WWDC 2023 Thread

[theorist9]

Not a good idea IMHO to lower the fan curve tho just to lower noise. I would think it is safe to do it the other way, i.e. turn it up.

This has been under discussion at Macs Fan Control's website. Mac's Fan Control said they tested it at 1100 RPM under various conditions, and the temps stayed low. They further remark that the MBP's run, by default, at 0 RPM under low-load conditions, so it seems odd this isn't available to the Studios, which have much larger heat sinks:

Mac Studio fan control and fan noise - CrystalIDEA Blog

crystalidea.com

However, various conditions isn't all conditions (which would include not just all use cases, but also all external temperatures--e.g., suppose your room is in the 80's). Thus there was a feature request, which they're working on, to run the Studio with either no fan, or at 1100 RPM, but revert to the OS's automatic control if the temps went above specified threshold values. That would seem to be the best of both worlds; though if you did this you'd probably want to have a separate, independent program running that will sound an alarm if the temps go over-threshold, in case the algorithm on Macs Fan Control fails.

Zero-RPM Speeds (Switch to 'Automatic' Control below lower Threshold) · Issue #582 · crystalidea/macs-fan-control

My M1 Max MBP 16" when in automatic mode (I.e. MFC not controlling the fans) sets the fan speed to zero RPM. MFC sets a minimum of 1500 RPM. Can a feature be implemented that allows the zero RPM mo...

github.com

 

[theorist9]

Max Tech compared the new M2 Ultra Mac Pro against the old Intel model.



Unsurprisingly, the new model makes the previous one look like an old, noisy beast. Except for a slight uptick during a Cinebench torture test, the Apple Silicon Mac Pro didn't ramp the fans and was functionally silent. Comparatively, if you've ever been anywhere within the proximity of an Intel Mac Pro, then anything mildly stressful turns it into a jet engine. Obviously, the new model should perform better than the previous one, but it does so silently, which isn't something the Intel era was known for. My puny 2018 Mac mini with a 4-core i3 makes more noise than any of the Apple Silicon Macs. For most users, the Mac Studio is probably a better option, but if you need slots, then the new Mac Pro should fill that niche.

I didn't watch the video, but some questions/points:

1) The M2 Ultras on the Studio and MP obviously have the same number of PCIe lanes. While the M2 Ultra doesn't have enough lanes to support the max possible bandwidth from the MP's internal SSD plus all the PCIe cards you can plug into it (not even close), I recall reading that the Ultra does have more than enough lanes to support the internal SSD plus the max bandwidth of all the external ports on the Studio. If that's correct (I don't know if it is), that would mean the MP offers more external bandwidth than the Studio (the idea being that, unlike the Studio, the MP can make use of all the M2 Ultra's PCIe bandwidth). It would be nice to get some numbers on this.

2) A music producer, Neil Parfitt, made the interesting point that the audio cards he uses have very noisy fans. Thus his MP needs to be in a separate room regardless, making the inherent quietness of the new MP irrelevant for his use case.

3) One way Apple could, in a very obvious way, distinguish the MP from the Studio, is to make the MP generationally upgradeable. E.g., allow the M2 Ultra MP to be upgraded to, say, an M4 Ultra or M4 Extreme MP. Then you don't have to junk the expensive case when you upgrade, making the additional $ you spent for the case be much more justifiable. Take a look at this pic from iFixit's teardown of the 2019 MP (https://www.ifixit.com/Teardown/Mac+Pro+2019+Teardown/128922). Granted, the AS version is somewhat different but, based on this, it doesn't seem like you''d need to replace much more than the motherboard to upgrade from an older AS version to a new one (assuming they stay with the same case design).

 

[exoticspice1]

Has Apple really given up on the scientific professionals? Steve Jobs in his keynotes hammer the Mac Pro as the best tool for these sort of industries.

To me it's looks like Apple is just mainly focusing on AV pros and a very niche area that is large rendering models thanks to the inherant advantage of having unified memory.

