I keep asking myself "why a Mac Pro in the Apple Silicon age"?

[Joelist]

I have to admit that, to me, it doesn't really make sense for there to be a new Mac Pro or even for the current model to continue. This is why I believe this (and I could easily be wrong):

1) Apple Silicon is not built around the idea of modular components. It is an integrated SOC and leverages SOC properties (like being physically close - indeed on die) to increase efficiency and performance. The closest thing we've seen to modular has been the Ultra, and that is simply two Max SOCs stitched together.

2) Are we actually seeing scenarios where an M2 Ultra is inadequate (honest question I don't know)? If yes they have to be pretty corner case in nature because even the current Max SOCs are monsters performance wise and assuming the M2 Ultra scales better it will be something to behold. I guess this leads to...

3) What does a Pro do that a Studio can't? Before they rolled out the Studios I understood the case more but the Studio with its monster cooling system and large port count plus the Ultra seems to sit in a similar spot in the lineup.

4) With the reliance Apple has on their SOCs getting good yields on large runs I am having problems understanding why they would commit to very specific custom chips that inherently would have small runs. Consider that the same fab run for the Max supplies the MacBook Pro and the Studio (both Max and Ultra since an Ultra is two Maxxes). In the same vein the run for the Pro does MacBook Pros and Mac Minis. And so on.

I guess that to me IF there were a "Pro" it would take the form of a chassis that held two Ultras stitched together (essentially 4 Maxxes).

 

[Citysnaps]

There are commercial/industrial/laboratory/aerospace/etc. applications where a computer with a general purpose bus and slots, such as PCIe, are needed, similar to the current Mac Pro (in both floor standing and rack mount versions).

If Apple wishes to continue appealing to those markets then an updated Mac Pro would be necessary.

 

[B01L]

Well, there's the whole PCIe slot thing; said slots are used for more than just discrete GPUs, and there are add-in cards that want more bandwidth than TB4 can provide...

 

[Joelist]

But again what are these specialist cards they need?

 

[B01L]

Networking cards...
Fiber channel cards...
Audio I/O cards...
Audio DSP cards...
Video I/O cards...
Raid cards...

And those are basically the general purpose cards, not even any specialty cards...!

 

[Citysnaps]

In addition to what was written above, here's a few I've been noodling over the years.

High speed, high dynamic range data acquisition (analog and digital) cards,

Special purpose high speed signal processing cards (real time FFTs, correlators, digital filtering, etc)

Digital radio for defense/aerospace... ie, multichannel SDR - software defined radios (a term I hate, as processing is done in ASICs or FPGAs, not in software)

Large high speed bulk memory and controllers to capture RF spectrums via A/Ds sampling a wideband RF/IF for signal exploitation/recognition purposes, hypothesis testing, etc.

Various image/signal acquisition/processing/acceleration and analysis applications from medical research to defense/aerospace applications.

Industrial control along with data acquisition for factories.


Some of the above are off the beaten path, but I view an updated rack mount MacPro with slots and a high speed network interface for control as a blank piece of paper with great potential.

 

[leman]

There is always demand for higher performance computing. One big market that Apple seems to be targeting explicitly is media creation, where they have the ambition to position Apple Silicon as the best hardware for production raytracing and video editing. And of course any domain that needs fast and flexible CPU performance (software development, engineering, sciences etc.).

On the other hand it's about brand perception. It is probably in Apple's interests to deliver something impressive for the high performance market even if they won't make money from it (or maybe even take a loss), as it will reinforce Mac reputation as powerful hardware. From financial perspective alone it makes sense to forego desktop altogether and just focus on ultra-compact laptops. That's where the money is and it's also the easiest, as one doesn't have to worry that much about scaling and high-performance applications. But this strategy will probably lead to inevitable decline and ultimately a demise of the Mac. Right now, a big reason for Mac popularity is the developer community. Alienate those people and you are only left with consumers.

What's important is that the current Apple Silicon is not nearly enough to command respect in the high performance desktop market. M1 Ultra is not any faster than any enthusiast class x86 CPU — and slower in the GPU department — while coming at a significant premium. Its selling point is lower power consumption as well as a compact chassis, which is not something many people in this market segment care that much about. Apple needs to ship something that would take the wind out of the power-hungry x86 performance-wise, as a true competitor to workstation chips like Xeon Max or Threadripper.

