MP 7,1 Mac Pro Apple Silicon is going to eat our lunch

[singhs.apps]

Given new MBPs back to Titanium Book like-alike, 24-inch iMac (and new MBA rumoured) back in candy colors, isn't Apple in some sort of nostalgic? Of course, I'm no genius or fool at Apple. Your guess may be as good as many others

I was looking back at the changes Apple has been making in it’s designs..starting from a proper Mac Pro (2019)… I did notice a trend of ‘back to the Mac’ vibes from Apple.
The new iMacs and Mac book pros…
Even the new iPhone 14 is rumoured to look like the classic iPhone 4 (best iPhone design imo)

In that by now famous interview, back in 2018..Apple did use the term ‘macintosh’ for a new Mac Pro..a name I hadn’t heard in years…

 

[singhs.apps]

Capture cards, mainly.

What kind of capture cards .. Audio-/Video ? Would external capture cards plugged via thunderbolt 4 not suffice ?

 

[Boil]

With Thunderbolt 4, Apple has no need for PCIe slots...

 

[mattspace]

Sony/AMD created a "unified memory architecture" in PS4 ten years ago but AMD (or Intel) didn't attempt to bring the architecture to PC. Assume people in there knew it didn't make business sense. I don't know if it makes more sense today in PC world. We're likely to see first hints from PC companies soon.

They created it as a way to compensate for putting lower capability hardware into a console made to a pricepoint - the PC world, with, from Sony's perspective "more primitive" architecture still achieved higher outcomes.

Efficiency and elegance isn't a guarantee of superior performance.

 

[mattspace]

With Thunderbolt 4, Apple has no need for PCIe slots...

They said the same thing about Thunderbolt 2, and it was just as misguided.

Thunderbolt 4 is just Thunderbolt 3, with updated USB and Displayport alt modes. It's still the same 1/4 the bandwidth of a PCI3 slot, and PCI 5 slotted motherboards are on the market already.

PCI5 - 2x -> PCI4 - 2x -> PCI3 -> 4x -> Thunderbolt 4

So, 1/16th the bandwith of a slot that every cheap consumer motherboard is going to have in 6-12 months.

 

[Melbourne Park]

They said the same thing about Thunderbolt 2, and it was just as misguided.

Thunderbolt 4 is just Thunderbolt 3, with updated USB and Displayport alt modes. It's still the same 1/4 the bandwidth of a PCI3 slot, and PCI 5 slotted motherboards are on the market already.

PCI5 - 2x -> PCI4 - 2x -> PCI3 -> 4x -> Thunderbolt 4

So, 1/16th the bandwith of a slot that every cheap consumer motherboard is going to have in 6-12 months.

The difference, perhaps, with Apple, is that they'd prefer not to have customer add ons. An exception may be, adding Apple own ones.

Ultimately a computer operates as fast as its tasks' demand, less its bottlenecks. There's little point of having a faster interface, if by removing a bottleneck, the bottleneck just shifts to somewhere else, and things don't speed up.

There may be ways around the current expensive and incremental improvements. Somehow though I don't see the PC makers and their suppliers, liking that business model. But perhaps Apple approach might create some change. There's enough potential players out there with the ability to do something alternative.

 

[singhs.apps]

With Thunderbolt 4, Apple has no need for PCIe slots...

Hehe… if that were the case we should have seen droves of such thunderbolt devices by now in the rest of the world, AKA 90% of the PC market.

Sorry, but with PCIe 6 on the horizon a few years from now (looks like the standards group is trying to make up for lost time after dragging their feet for almost a decade ), internal expansion on desktops is here to stay.

Thunderbolt is not even in the same league.

 

[goMac]

Sony/AMD created a "unified memory architecture" in PS4 ten years ago but AMD (or Intel) didn't attempt to bring the architecture to PC. Assume people in there knew it didn't make business sense. I don't know if it makes more sense today in PC world. We're likely to see first hints from PC companies soon.

Ryzen is a unified memory architecture. Most people don't use it that way, they add a discrete GPU on.