3D rendering that meets or beats Nvidia GPUs, Animation, AI training is what I hope Apples tackles next because it's sorely missing on the Mac. It would nice to see some benchmarks in regards to scientific areas, not just Video editing. Maybe that's what the new Mac Pro is capable of...

That's what I don't like about Mac reviewers, no substance. Just marketing.

 

[theorist9]

Has Apple really given up on the scientific professionals? Steve Jobs in his keynotes hammer the Mac Pro as the best tool for these sort of industries...

It would nice to see some benchmarks in regards to scientific areas, not just Video editing.

Not generally, but some time ago they did abandon the subset of researchers that do GPU computing using NVIDIA/CUDA.

But the Mac remains the only platform on which you can run fully-featured versions of MS Office and Adobe CC natively, and also have access to a native Unix terminal; this makes it very appealing for those who need both.

Apple Silicon does offer a Neural Engine, which is dedicated circuitry for accelerating ML/AI. However, it first appeared several years ago on the A11, where it clearly wasn't intended for ML/AI data science research projects. I don't know if the capabilities of the Neural Engine in the M-series chips have been expanded specifically with ML/AI researchers in mind. And I'll leave it to others to comment on how effective this is vs. what NVIDIA offers for such tasks.

I did post benchmarks with Mathematica. Generally, the performance with Apple Silicon has been disappointing. Comparing the M1 series with a 2019 i9 Intel iMac, it's sometimes slower for numerical work, and only modestly faster (~10%, on average) for symbolic tasks. I don't know the reason for this.

 

[theorist9]

Returning to the question of whether the MacPro makes sense—which we were discussing earlier in this thread—I've come across two new points:

One was from an interesting discussion I was having with poster joevt at MR about the difference in bandwidth between the M2 Ultra Studio and the M2 MP. I found this in Apple's documentation for the MP (https://support.apple.com/en-euro/HT213663), indicating it has 24 PCIe lanes left over after subtracting those used for the SSD:

"PCIe bandwidth​

The M2 Ultra chip provides 32 lanes of PCIe gen 4 to the system, with 8 lanes dedicated to the internal SSD. The M2 Ultra chip connects to the PCIe slots through a PCIe switch and provides 24 lanes of gen 4 bandwidth. Pool A provides a maximum of 16 lanes of gen 4 bandwidth and Pool B provides a maximum of 8 lanes of gen 4 bandwidth."

But then joevt noticed this:
"Each built-in Thunderbolt port in Mac Pro [sic] is managed by its own controller integrated in the M2 Ultra chip and doesn't share bandwidth with the PCIe slots."

The Ultra Studio has 6 built-in TB4 ports, while the MP has 8. The documentation says the MP uses PCIe for all the non-TB ports. Thus, comparing the two machines for I/O, we have:

M2 Ultra Studio: 6 x TB4 + Ethernet (10 Gb/s) + 2 x USB-A (2 x 5 Gb/s = 10 Gb/s)
M2 Mac Pro 8 x TB4 + 24 x PCIe4 (24 lanes x 15.754 Gb/s/lane = 378 Gb/s)

That's an enormous difference in I/O bandwidth. For instance, since the 8 x TB are separate, you could get, say, a HighPoint SSD7540 PCI-Express PCIe 4.0 x16 8-Port M.2 NVMe RAID Controller (claimed peak transfer speed of 28,000 MB/s = 224 Gb/s*, and still have more bandwidth left over than you'd get from an entire M2 Ultra Studio. [*Another vendor, OWC, claims a 26,926 MB/s peak speed for their PCIe x16 storage, which is in the same ballpark.]

[Still, part of me is puzzled by this: If there really were such a substantial difference, you'd think Apple would emphasize that in their marketing materials for the MP (i.e., not just the far greater variety of possible interfaces b/c of PCIe cards, but the far greater I/O bandwidth as well). Particularly since they've struggled to differentiate its capabilities from those of the Ultra Studio.]