Of course, it is still possible that Apple decides to focus on the consumer and laptop market and forego high-performance applications. However, I do not believe that this will happen. They have been investing substantial amount of effort there recently: raytracing, machine learning, scientific computing (steadily building up their vector/matrix coprocessor as well as revamping the work on numerical libraries that have been self stale for many years). It would be weird for Apple to spend all that engineering time and money on these things if they planned to just drop this market.

 

[Citysnaps]

Also... Apple has already invested a lot of $ in the two current MacPro chassis (floor standing and rack mount versions).

There's no need to change those as they're well-engineered and exude quality; especially the rack mount version. A daughter card bus with options to upgrade AS processors over time would be nice. The current 1.4 kWatt power supply might be able to be downsized due to AS.

 

[leman]

Also... Apple has already invested a lot of $ in the two current MacPro chassis (floor standing and rack mount versions).

There's no need to change those as they're well-engineered and exude quality; especially the rack mount version. A daughter card bus with options to upgrade AS processors over time would be nice. The current 1.4 kWatt power supply might be able to be downsized due to AS.

About that: they just re-published the 2019 chassis patent, it doesn't seem like there are any changes to the text: https://patentscope.wipo.int/search/en/detail.jsf?docId=EP391888150&_cid=P22-LFB9H7-42754-3

Question to @cmeier: why would one re-publish a patent like that?

 

[Cmaier]

About that: they just re-published the 2019 chassis patent, it doesn't seem like there are any changes to the text: https://patentscope.wipo.int/search/en/detail.jsf?docId=EP391888150&_cid=P22-LFB9H7-42754-3

Question to @cmeier: why would one re-publish a patent like that?

Re-publishing isn’t a thing. Your link goes to a european patent, filed in 2020. It’s an entirely new patent, technically. The difference will be in the list of claims, not in the “text” (which we call the “patent specification.”). It ”claims priority” to the 2019 patent application.

The way the patent system works is that your patent application includes a specification and claims. At any point before your application issues as a granted patent, you can file what’s called a “continuation” of your application, which may eventually issue as a different patent. There are three types of continuations: continuations, continuations-in-part, and divisionals.

In a normal continuation or divisional, the text of the specification has to be identical (other than typo fixes or the like). The claims, however, must be different. In these types of continuations, the priority date of your patent (the date before which prior art must have been published in order to invalidate your claims) is the priority date of the original application.

In a continuation-in-part, you can add stuff to your specification. However, the priority date will be determined on a claim-by-claim basis. The date will be the date on which the material in the claim was first added to an application in the chain of continuations-in-part.

There are a couple of other things that can cause new versions of patents - reissues and reexaminations, but that’s not what’s going on here.

 

[leman]

Re-publishing isn’t a thing. Your link goes to a european patent, filed in 2020. It’s an entirely new patent, technically. The difference will be in the list of claims, not in the “text” (which we call the “patent specification.”). It ”claims priority” to the 2019 patent application.

The way the patent system works is that your patent application includes a specification and claims. At any point before your application issues as a granted patent, you can file what’s called a “continuation” of your application, which may eventually issue as a different patent. There are three types of continuations: continuations, continuations-in-part, and divisionals.

In a normal continuation or divisional, the text of the specification has to be identical (other than typo fixes or the like). The claims, however, must be different. In these types of continuations, the priority date of your patent (the date before which prior art must have been published in order to invalidate your claims) is the priority date of the original application.

In a continuation-in-part, you can add stuff to your specification. However, the priority date will be determined on a claim-by-claim basis. The date will be the date on which the material in the claim was first added to an application in the chain of continuations-in-part.

There are a couple of other things that can cause new versions of patents - reissues and reexaminations, but that’s not what’s going on here.

Thanks, very informative!

 

[Cmaier]

Thanks, very informative!

I should also point out that your expiration date, for a continuation, is based on the date of the oldest application. So you can’t use this as a hack to add more years.

 

[Joelist]

Networking cards...
Fiber channel cards...
Audio I/O cards...
Audio DSP cards...
Video I/O cards...
Raid cards...