But saying there are no UMA PC architectures is dead wrong. Ryzen was literally based off of the chip in the PS4, and directly brought the PS4 tech to PC. And projects like Infinity Fabric are working to pull dGPUs into the UMA space as well.

Also, the PS5 and Xbox (because of Ryzen) are also UMA.

Also, it's widely expected that Nvidia will enter the PC space with UMA after their ARM acquisition. UMA requires graphics and CPU to both exist in the same company, which AMD has had a monopoly on. It's expected that Nvidia's moves into CPUs and Intel's moves into GPUs are both first moves into shipping UMA products.

(AMD even has a nice support document clarifying the UMA configuration options for Ryzen: https://www.amd.com/en/support/kb/faq/pa-280)

Apple indeed did it before or at least very similar in Class Intel Mac Pro where they had the CPU tray for housing the CPU and DIMMs.

Apple Silicon is nothing like the trays in a Mac Pro. For Unified Memory to work, all GPUs and all CPUs in the system would need to use the same bank of RAM. Putting different RAM on different trays doesn't solve that. You're just putting multiple Mac Minis in a single case. Nothing makes that unified.

Basically it's adding a bunch of discrete SoCs into the same box, and then handwaving and going "and unification magic happens somehow."

Apple could use something like Infinity Fabric to link multiple SoCs together, like AMD is doing. But that's not true unified memory space, it just makes the CPUs and GPU's act like they're in a single unified space. Still interesting, but it's not the same performance outcome.

It's also not really necessary. The rumored "quad" M1 Max chip will be plenty hot and plenty fast. And probably extremely expensive.

 

[Melbourne Park]

....

Basically it's adding a bunch of discrete SoCs into the same box, and then handwaving and going "and unification magic happens somehow."

Apple could use something like Infinity Fabric to link multiple SoCs together, like AMD is doing. But that's not true unified memory space, it just makes the CPUs and GPU's act like they're in a single unified space. Still interesting, but it's not the same performance outcome.

It's also not really necessary. The rumored "quad" M1 Max chip will be plenty hot and plenty fast. And probably extremely expensive.

Qualification is that I know little about computer architecture. But the memory that Apple uses is still physical solid state memory. LPDDR5-6400 SDRAM chips in fact. The LP stands for low power. On 8 channels with a bandwidth of 200GB/s on the Pro and 400GB/s for the M1 Max. I imagine one might connect multiple processors memory via higher bandwidth memory connectors that might allow multiple processors to share their unified memory. Even just changing from low power LP memory would assist. And then there is DDR6 memory which is faster memory, as is DDR6X memory. So for desktops, Apple could change the memory type to get greater speed and band width, which would assist in achieving a unified memory using more than one M type processor.

I presume Apple will scale their processors, and doing so should remove the need for adding in costly 3rd party GPUs. I hope their desktop Mac Pro architecture allows buyers to have a scaleable upgrade path directly from Apple, as I suspect GPUs may not be part of the Apple M architecture's future. If Apple does that, it will cost the Pro user less than buying top end professional grade GPUs. But whether such scalability would compete with somewhat similar to Pro grade GPU but much cheaper games style GPUsy, would make a big difference in affordability.

I guess that will depend on whether Apple is designing in customer based scalability. IMO the justification for buying Mac Pro desktops (excluding the 6,1 models) was based on two criteria - getting the performance you needed for the next two years, and then being able to upgrade later on to attain higher levels of performance when you'll then need the extra performance. I hope Apple keeps providing such hardware. We don't know how they'll do it though, or if they do it.

 

[goMac]

I imagine one might connect multiple processors memory via higher bandwidth memory connectors that might allow multiple processors to share their unified memory.

So yes, this is what Infinity Fabric very roughly is.

The problem is at this point it's not Unified Memory. It's Discrete Unified Memory which is... not a thing. That's just discrete memory.

Like I said, it's plausible, but the performance might not be there. The processors will spend so much time synchronizing memory that they may not be able to perform well. For example, the multiple GPUs may have to spend a lot of time synchronizing with each other while drawing a frame, which is not optimal.