The other point was made by a different poster at MR who mentioned he does audio production, and has ≈$30k in PCIe cards that he was happy to be able to transfer over to his new AS Mac Pro. Yes, you could use TB-attached PCIe boxes; two of these would hold just as many PCIe cards as the MP (not including the MP's I/O card) (two full-length plus four half-length). But they're $1,200 each, for a total of $2,400, and thus approach the $3k differential for the MP, while not being as clean a solution. Plus, if the above is correct, you lose I/O by using TB for this:

Echo III Rackmount System (3 PCIe slots for full size cards + Module support)

Rackmount Thunderbolt 3 adapter box with three PCIe expansion slots for audio, video, fibre, RAID and ethernet PCIe cards.

www.sonnetstore.com

 

[dada_dave]

Returning to the question of whether the MacPro makes sense—which we were discussing earlier in this thread—I've come across two new points:

One was from an interesting discussion I was having with poster joevt at MR about the difference in bandwidth between the M2 Ultra Studio and the M2 MP. I found this in Apple's documentation for the MP (https://support.apple.com/en-euro/HT213663), indicating it has 24 PCIe lanes left over after subtracting those used for the SSD:

"PCIe bandwidth​

The M2 Ultra chip provides 32 lanes of PCIe gen 4 to the system, with 8 lanes dedicated to the internal SSD. The M2 Ultra chip connects to the PCIe slots through a PCIe switch and provides 24 lanes of gen 4 bandwidth. Pool A provides a maximum of 16 lanes of gen 4 bandwidth and Pool B provides a maximum of 8 lanes of gen 4 bandwidth."

But then joevt noticed this:
"Each built-in Thunderbolt port in Mac Pro [sic] is managed by its own controller integrated in the M2 Ultra chip and doesn't share bandwidth with the PCIe slots."

The Ultra Studio has 6 built-in TB4 ports, while the MP has 8. The documentation says the MP uses PCIe for all the non-TB ports (other than HDMI, which is connected directly to the GPU). Thus, comparing the two machines for I/O, we have:

M2 Ultra Studio: 6 x TB4 + Ethernet (10 Gb/s) + 2 x USB-A (2 x 5 Gb/s = 10 Gb/s)
M2 Mac Pro 8 x TB4 + 24 x PCIe4 (24 lanes x 15.754 Gb/s/lane = 378 Gb/s)

That's an enormous difference in I/O bandwidth. For instance, since the 8 x TB are separate, you could get, say, a HighPoint SSD7540 PCI-Express PCIe 4.0 x16 8-Port M.2 NVMe RAID Controller (claimed peak transfer speed of 28,000 MB/s = 224 Gb/s*, and still have more bandwidth left over than you'd get from an entire M2 Ultra Studio. [*Another vendor, OWC, claims a 26,926 MB/s peak speed for their PCIe x16 storage, which is in the same ballpark.]

[Still, part of me is puzzled by this: If there really were such a substantial difference, you'd think Apple would emphasize that in their marketing materials for the MP (i.e., not just the far greater variety of possible interfaces b/c of PCIe cards, but the far greater I/O bandwidth as well). Particularly since they've struggled to differentiate its capabilities from those of the Ultra Studio.]

While they may have separated the Thunderbolt and PCIe lanes, they don't actually have all that bandwidth*:

View attachment 24184

The other point was made by a different poster at MR who mentioned he does audio production, and has ≈$30k in PCIe cards that he was happy to be able to transfer over to his new AS Mac Pro. Yes, you could use TB-attached PCIe boxes; two of these would hold just as many PCIe cards as the MP (not including the MP's I/O card) (two full-length plus four half-length). But they're $1,200 each, for a total of $2,400, and thus approach the $3k differential for the MP, while not being as clean a solution. Plus, if the above is correct, you lose I/O by using TB for this:

Echo III Rackmount System (3 PCIe slots for full size cards + Module support)

Rackmount Thunderbolt 3 adapter box with three PCIe expansion slots for audio, video, fibre, RAID and ethernet PCIe cards.

www.sonnetstore.com

That's the problem. It's a mirage of providing basically PCIe housing, but not the full bandwidth to supply it all. If you don't need all that bandwidth simultaneously, which likely you won't to be fair a lot of the time, maybe, depending on your use case, then you might never notice. But if you do want that, then you aren't getting it and, in my opinion, you're paying Apple like you do get all that extra bandwidth.

