Could we see M3 before A17?

[leman]

I know you've talked about horizontal scaling before but I just don't buy it. The Mac Pro isn't a server. Any Apple Silicon Mac Pro will be used to run software similar to what your Max or Ultra chip will run but users expect it to run much faster or run much bigger models. If you scale horizontally, software won't run faster unless they're rewritten from the ground up to take advantage of horizontal scaling - which very few software can make use of.

There is very little financial incentive for Apple to rearchitect most of macOS and rewrite most of their software to use horizontal scaling just for a niche machine that is already dying in relevance.

Hence, I'm still on the side of 4x Max dies glued together for a Mac Pro chip. This chip won't be cheap in R&D cost and it won't be cheap to manufacture. This is why I'm speculating that Apple will create an Apple Silicon Cloud to expand the market of any "Extreme" chip.

Between these options: cancel the Mac Pro, use only the Ultra, glue 4x max together, and horizontally scale Ultra chips, I think the horizontal scaling is the least likely option.

Gluing together four dies is also horizontal scaling, it’s just more software-friendly because you can have faster die-to-die communication. The largest issue with separate compute boards is the high synchronization overhead. But if this can be solved (e.g. with a proprietary low-latency high-bandwidth bridge), then the distinction between the two approaches disappear. But building such a bridge is far from trivial, if it’s even possible.

Vertical scaling with large SoCs sounds like the most reasonable approach to me.

 

[senttoschool]

Gluing together four dies is also horizontal scaling, it’s just more software-friendly because you can have faster die-to-die communication. The largest issue with separate compute boards is the high synchronization overhead. But if this can be solved (e.g. with a proprietary low-latency high-bandwidth bridge), then the distinction between the two approaches disappear. But building such a bridge is far from trivial, if it’s even possible.

Vertical scaling with large SoCs sounds like the most reasonable approach to me.

It isn't the same. What we want is to present software with one CPU, one GPU, one Neural Engine, one decoder/encoder, etc.

If you horizontally scale via plugin modules, it either won't be able to present to software as one chip or if it can, it'll be very slow for certain applications that depend on low latency.

 

[leman]

It isn't the same. What we want is to present software with one CPU, one GPU, one Neural Engine, one decoder/encoder, etc.

If you horizontally scale via plugin modules, it either won't be able to present to software as one chip or if it can, it'll be very slow for certain applications that depend on low latency.

Absolutely. I don't think we are in no disagreement at all, in fact we seem to be saying exactly the same thing. My point is that if the problem of low-overhead communication between plugin modules can be solved, then the difference between an SoC and plugin modules disappears. But since such technology appears unlikely, it becomes important to extract as much performance as possible from one SoC. This means (as you say) scaling the SoC horizontally (more dies with fast die-to-die interface) and vertically (higher clocks).

As I don't believe we will see an Extreme (for dies) SoC this iteration, the most feasible option becomes overclocking the hell out of M2 Ultra. Or maybe a limited run 3nm chip. Looking forward to seeing what they go for.

 

[senttoschool]

Absolutely. I don't think we are in no disagreement at all, in fact we seem to be saying exactly the same thing. My point is that if the problem of low-overhead communication between plugin modules can be solved, then the difference between an SoC and plugin modules disappears. But since such technology appears unlikely, it becomes important to extract as much performance as possible from one SoC. This means (as you say) scaling the SoC horizontally (more dies with fast die-to-die interface) and vertically (higher clocks).

As I don't believe we will see an Extreme (for dies) SoC this iteration, the most feasible option becomes overclocking the hell out of M2 Ultra. Or maybe a limited run 3nm chip. Looking forward to seeing what they go for.

I don't think anyone has solved this problem yet. It seems like it's just a matter of physics. It takes longer to transfer things when the distance is longer. A direct connection will always beat out a module-like architecture.

Even Nvidia's NVLink, which connects multiple GPUs together, does not present the system with 1 GPU despite the fact that GPUs aren't as latency sensitive.

 

[senttoschool]

the most feasible option becomes overclocking the hell out of M2 Ultra

This is the easiest option, but least desirable. I suppose Apple can overclock the chip up to 255w because the highest Xeon they support uses up to 255w. But even then, maybe you'll get 30% more performance but it comes with the expense of losing some of the marketing halo around Apple Silicon being generally extremely efficient from top to bottom.