So you go back to: An M1 Max Quad would likely be fast enough you don't need to have the fuss of doing this.

And if you're going to go down this route, why bother with Unified Memory? Just go back to discrete GPUs. What you're describing is basically discrete GPUs anyway. Why try to bring Unified Memory into things, just to immediately break the performance advantages of Unified Memory.

 

[4wdwrx]

Yup, and PowerPC is still used on Mars rovers and other deep space vehicles, doesn't mean its relevant for desktop processing, which is what Apple, and everyone who uses computers in the same physical place as themselves, does.

Even if the entire rest of the industry, including Intel goes to ARM-derived designs, it's Apple Silicon that's going to end up as the new PowerPC, not ARM in general.

Very true.

I also think Apple is transitioning from the common "computers". Apple has been cohesively integrate all their devices. Now that Macs are also ARM based, it can now finally do the same.

Apple computers will be less of a "computer" or "laptop" that we think of and more as IoT devices.

Like a Mars rover to the mission. Apple is creating IoT devices to provide a certain experience or "mission", rather than a "computer" for different things, with no real "mission".

 

[singhs.apps]

So yes, this is what Infinity Fabric very roughly is.

The problem is at this point it's not Unified Memory. It's Discrete Unified Memory which is... not a thing. That's just discrete memory.

Like I said, it's plausible, but the performance might not be there. The processors will spend so much time synchronizing memory that they may not be able to perform well. For example, the multiple GPUs may have to spend a lot of time synchronizing with each other while drawing a frame, which is not optimal.

So you go back to: An M1 Max Quad would likely be fast enough you don't need to have the fuss of doing this.

And if you're going to go down this route, why bother with Unified Memory? Just go back to discrete GPUs. What you're describing is basically discrete GPUs anyway. Why try to bring Unified Memory into things, just to immediately break the performance advantages of Unified Memory.

So what do you think apple will do ?

 

[Sophisticatednut]

So yes, this is what Infinity Fabric very roughly is.

The problem is at this point it's not Unified Memory. It's Discrete Unified Memory which is... not a thing. That's just discrete memory.

Like I said, it's plausible, but the performance might not be there. The processors will spend so much time synchronizing memory that they may not be able to perform well. For example, the multiple GPUs may have to spend a lot of time synchronizing with each other while drawing a frame, which is not optimal.

So you go back to: An M1 Max Quad would likely be fast enough you don't need to have the fuss of doing this.

And if you're going to go down this route, why bother with Unified Memory? Just go back to discrete GPUs. What you're describing is basically discrete GPUs anyway. Why try to bring Unified Memory into things, just to immediately break the performance advantages of Unified Memory.

We have no reason to believe unified memory and discrete memory can't exist at the same time.
IF our CPUs can have L1,L2 and L3 cash memory + RAM, why can't we have the same principles with Ram(L1+L2 cash)? Apples unified Memory is essentially working as L3 Cash to the CPU/GPU with 16-64 GB worth of it.
L1 cash can peak at 2.3 TB/s L3 Cash at 600 GB/s respectively.
And L3 cash is usually only 20Mb

Apple have essentially implemented their own version of infinity fabric with 200-400Gbps

If we add normal DDR5 ram modules, they would work as L2 memory for the faster Unified memory.
DDR5 supports a speed of 51.2 GB/s per module. This is still way faster than running out of Unified memory and writing from the SSD.

1. L1 cash 2.3 TB/s
2. L2 cash 1 TB/s
3. 16-128Gb Unified memory 400 GB/s (effectively L3cash)lower latency
4. 16-1Tb DDR5 memory 100 GB/s(Dual channels) higher latency
5. 500Gb PCIe 4 SSD storage 7.4GB/s extremely high latency

Why on earth do you think it's a good thing to write directly from storage instead of using DDR5 sticks?

 

[goMac]

We have no reason to believe unified memory and discrete memory can't exist at the same time.

They... can't? Like seriously, they can't.

The point of unified memory is that _all memory is unified._ That's... how it works. That's the performance boost.