*correction I think I misread your post: they don’t have the bandwidth they claim for the slots they have but yes they do have 24 lanes of PCIe, some of which itself is taken up by default giving you I believe only 16 lanes of PCIe internally by default. Yes it does have the full 8 thunderbolt ports as opposed to 6 for some reason in the Studio, presumably product segmentation since the Studio having the same chip would be quite capable of supporting that - I guess it keeps the chassis body the same between the Max and Ultra which is a nice cost saving measure… for Apple.

 

[theorist9]

While they may have separated the Thunderbolt and PCIe lanes, they don't actually have all that bandwidth:

I remember Hector's post. I found his presentation a bit confusing, but I thought his complaint was that you can't count the number of slots to get the bandwidth, because PCIe lanes are shared among slots, which is true.

I.e., I thought his complaint was that the MP doesn't offer (2 x PCIe4 x16) + ( 4 x PCIe4 x8) + PCIe3 x4 (equiv. to PCIe4 x2) = 66 x PCIe4 lanes of bandwidth:

But even if the MP instead offers 24 extra PCIe4 lanes beyond what's available from the Ultra Studio, as I've inferred from Apple's documentation, that's still enormous—it would give the MP more than double the Ultra Studio's I/O.

Or if my 24-lane interpretation of Apple's documentation is incorrect, then what is the extra number of PCIe4 lanes of I/O bandwidth the MP offers beyond that available from the Ultra Studio?

 

[dada_dave]

I remember Hector's post. I found his presentation a bit confusing, but I thought his complaint was that you can't count the number of slots to get the bandwidth, because PCIe lanes are shared among slots, which is true.

I.e., I thought his complaint was that the MP doesn't offer (2 x PCIe4 x16) + ( 4 x PCIe4 x8) + PCIe3 x4 (equiv. to PCIe4 x2) = 66 x PCIe4 lanes of bandwidth:
View attachment 28011

But even if the MP instead offers 24 extra PCIe4 lanes beyond what's available from the Ultra Studio, as I've inferred from Apple's documentation, that's still enormous.

Or if my 24-lane interpretation of Apple's documentation is incorrect, then what is the extra number of PCIe4 lanes of I/O bandwidth the MP offers beyond that available from the Ultra Studio?

Aye if you read my edit, I misunderstood your post initially.

It’s not that enormous though for the use cases that the Pro is intended for. That’s basically 1 full length card that could take up the available bandwidth with another half card if you take out Apple’s IO. Running all those cards simultaneously was supposed to be the wheelhouse of the Pro. And it can’t do it.

Keep in mind the silicon is the same, you’re spending 3K on just the body. For some, maybe worth it! But for a much narrower audience than before. And that worries me for its long term prospects. The user has to have a large number of internal cards that don’t actually use up that much bandwidth or aren’t expected to be run at the same time.

 

[theorist9]

Aye if you read my edit, I misunderstood your post initially.

It’s not that enormous though for the use cases that the Pro is intended for. That’s basically 1 full length card that could take up the available bandwidth with another half card if you take out Apple’s IO. Running all those cards simultaneously was supposed to be the wheelhouse of the Pro. And it can’t do it.

Keep in mind the silicon is the same, you’re spending 3K on just the body. For some, maybe worth it! But for a much narrower audience than before. And that worries me for its long term prospects. The user has to have a large number of internal cards that don’t actually use up that much bandwidth or aren’t expected to be run at the same time.

*correction I think I misread your post: they don’t have the bandwidth they claim for the slots they have but yes they do have 24 lanes of PCIe, some of which itself is taken up by default giving you I believe only 16 lanes of PCIe internally by default. Yes it does have the full 8 thunderbolt ports as opposed to 6 for some reason in the Studio, presumably product segmentation since the Studio having the same chip would be quite capable of supporting that - I guess it keeps the chassis body the same between the Max and Ultra which is a nice cost saving measure… for Apple.

