thunderbolt 5, and therefore the best way to connect such a cluster is TB5. If you can. Which you couldn't before 26.2, but now can.

Sure, prior to this you could attach a true IB or Ethernet NIC to each Studio. But they'd have to be PCIe NICs connected to the Studios through TB5. In such a configuration, TB5's raw performance would still be the theoretical upper limit, but you'd be losing performance to the overheads of the additional protocol stacks. For any cluster small enough to not need network switches, why not cut out as many middleman protocols as you can?

So it just doesn't matter whether there's faster IB or other Verbs RDMA solutions available on other platforms. If you have a need to specifically construct a Mac Studio cluster, this is the best way to do it. Apple's RDMA doesn't need to offer anything more than that to justify itself, no matter how many irrelevant nitpicks you invent.