And those are basically the general purpose cards, not even any specialty cards...!

Okay.

Networking is already built in. So are Audio and Video (in fact AS has specific hardware video accelerators - basically every M series except for the original M1 has built in Afterburner). Raid cards are interesting but again the storage you can get on card is already pretty huge.

M2Max is already as fast or faster than basically everything in the Intel and AMD lineups in CPU terms and GPU it stands toe to toe with top end discrete GPUs. All this at a small fraction of the heat and power - which are important because it means no need for elaborate cooling systems and the like.

Something else that comes to mind is Apple Silicon's memory bandwidth (400 GB/s). Any expansion technology has to at minimum support that. Otherwise performance will drop.

So, assuming Apple has a way for expansion not to kill performance (because of increased physical distance and potentially lower bandwidth), possibly a "Pro" could take the form of adding logic boards?

Suppose the base model has M2 Ultra with RAM and storage and other goodies on a logic board. What you can buy is more of that same board and slot them in and they would all work together. And when M3 comes out if you want you just pull out the M2 boards and insert M3s?

 

[dada_dave]

Okay.

Networking is already built in. So are Audio and Video (in fact AS has specific hardware video accelerators - basically every M series except for the original M1 has built in Afterburner).

Different types of cards. More variety, cards designed to interoperate with specific equipment, basically everything on @Citysnaps list. Apple’s on chip accelerators are indeed fantastic for their use cases. Other people have more use cases and there’s a limit to what Apple can provide. PCIe modularity bridges that gap.

Raid cards are interesting but again the storage you can get on card is already pretty huge.

8TB is far from huge for a lot of use cases.

M2Max is already as fast or faster than basically everything in the Intel and AMD lineups in CPU terms and GPU it stands toe to toe with top end discrete GPUs. All this at a small fraction of the heat and power - which are important because it means no need for elaborate cooling systems and the like.

Simply not true. M1 Ultra and M2Max are fantastic chips, but @leman is talking about CPU/GPU performance levels beyond that. It does not require dGPUs but it does require bigger, more powerful chips than we’ve seen so far.

Something else that comes to mind is Apple Silicon's memory bandwidth (400 GB/s). Any expansion technology has to at minimum support that. Otherwise performance will drop.

That’s not how things work. Even the CPU doesn’t access all that bandwidth and the kinds of stuff we’re talking about is rarely if ever bottlenecked by PCIe. They can be bottlenecked by thunderbolt but as I understand for most of these cards it’s about decreased cost/added convenience of internal PCIe connection relative to thunderbolt rather than an issue of bandwidth or latency.

So, assuming Apple has a way for expansion not to kill performance (because of increased physical distance and potentially lower bandwidth), possibly a "Pro" could take the form of adding logic boards?

Suppose the base model has M2 Ultra with RAM and storage and other goodies on a logic board. What you can buy is more of that same board and slot them in and they would all work together. And when M3 comes out if you want you just pull out the M2 boards and insert M3s?

I believe people have indeed made suggestions similar to this on this forum for various types of expanded performance but as said this is not just about dGPUs.

As @leman said Apple forgo this market but it tends to be lucrative on a per unit basis and it tends to have a halo effect.

 

[Nycturne]

Networking is already built in. So are Audio and Video (in fact AS has specific hardware video accelerators - basically every M series except for the original M1 has built in Afterburner). Raid cards are interesting but again the storage you can get on card is already pretty huge.

Audio/Video that’s built in isn’t what pros use. Granted, someone using a MBP for audio can use a USB or Thunderbolt audio interface for the purpose, but in a studio, being able to to build it in as a PCIe card and rack mount the machine it is in with the rest of the equipment is pretty standard.

When we talk about video, it’s not always acceleration, but input and output and being able to deal with multiple streams at the same time for capture, compositing and output. Something that a multi-port PCIe card is built for. See above for PCIe vs external.

Networking? Let me know when Apple offers SFP out of the box to support fiber connectivity. Again, see above for PCIe vs external.