But even if the Ultra can run 30% faster in a Mac Pro than a Studio, it still doesn't seem like a good sell to buyers in my opinion. If I'm buying a Mac Pro, I'd at least want 70-80% faster speed than the Ultra and the ability to use hundreds of GBs of RAM.

 

[leman]

This is the easiest option, but least desirable. I suppose Apple can overclock the chip up to 255w because the highest Xeon they support uses up to 255w. But even then, maybe you'll get 30% more performance but it comes with the expense of losing some of the marketing halo around Apple Silicon being generally extremely efficient from top to bottom.

But even if the Ultra can run 30% faster in a Mac Pro than a Studio, it still doesn't seem like a good sell to buyers in my opinion. If I'm buying a Mac Pro, I'd at least want 70-80% faster speed than the Ultra and the ability to use hundreds of GBs of RAM.

Here I disagree. If they can make M2 Ultra 30% faster that M2 Max in single core, it will have better single core performance than top gaming chips from Intel or AMD while offering competitive performance with workstation chips. That’s something nobody else has. Even the new Intel WS Xeons take a noticeable performance hit in single core workloads. And it doesn’t matter a bit whether the Mac Pro CPU draws 60 watts or 200 watts - it will still be more efficient than Xeons or Threadrippers with the same performance.

Also, they have much more power at their disposal than 255 watts. The system is capable of continuously dissipating well over 600 watts. They just need a beefy enough radiator

 

[Pressure]

I still believe it makes more sense to make a larger SoC (650-700mm²) for the Mac Pro.

The trouble of "gluing" more than two Mx Max chips together just seems so inelegant and introduces scaling issues. It also takes useable chips away from their best selling line-up.

 

[Serqetry]

This is the easiest option, but least desirable. I suppose Apple can overclock the chip up to 255w because the highest Xeon they support uses up to 255w. But even then, maybe you'll get 30% more performance but it comes with the expense of losing some of the marketing halo around Apple Silicon being generally extremely efficient from top to bottom.

I just don't believe there's much chance of Apple going with this. If they considered it an option to overclock at the cost of making a computer big, loud, hot, and power hungry, we didn't need to wait 3+ years for a new Mac Pro. I think Apple is far more likely to decide the Mac Pro has no future than go down this dead end path.

Not to mention the Mac Studio could have been pushed a lot harder if they thought doing that was acceptable. And they could have just called the Studio the new "Mac Pro".

Remember, the same dumb M2 Ultra rumors from Mark Gurman say it won't have upgradable RAM or support pcie GPUs. If this were all true, it would probably be more poorly received than the Trash Can Mac Pro. They'd be better off just updating the Mac Studio with M2 and making the SSDs upgradable if that's the best they can do.

 

[leman]

I still believe it makes more sense to make a larger SoC (650-700mm²) for the Mac Pro.

The trouble of "gluing" more than two Mx Max chips together just seems so inelegant and introduces scaling issues. It also takes useable chips away from their best selling line-up.

Quite the opposite. Multi-die systems are an extremely elegant solution and do not introduce any scaling issues with appropriate technology. But most importantly, they allow a much better utilisation of production resources. The chip foundry has only limited production capacity. The larger the die, the higher the change that it will be defective. So making smaller dies and gluing them together ends up producing more useable devices of both types. Not to mention that one can quickly balance supply and demand and you can decide where your dies go.

There is a reason why everyone is pushing hard for multi-die systems. AMD has been using it for some time on their desktop CPUs with great success and has recently introduced it on their GPUs. Intel uses it for new Xeons to glue together four dies. Nvidia is reportedly planning to use this tech in their upcoming GPUs.

I just don't believe there's much chance of Apple going with this. If they considered it an option to overclock at the cost of making a computer big, loud, hot, and power hungry, we didn't need to wait 3+ years for a new Mac Pro. I think Apple is far more likely to decide the Mac Pro has no future than go down this dead end path.

It is also possible that M1 was not capable of running higher clocks without stability issues while M2 is a tweaked design capable of sustained much higher clocks.

 

[Serqetry]

It is also possible that M1 was not capable of running higher clocks without stability issues while M2 is a tweaked design capable of sustained much higher clocks.

Again, if this were the case, you'd think it would be a priority for Apple to just put out an updated M2 Mac Studio where they could easily run it at a higher clock speed, and then scrap the Mac Pro. In fact they probably would have made the M1 Studio a little taller with an even more powerful cooler if they were even considering this path in the future, although what I've seen it sounds like the Studio as-is would be perfectly capable of handling a lot more heat.