You can't have unified memory and discrete memory because _that's not a unified model._ ALL memory must be unified in the unified model.

That's why when you add a discrete card on a Ryzen system, it drops out of pure unified mode. Because you can't have unified and discrete. If you have discrete then you aren't unified. They're literally words that mean the opposite thing.

You're right that they could implement Infinity Fabric. That would give them the behavior of Unified Memory, _but not the performance._ And in since Infinity Fabric works with discrete GPUs, why bother at that point? They could just go back to a discrete model. Pure unified memory does absolutely nothing for you when you start adding discrete components. If they want to do Infinity Fabric, just go back to discrete GPUs on Infinity Fabric. Toss out unified memory in hardware completely. It's simpler that way.

Heck, they're most the way now. The current Mac Pro already supports Infinity Fabric on the GPUs already.

The entire point of Apple's Unified Memory implementation is that there is no sync of different devices and that's extremely performant. Building an entire "sync a bunch of different devices" thing on top of it defeats the point.

(Should be noted that Infinity Fabric and UMA do _not_ work the same in the Metal APIs. So that would cause compatibility issues I doubt Apple wants to deal with.)

So what do you think apple will do ?

They'll ship a Mac Pro that has a single SoC CPU and... that's it. People are making it complicated. It's not going to be complicated. It's going to be a single chip.

 

[kvic]

Ryzen is a unified memory architecture. Most people don't use it that way, they add a discrete GPU on.

But saying there are no UMA PC architectures is dead wrong. Ryzen was literally based off of the chip in the PS4, and directly brought the PS4 tech to PC. And projects like Infinity Fabric are working to pull dGPUs into the UMA space as well.

Also, the PS5 and Xbox (because of Ryzen) are also UMA.

I suggest people read one of my previous posts on the background of "unified memory architecture", post #75 in this thread.

PS4 was innovative of its time not simply because it's "unified memory architecture" but it's also high performance for its intended purpose. Sony continued its success of the architecture in PS5. Microsoft saw and copied PS4 and created Xbox Series X.

I believe AMD (and Intel) didn't push PS4-like architecture to PC because 1) as-is the architecture performs poorly in general-purpose computing; 2) DDR technology back then was not up to the job; 3) such an architecture were expensive and not flexible.

x86-64 processors with integrated GPU are technically "unified memory architecture". This includes most Intel client processors since Sandy Bridge, and AMD APUs which is basically x86-64 CPU plus an integrated GPU on the same die. All of sudden, UMA becomes a buzzword and vogue thanks to Apple Marketing, Fans & Co. For UMA to be meaningfully useful today, one predicate is high memory bandwidth.

Apple Silicon is nothing like the trays in a Mac Pro. For Unified Memory to work, all GPUs and all CPUs in the system would need to use the same bank of RAM. Putting different RAM on different trays doesn't solve that. You're just putting multiple Mac Minis in a single case. Nothing makes that unified.

[rest of the post snipped]

I'm afraid you completely missed the gist of the discussion you quoted

We have no reason to believe unified memory and discrete memory can't exist at the same time.
IF our CPUs can have L1,L2 and L3 cash memory + RAM, why can't we have the same principles with Ram(L1+L2 cash)? Apples unified Memory is essentially working as L3 Cash to the CPU/GPU with 16-64 GB worth of it.
L1 cash can peak at 2.3 TB/s L3 Cash at 600 GB/s respectively.
And L3 cash is usually only 20Mb

Apple have essentially implemented their own version of infinity fabric with 200-400Gbps

That's not Infinity Fabric. That's LPDDR5 memory bandwidth and nothing more.

I don't believe Apple will bother with DIMMs, at least not in the short term. Really, a tiny number of users will need terabytes of RAM..why Apple have to bother with?

In future, if Apple have to add DIMMs to Mac Pro, I have an interesting idea on how they may add it. I haven't seen it discussed anywhere on Internet yet. Save it for another day.

 

[singhs.apps]

And you don’t see Apple running into yield issues ?