And as for RAID, it’s not always about sheer storage (although being able to include a 4 drive RAID inside the machine is a boon for those still using spinning platters like the Pegasus R4i allows). PCIe NVME RAID is getting pretty gnarly, and it isn’t hard to saturate a 16 lanes of PCIe 4.0 these days using NVME. On that front, you can start seeing bandwidth needs that are on the order of around 8x what Thunderbolt is capable of delivering. We’re starting to see NVME RAID cards get over 30GB/s (versus the 5GB/s of the 4x PCIe 3.0 lanes Thunderbolt provides, which is generally less in practice). And that’s Gigabytes/sec, rather than Gigabits/sec.

So, assuming Apple has a way for expansion not to kill performance (because of increased physical distance and potentially lower bandwidth), possibly a "Pro" could take the form of adding logic boards?

As already pointed out, this isn’t not how it works. Ultimately, the key thing would be providing the PCIe lanes required from the SoC out to the slots, and handling muxing of slots to share PCIe lanes as needed, same as any Intel/AMD system. Apple already does include a few PCIe lanes off the SoC for a couple peripherals like onboard Ethernet (although I think the PCIe for Thunderbolt is muxed before it leaves the SoC). So it’d be a matter of scaling that up.

But it doesn’t require Apple add any tech they don’t already support. They already have PCIe 4 in their SoCs (which is odd when almost nothing they hook up to their PCIe lanes needs PCIe 4), and I wouldn’t be surprised to see a Mac Pro that makes PCIe lanes available for cards.

 

[Citysnaps]

Here's how Neil Parfitt (composer, producer, sound designer) uses PCIe cards in his rack mount MacPro:

 

[Citysnaps]

And for those new here, here's a vid of Parfitt unboxing and setting up his rack mount MacPro that I posted last year. Hopefully Apple is talking to him as he has a few issues that bug him about the computer.

 

[Colstan]

Here's how Neil Parfitt (composer, producer, sound designer) uses PCIe cards in his rack mount MacPro:

Hush, you're ruining the narrative. According to the forum mavens at MacRumors, Apple doesn't care about the needs of professionals and only designs iToys and releases new emojis.

And for those new here, here's a vid of Parfitt unboxing and setting up his rack mount MacPro that I posted last year. Hopefully Apple is talking to him as he has a few issues that bug him about the computer.

This doesn't make anymore more sense than the last time you posted it. In fact, it makes even less sense. The all-knowing forum members over at MacRumors have told us that the only use for PCIe slots are for third-party graphics cards purchased directly from Newegg, in a midnight fire sale, preferably during a full moon. If the Apple Silicon Mac Pro doesn't feature heavily optimized drivers for CUDA, which ideally only works inside of Boot Camp for Arm, then Apple is doomed and Tim Cook needs to be fired.

Also, the new System Settings in macOS Ventura is terrible and Tim Cook needs to be fired.

 

[Citysnaps]

Hush, you're ruining the narrative. According to the forum mavens at MacRumors, Apple doesn't care about the needs of professionals and only designs iToys and releases new emojis.

This doesn't make anymore more sense than the last time you posted it. In fact, it makes even less sense. The all-knowing forum members over at MacRumors have told us that the only use for PCIe slots are for third-party graphics cards purchased directly from Newegg, in a midnight fire sale, preferably during a full moon. If the Apple Silicon Mac Pro doesn't feature heavily optimized drivers for CUDA, which ideally only works inside of Boot Camp for Arm, then Apple is doomed and Tim Cook needs to be fired.

Also, the new System Settings in macOS Ventura is terrible and Tim Cook needs to be fired.

I just hope Apple doesn’t give up on producing a computer with a bus and slots. I’m out of the biz now, and as much as it’s lust-worthy appreciating extremely well-designed and built equipment, and the possibilities of how it can be used, I really have no need a MacPro computer (a Mac Studio meets my needs). But I still recognize its utility across a wide range of disciplines and possibilities in commercial and aerospace applications. Having designed systems in the past using both PC-based rack mount computers, and repurposed Macs for rack mounting, I’d much rather start with a Mac.

 

[throAU]

Whilst the market segment will likely continue to shrink - there is always a market for more powerful hardware.

The use cases may not be there yet but without the hardware there will never be the next paradigm shift to demand it.

Higher end hardware raises the bar for what is possible. Just because most of us don’t need it, doesn’t make it pointless.