Whatever they decide to do, they're going to have to do a lot better than just running M2 Ultra a little bit faster if anyone is going to take a new "Mac Pro" seriously.

 

[leman]

Again, if this were the case, you'd think it would be a priority for Apple to just put out an updated M2 Mac Studio where they could easily run it at a higher clock speed, and then scrap the Mac Pro. In fact they probably would have made the M1 Studio a little taller with an even more powerful cooler if they were even considering this path in the future, although what I've seen it sounds like the Studio as-is would be perfectly capable of handling a lot more heat.

Depends on how much faster these chips can run. Studio can probably easily deal with 150-200 watts of power, but nothing beyond that. And of course, Studio doesn't have the internal expandability which is important to some workstation users.

Whatever they decide to do, they're going to have to do a lot better than just running M2 Ultra a little bit faster if anyone is going to take a new "Mac Pro" seriously.

Why not? I saw some leakers claiming that M2 Ultra can run at 4.2 Ghz. That would put its single-core performance on par with fastest enthusiast gaming processors while making the multi-core performance competitive with the 32-core Threadrippers or the new Intel Xeon w9-3475X. For around $8k, it would be a very competitively priced CPU workstation with unmatched combination of single- and multi-core performance.

 

[dmccloud]

Again, if this were the case, you'd think it would be a priority for Apple to just put out an updated M2 Mac Studio where they could easily run it at a higher clock speed, and then scrap the Mac Pro. In fact they probably would have made the M1 Studio a little taller with an even more powerful cooler if they were even considering this path in the future, although what I've seen it sounds like the Studio as-is would be perfectly capable of handling a lot more heat.

Whatever they decide to do, they're going to have to do a lot better than just running M2 Ultra a little bit faster if anyone is going to take a new "Mac Pro" seriously.

The studio is not and never will be a replacement for a Mac Pro. The users who want a Mac Pro want that form factor for expandability, which of often of more importance than raw power for that subset of users. Furthermore, given how large the heatsink already is for the M1 Ultra in the Studio, there really isn't much room to upgrade the SoC at all without making the chassis significantly larger. With the Mac Pro (especially if they keep the current case), they have more than enough space for cooling and expansion.

 

[Serqetry]

The users who want a Mac Pro want that form factor for expandability, which of often of more importance than raw power for that subset of users.

Yeah and Apple doesn't really seem to value expandability anymore. And all the rumors around the new Mac Pro claim that it will not be expandable. So... which is it? Rumors are true (M2 Ultra, same old case), or they aren't (expandability)? If the rumors aren't true, then all bets are off on what kind of Apple Silicon we're going to get in it.

 

[Boil]

Mac Studio = no PCIe slots

ASi Mac Pro = PCIe slots

This is why the Mac Studio is not the Mac Pro replacement, the lack of PCIe slots...

Remember, PCIe slots are used for more than just discrete GPUs...

 

[jaehaerys48]

I'm betting on 10 water cooled Power Mac G5 processors strapped together.

 

[Serqetry]

Mac Studio = no PCIe slots

ASi Mac Pro = PCIe slots

This is why the Mac Studio is not the Mac Pro replacement, the lack of PCIe slots...

Remember, PCIe slots are used for more than just discrete GPUs...

Yeah... and what's so important about PCIe slots if you can't put a GPU in it?

If you think people are dying for a new Mac Pro so they can fill it full of network cards, I think you're probably mistaken. PCIe slots are not a good enough reason for Apple to make a new Mac Pro.

I would bet Apple views PCIe slots as outdated technology.

 

[Boil]

Yeah... and what's so important about PCIe slots if you can't put a GPU in it?

If you think people are dying for a new Mac Pro so they can fill it full of network cards, I think you're probably mistaken. PCIe slots are not a good enough reason for Apple to make a new Mac Pro.

I would bet Apple views PCIe slots as outdated technology.

Do you really think PCIe slots are only used for GPUs & networking cards...?

 

[Serqetry]

Do you really think PCIe slots are only used for GPUs & networking cards...?

Of course not. But how about you tell me exactly what PCIe cards are so essential to a Mac that Apple has to release a new Mac Pro that supports them? The advantage to being able to use GPU cards is very obvious, but I think we know the chances of Apple implementing that in an ASi Mac Pro is slim.