 

[goMac]

I believe AMD (and Intel) didn't push PS4-like architecture to PC because 1) as-is the architecture performs poorly in general-purpose computing; 2) DDR technology back then was not up to the job; 3) such an architecture were expensive and not flexible.

x86-64 processors with integrated GPU are technically "unified memory architecture". This includes most Intel client processors since Sandy Bridge, and AMD APUs which is basically x86-64 CPU plus an integrated GPU on the same die. All of sudden, UMA becomes a buzzword and vague thanks to Apple Marketing, Fans & Co. For UMA to be meaningfully useful today, one predicate is high memory bandwidth.

Ryzen is a literal UMA system in the same way Apple Silicon is. It's not marketing. I don't know what to tell you. It's the same CPU series that was in the PS4. AMD did bring the PS4 architecture to PC. It's Ryzen.

Thats why Apple couldn't claim to be the first PC architecture with UMA. They're not.

The Xbox Series X isn't a copy. It's all Ryzen CPUs. They all just bought Ryzen CPUs. In fact, the Xbox Series X definitely didn't copy because the Xbox One also used a Ryzen CPU with UMA.

Intel doesn't meet the definition of UMA. It's not a unified memory addressing space. There's a firm definition here that doesn't have to do with marketing.

 

[goMac]

And you don’t see Apple running into yield issues ?

No.

The rumor is that they're going to do a chiplet approach. Specifically: they're just going to arrange 2 to 4 M1 Maxes dies on a single chip with a single bank of unified memory.

So as long as they have M1 Max yield, they'll have yield.

I believe AMD is using a similar approach on newer Ryzen CPUs.

 

[singhs.apps]

No.

The rumor is that they're going to do a chiplet approach. Specifically: they're just going to arrange 2 to 4 M1 Maxes dies on a single chip with a single bank of unified memory.

So as long as they have M1 Max yield, they'll have yield.

I believe AMD is using a similar approach on newer Ryzen CPUs.

This is a bit confusing now.
All this while I was under the impression that Apple will design the quad/jade/whatever Mac Pro AS the way AMD has been doing for its Threadripper and Epyc CPUs - connect the chiplets via ‘infinity Fabric’.

Your post above also suggests Apple will take the same approach… but this wouldn’t violate your strict criteria for UMA ?
All of AMD’s Desktop/laptop/sever CPUs do use discreet memory.

In effect the base 4/8 (and now the suggested 16 core ) Ryzen CPUs would be the M1pro/Max equivalent and Apple will use Threadripper/Epyc style interconnected M1 ‘chiplets’

 

[goMac]

This is a bit confusing now.
All this while I was under the impression that Apple will design the quad/jade/whatever Mac Pro AS the way AMD has been doing for its Threadripper and Epyc CPUs, connect the chiplets via ‘infinity Fabric’

Your post above also suggests Apple will take the same approach… but this wouldn’t violate your strict criteria for UMA ?
All of AMD’s Desktop/laptop/sever CPUs do use discreet memory.

Quick background: UMA, in the hardware sense, means every chip has direct access to the same data in the same memory. Specifically, it usually implies that data does not have to be copied between devices because they all use the same memory.

The performance advantage is that data never has to be copied or moved. The GPU and CPU can work on the same data in the same place in memory. Anything the CPU has loaded is implicitly accessible by the GPU instantaneously.

This is why M1 Max is able to outperform a lot of other GPUs even though it's only about 10 tflops. It's not actually that strong of a GPU. But because it has instant access to anything in RAM, it doesn't need to wait to load data. It can instantaneously start working while a bigger but more powerful Nvidia GPU might be idling waiting for data to load.

To your question: A chiplet design still shares a single bank of RAM. There aren't multiple banks of RAM that all need to be linked or synced. So it still meets the definition.

A larger Infinity Fabric system that involves discrete components breaks what Apple is doing. Because you're waiting for data to load and sync, you lose your performance advantage. And because M1 Max's GPU isn't actually that fast, and you've tossed out it's only advantage, you might as well just go back to discrete GPUs.