It's 32 lanes, 8 of which are taken up internally by the SSD, giving you 24 lanes of PCIe I/O (some of which can be used for additional internal connections, like a SATA drive, but logically that counts as added I/O, beyond what the TB ports supply, just as much as that provided by the PCIe slots, since they only differ in where they are located inside the machine).

While I understand the MP's I/O can be considered disapponting, the fact that you can't run all 68 lanes of the available PCIe slots at full bandwidth is a bit of a red herring when comparing the MP's capabilities to that of the Ultra Studio, since you're still talking more than twice the I/O bandwidth, which should not be trivialized. I.e., both of these statements are true:

1) As a workstation-class machine, the AS MP's I/O, like its maximum RAM, falls short when compared with what PC workstations offer, and also compared to what would be expected from the sucessor to the Intel MP (I think this is what you're saying, except I added the part about the RAM).

2) In assessing whether the AS MP provides sufficient added value over the Studio to make it worth the $3k, it offers not only six PCIe slots (that would cost ≈$2,400 to accommodate with professional-class TB-attached boxes), but also over twice the I/O bandwidth. This makes it a good value for anyone that uses PCIe cards, and needs significantly more bandwidth than the Ultra Studio provides (this is what I'm saying). I.e., I wanted to counter the CW that the machine is pointless. It's niche, and perhaps disappointing compared to the expectations that were out there for an MP "Extreme", but not pointless.

And their marketing of this capability seems strangely reticent. On their marketing page for the MP they say "twice the bandwidth", leaving out of what (i.e., of the Ultra Studio), which surely was a deliberate decision on their part (given how much attention they pay to marketing). Is it because they don't want to diss the Ultra Studio (which is probably a much bigger seller than the MP)?

 

[leman]

Well, it’s effectively a single 16x slot worth of bandwidth (pool A). More than what a standard consumer CPU gives you (since 16x is taken by a GPU), but laughably little compared to workstation systems. I don’t think Apple misses their chance in advertising the additional I/O capabilities of the tower, PCI expandability is pretty much the only reason for its existence, and their marketing reflects it fairly well IMO. In the end, it’s a fancy expensive enclosure for people that either have to work with large (and very fast) storage arrays, or have a lot of specialized cards that don’t need too much aggregated bandwidth. In other words, it’s probably limited to a certain class of video/audio professional.

 

[theorist9]

Well, it’s effectively a single 16x slot worth of bandwidth (pool A). More than what a standard consumer CPU gives you (since 16x is taken by a GPU), but laughably little compared to workstation systems. I don’t think Apple misses their chance in advertising the additional I/O capabilities of the tower, PCI expandability is pretty much the only reason for its existence, and their marketing reflects it fairly well IMO. In the end, it’s a fancy expensive enclosure for people that either have to work with large (and very fast) storage arrays, or have a lot of specialized cards that don’t need too much aggregated bandwidth. In other words, it’s probably limited to a certain class of video/audio professional.

How do you get just an extra 16x from the difference between these two?
M2 Ultra Studio: 6 x TB4 + Ethernet (10 Gb/s) + 2 x USB-A (2 x 5 Gb/s)
M2 Mac Pro 8 x TB4 + 24 x PCIe4

 

[dada_dave]

It's 32 lanes, 8 of which are taken up internally by the SSD, giving you 24 lanes of PCIe I/O (some of which can be used for additional internal connections, like a SATA drive, but logically that counts as added I/O, beyond what the TB ports supply, just as much as that provided by the PCIe slots, since they only differ in where they are located inside the machine).

While I understand the MP's I/O can be considered disapponting, the fact that you can't run all 68 lanes of the available PCIe slots at full bandwidth is a bit of a red herring when comparing the MP's capabilities to that of the Ultra Studio, since you're still talking more than twice the I/O bandwidth, which should not be trivialized. I.e., both of these statements are true:

1) As a workstation-class machine, the AS MP's I/O, like its maximum RAM, falls short when compared with what PC workstations offer, and also compared to what would be expected from the sucessor to the Intel MP (I think this is what you're saying, except I added the part about the RAM).