Apple sure doesn't seem to think slots are essential, or even expansion in general. The only machine they make that uses them has been consistently ignored for long stretches of time. It seems pretty obvious to me the only thing they consider essential these days for expansion is thunderbolt ports.

 

[Boil]

Of course not. But how about you tell me exactly what PCIe cards are so essential to a Mac that Apple has to release a new Mac Pro that supports them? The advantage to being able to use GPU cards is very obvious, but I think we know the chances of Apple implementing that in an ASi Mac Pro is slim.

Apple sure doesn't seem to think slots are essential, or even expansion in general. The only machine they make that uses them has been consistently ignored for long stretches of time. It seems pretty obvious to me the only thing they consider essential these days for expansion is thunderbolt ports.

This has been covered numerous times...

8K video I/O cards require more bandwidth than TB provides, as do RAID cards...

Folks running DAWs with a bunch of audio I/O & audio DSP cards like a lot of slots, and usually in a rack mount variant (such as the 7,1 rack mount variant)...

6,1 Mac Pro aside, every single Power Mac / Mac Pro has had expansion slots, pretty much a defining characteristic of the product...

 

[Serqetry]

This has been covered numerous times...

8K video I/O cards require more bandwidth than TB provides, as do RAID cards...

Folks running DAWs with a bunch of audio I/O & audio DSP cards like a lot of slots, and usually in a rack mount variant (such as the 7,1 rack mount variant)...

6,1 Mac Pro aside, every single Power Mac / Mac Pro has had expansion slots, pretty much a defining characteristic of the product...

Yes I think it would be great if a new Mac Pro could use all those things. I just don't think Apple cares enough about any of that to consider it essential. Certainly they don't care about supporting 3rd party 8k video cards.

I guess we'll just have to wait and see what happens. I think either the new Mac Pro will be amazing and nothing like what we're hearing with the rumors, or it will be a huge disappointment. And I just don't believe Apple will put out a disappointing machine that is barely more powerful than a Mac Studio and be like "but look it has PCIe slots!!!"

 

[quarkysg]

Yes I think it would be great if a new Mac Pro could use all those things. I just don't think Apple cares enough about any of that to consider it essential. Certainly they don't care about supporting 3rd party 8k video cards.

Wouldn't it be that the vendor providing the 8K capture card would have to support said card by writing drivers for it?

Apple would just need to support the vendor providing the capture card by way of OS API support.

This should be true for all plug in cards or devices via ports.

 

[senttoschool]

Also, they have much more power at their disposal than 255 watts. The system is capable of continuously dissipating well over 600 watts. They just need a beefy enough radiator

The 600 watts is spread over the CPU, and the GPU(s), and the rest of the components, right?

Since Apple Silicon is a SoC, it'd sit where the CPU sits and that's where all the heat will concentrated. Hence, I used the 255w figure. I'm sure Apple could make modifications to direct better airflow to the SoC than they do right now to the Xeon CPUs but absent this certainty, I used 255w.

 

[leman]

The 600 watts is spread over the CPU, and the GPU(s), and the rest of the components, right?

Since Apple Silicon is a SoC, it'd sit where the CPU sits and that's where all the heat will concentrated. Hence, I used the 255w figure. I'm sure Apple could make modifications to direct better airflow to the SoC than they do right now to the Xeon CPUs but absent this certainty, I used 255w.

We have consumer hardware on the market that dissipates over 600 watts in a PC tower, so none of that should be a problem. Mac Pro has excellent airflow. It's just a question of putting a large enough radiator on the SoC and directing the airflow to it (which can be easily done).

 

[steve123]

We have consumer hardware on the market that dissipates over 600 watts in a PC tower, so none of that should be a problem. Mac Pro has excellent airflow. It's just a question of putting a large enough radiator on the SoC and directing the airflow to it (which can be easily done).

I think you just hit the nail on the head about why Apple will not entertain the ideas suggested here about increasing power consumption. Dissipating 600W does not seem like something a company concerned about waste would do.

 

[iPadified]

I think you just hit the nail on the head about why Apple will not entertain the ideas suggested here about increasing power consumption. Dissipating 600W does not seem like something a company concerned about waste would do.

Is not the cooler in MP2019 cooling 400W. Seem enough.

Too many threads end in Mac Pro speculations at the moment. Apple please stop the pain: release it now or kill it (like the iMac 27 inch) so we can move on to complain about the complete absence of an MP or the absence of GTX cards, cheap RAM and SSD upgrades and of course the price and bean counter Tim.