2) In assessing whether the AS MP provides sufficient added value over the Studio to make it worth the $3k, it offers not only six PCIe slots (that would cost ≈$2,400 to accommodate with professional-class TB-attached boxes), but also over twice the I/O bandwidth. This makes it a good value for anyone that uses PCIe cards, and needs significantly more bandwidth than the Ultra Studio provides (this is what I'm saying). I.e., I wanted to counter the CW that the machine is pointless. It's niche, and perhaps disappointing compared to the expectations that were out there for an MP "Extreme", but not pointless.

And their marketing of this capability seems strangely reticent. On their marketing page for the MP they say "twice the bandwidth", leaving out of what (i.e., of the Ultra Studio), which surely was a deliberate decision on their part (given how much attention they pay to marketing). Is it because they don't want to diss the Ultra Studio (which is probably a much bigger seller than the MP)?

View attachment 28014
View attachment 28015

How do you get just an extra 16x from the difference between these two?
M2 Ultra Studio: 6 x TB4 + Ethernet (10 Gb/s) + 2 x USB-A (2 x 5 Gb/s)
M2 Mac Pro 8 x TB4 + 24 x PCIe4

The M2 Ultra has 32 lanes and 8 TB4 controllers regardless of whether it is in a Studio enclosure vs Pro enclosure. x8 are taken by the SSD in both. The reason @leman and I say it's really an extra 16x is the by x8 from pool B is by default already used up for the most part, some of it builtin. It includes things like USB, WiFi, Ethernet, etc ... So there is really only 16x lanes of PCie free to actually use for a card of your choice. Again, that's effectively 1 full card's worth of bandwidth. That's nothing for the use case it was designed for and the Ultra Studio already has the expensive part of that builtin (the silicon with the PCIe lanes and thunderbolt controllers). The part the Mac Pro adds is by far the cheaper part.

The major reason TB4 enclosures cost so much is they have to pay for extra thunderbolt controllers on the receiving end. That's expensive (less than it used to be but still a lot) and much, much more expensive than adding physical PCIe slots which are dirt cheap to add (by comparison), and yet as you note 2 thunderbolt enclosures are still less expensive than buying the Mac Pro. If we're going by value, the Mac Pro should be *less* not *more* than two thunderbolt enclosures plus Studio, especially considering the fact that those PCIe slots aren't "real". You're housing all the extra cards in slots but you aren't actually getting the bandwidth, so saying you have all these slots is the red herring and Apple is charging you like you are getting real slots with actual bandwidth. You aren't.

The Studio for its form factor and price is decent for connectivity. The Mac Pro is not and in my opinion not worth it for just getting a single slot's worth of internal connectivity and two extra TB slots. That's not 3K except for a very limited audience.

 

[theorist9]

x8 from pool B is by default already used up for the most part, some of it builtin. It includes things like USB, WiFi, Ethernet, etc ...

I'm afraid I'm not following how it's "used up for the most part". The bandwidth consumed by USB/WiFi/Ethernet/etc should be only a small fraction of the bandwidth provided by the x8 in pool B, leaving the ovewhelming majority of it available. I'd estimate all that stuff, collectively, consumes no more than the bandwidth of about 1 TB port*. Then the MP exceeds the Studio by about PCIe4 x24 + 1** x TB4.

*To avoid misinterpretation: I know it does't use the TB controllers for this. I'm just estimating the total bandwidth consumed by these is, together, no more than about 1 TB port's worth.
**From (8–6)–1

especially considering the fact that those PCIe slots aren't "real". You're housing all the extra cards in slots but you aren't actually getting the bandwidth, so saying you have all these slots is the red herring and Apple is charging you like you are getting real slots with actual bandwidth. You aren't.

Of course they're real. They represent added bandwidth not available from the Studio. They'd only not be "real" if they represented no additional bandwidth, i.e., if their bandwidth were coming from the TB buses. But as Apple says: "Each built-in Thunderbolt port in Mac Pro [sic] is managed by its own controller integrated in the M2 Ultra chip and doesn't share bandwidth with the PCIe slots." You were saying the added bandwidth isn't as much as you think this class of machine should provide. I don't disagree, but that's entirely different from not being real.

If your position is that every slot must have fully dedicated PCIe lanes for the PCIe bandwidth to be considered real, then the Intel MP's PCIe bandwidth wouldn't be real either, since it also pooled bandwidth:

Mac Pro (2019 - Intel): Limits and possibilities of PCIe slots

The 2019 Mac Pro allows to connect PCIe cards. Apple gives some recommendations on how to install and configure them. But below are some more recommendations. How to open the "Expansion Slot Util...

softron.zendesk.com

 

[dada_dave]

I'm afraid I'm not following how it's "used up for the most part". The bandwidth consumed by USB/WiFi/Ethernet/etc should be only a small fraction of the bandwidth provided by the x8 in pool B, leaving the ovewhelming majority of it available. I'd estimate all that stuff, collectively, consumes no more than the bandwidth of about 1 TB port*. Then the MP exceeds the Studio by about PCIe4 x24 + 1 x TB4.

*To avoid misinterpretation: I know it does't use the TB controllers for this. I'm just estimating the total bandwidth consumed by these is, together, roughly equivalent to that of 1 TB port.

88% of the bandwidth is used up according to Apple. It's in the manual you provided. Even if that weren't the case it would be 1 x 16 slot and 1x 8 slot. Which is still very, very little.

Of course they're real. They represent added bandwidth not available from the Studio. They'd only not be "real" if they represented no additional bandwidth, i.e., if their bandwidth were coming from the TB buses. But as Apple says: "Each built-in Thunderbolt port in Mac Pro [sic] is managed by its own controller integrated in the M2 Ultra chip and doesn't share bandwidth with the PCIe slots." You're saying the added bandwidth isn't as much as you think this class of machine should provide. I don't disagree, but that's entirely different from not being real.

If the position is that every slot must have dedicated PCIe lanes for the PCIe bandwidth to be considered real, then the Intel MP's PCIe bandwidth wouldn't be real either, since it also pooled bandwidth:

Mac Pro (2019 - Intel): Limits and possibilities of PCIe slots

The 2019 Mac Pro allows to connect PCIe cards. Apple gives some recommendations on how to install and configure them. But below are some more recommendations. How to open the "Expansion Slot Util...

softron.zendesk.com

Indeed and some of them weren't! But the Mac Pro currently truly only has 1 x 16 slot and 1 x 8 slot most of the latter of which is taken up by default. Additional slots on top of that exist physically but you cannot run them at spec, not together. That's what I mean by not "real". So arguing that it's worth it for all these PCIe slots isn't a valid argument in my book. It's only actually got the 1x16 and the 1x8. Further, yes, compared to how much internal expansion they had before, it is less than what used to be offered (4x16!), substantially so.

That said, truthfully the switch to AS means that there are fewer high bandwidth cards actually available for the Mac, so maybe it's the form factor itself that is obsolete. I hope not. I think an Extreme chip could justify its existence. But Apple priced the Ultra Mac Pro way too high for what it has to offer. Lower it by $1500 and one could argue its high but maybe worth it (I'm just pulling the number out of thin air). At $3K over the Studio only the most niche of the niche use cases can justify its purchase. That's not a healthy user base to maintain a product line. That's what concerns me. It feels like its been priced to justify killing it off in the next couple of generations "because so few buy it anyway" - well, yeah it isn't a compelling purchase for most people who might otherwise be interested in it ...

In summary: it basically has identical computational power to the Studio with a small amount of internal PCIe expansion for 1.75x the price. Again, will some find value? Sure and I’m not saying they’re “wrong” or “bad” for making the purchase. I’m saying it’s a smaller group than it should be and I fear smaller than is sustainable.

 

[leman]

How do you get just an extra 16x from the difference between these two?
M2 Ultra Studio: 6 x TB4 + Ethernet (10 Gb/s) + 2 x USB-A (2 x 5 Gb/s)
M2 Mac Pro 8 x TB4 + 24 x PCIe4

I think the big question is how much aggregated bandwidth is offered by the TB4 ports? Has this been clarified somewhere (sorry if I missed it).

My understanding of all this that the 8 ports of the tower don’t actually offer more bandwidth than those of the Studio, and the internal I/O is connected from the 8x pool. Which would leave 16x as the primary difference. Am I wrong about this?

 

[dada_dave]

I think the big question is how much aggregated bandwidth is offered by the TB4 ports? Has this been clarified somewhere (sorry if I missed it).

My understanding of all this that the 8 ports of the tower don’t actually offer more bandwidth than those of the Studio, and the internal I/O is connected from the 8x pool. Which would leave 16x as the primary difference. Am I wrong about this?

I do believe you are getting a full extra 2 TB ports as the M2 Ultra has 8 TB controllers on top of the PCIe. The reason the Studio doesn’t have them I think is to maintain physical consistency with the M2 Max version of the Studio (puts two 10Gb/s USB-C on the front + 4 TB4 in the back replaced with 2+4 full TB4 in the Ultra Studio) and of course product differentiation.

To recap PCIe in the Pro, M2 Ultra offers 32 slots of PCIe total, 8 of which is for the SSD. Of the remaining 24, it’s split into pool A (16) and pool B (8). Most of pool B’s 8 slots is also taken up by default but some of that may be Apple’s technically removable IO card in Slot 7 (4 slots). But some of that default Pool B can be used for additional storage.

So yes the primary difference is that internal pool of free 16 slots with maybe a second pool of free 1/4/8 slots depending on how generous you’re feeling and 2 extra TB ports for 1.75x the price.

 

[leman]

I do believe you are getting a full extra 2 TB ports as the M2 Ultra has 8 TB controllers on top of the PCIe.

I thought it was 3 TB controllers per Max? Where does one extra controller come from?

Also, even if those are full-featured controllers, does it really mean that the aggregated bandwidth adds up. Eight port adds up to what, 40GB/s bidirectional bandwidth? That's quite a lot!

 

[dada_dave]

I thought it was 3 TB controllers per Max? Where does one extra controller come from?

Also, even if those are full-featured controllers, does it really mean that the aggregated bandwidth adds up. Eight port adds up to what, 40GB/s bidirectional bandwidth? That's quite a lot!

I thought it was 4. Maybe I’m wrong.

Edit:

Hector Martin (@marcan@treehouse.systems)

Attached: 1 image So the weird thing is I don't see the extra PCIe controllers on the M2 Max die. I see 4 x4 lane blocks and 4 Thunderbolt blocks. Two of the x4 blocks would be for storage (8 NAND chips, x1 each). One of them is the regular APCIe which is used for WiFi/USB/SATA/Ethernet/etc (4...

social.treehouse.systems

I think it’s 4 unless I misunderstood the above. I believe they all get the full bandwidth but again maybe I’ve misunderstood something here. The 8 TB4 ports are roughly worth 20x of Gen 4 equivalent if I’ve done my calculations correctly. I believe TB4 is 40Gbs bidirectional not GBs.

 

[leman]

I thought it was 4. Maybe I’m wrong.

The laptop has only three ports. Wouldn’t it be weird to include four controllers but only use three for the most sold products?

I believe TB4 is 40Gbs bidirectional not GBs.

Eight ports in total would make it 40 GB/s

 

[dada_dave]

The laptop has only three ports. Wouldn’t it be weird to include four controllers but only use three for the most sold products?

yeah I dunno. The Max Studio has 4 TB4 ports and Hector said 4 controllers.

Eight ports in total would make it 40 GB/s

Which is 20x lanes of PCIe and … oh … yes I see what you wrote now more clearly, ooops