Apple's entire strategy revolves around data not having to synchronize between devices. Add back synchronization between multiple boards or SoCs, and you'll kill performance.

Infinity Fabric with discrete components is like UMA emulation. It lets you pretend you have UMA, but you don't get the performance advantages. Apple could actually build a tower that uses Infinity Fabric to emulate UMA. They could do that with x86 using something like Ryzen. But with their current SoCs it doesn't make sense because the design is built so strongly around data not needing to sync.

(Note: Intel Integrated Graphics don't meet the definition of UMA even though they use shared memory. Intel partitions the on board memory between the GPU and CPU instead of sharing it. So it's not unified.)

----------

Edit: One thing I thought of that's probably not helping this conversation is Infinity Fabric can mean a few different things. The Infinity Fabric that AMD would use to build a chiplet is not quite the same thing as the Infinity Fabric that would link multiple discrete components. AMD also uses the term Infinity Fabric to refer to syncing/enforcing cache coherency on a discrete components as well. They're all based on Infinity link technology, but have very different performance meanings.

 

[singhs.apps]

Makes sense.
What about CXl over PCI-e 5.0 and above ? Wouldn’t that mitigate discreet card bottle necks ?

I can understand Apple’s need to have UMA to help it’s GPU … but locking into an AIO Chip with zero internal expansion…errrrm.. we are back to the cube and cylinder days

 

[Sophisticatednut]

They... can't? Like seriously, they can't.

Yes they can. They are doing it right now.
CPU with L1+L2 cash is one level
Unified memory Is the second level
And your Storage is the last level.

The point of unified memory is that _all memory is unified._ That's... how it works. That's the performance boost.

You can't have unified memory and discrete memory because _that's not a unified model._ ALL memory must be unified in the unified model.

That is not how memory works at all. have you forgotten why we have RAM in the first place? what the purpose of separate "memory" instead of using the storage medium directly?

That's why when you add a discrete card on a Ryzen system, it drops out of pure unified mode. Because you can't have unified and discrete. If you have discrete then you aren't unified. They're literally words that mean the opposite thing.

how AMD does it is not even equivalent to how Apple does it. it's close to 4 times slower.( ithink it's about 70~Gbps9

You're right that they could implement Infinity Fabric. That would give them the behavior of Unified Memory, _but not the performance._ And in since Infinity Fabric works with discrete GPUs, why bother at that point? They could just go back to a discrete model. Pure unified memory does absolutely nothing for you when you start adding discrete components. If they want to do Infinity Fabric, just go back to discrete GPUs on Infinity Fabric. Toss out unified memory in hardware completely. It's simpler that way.

Heck, they're most the way now. The current Mac Pro already supports Infinity Fabric on the GPUs already.

Forget Infinity fabric, i used it as an equivalent system as an example.
Apple would have their 16-128Gb LPDDR5 unified memory. and from this you would load data from the "slow" DDR5 sticks. your DDR5 sticks would works as cash for the Unified memory system apple uses.

The entire point of Apple's Unified Memory implementation is that there is no sync of different devices and that's extremely performant. Building an entire "sync a bunch of different devices" thing on top of it defeats the point.

(Should be noted that Infinity Fabric and UMA do _not_ work the same in the Metal APIs. So that would cause compatibility issues I doubt Apple wants to deal with.)

They'll ship a Mac Pro that has a single SoC CPU and... that's it. People are making it complicated. It's not going to be complicated. It's going to be a single chip.

There won't be syncing between different devices. it will never be infinity fabric(
You are essentially arguing for an ineffective system, why you could have

1. CPU L1+L2 cash this is the fastest memory in the computer
2. LPDDR5 Unified memory system 16-128Gbps with 400Gbps bandwith, unheard of in the industry. feeding the GPU/CPU with information
3. DDR5 sticks 16gb-1Tb memory with 100Gbps bandwidth Buffer/Cashing data that will go to the Unified memory.
4. SSD your normal slow PCIe4 storage at 7Gbps feeding the RAm buffer/cash instead of slowly feeding the limited Unified memory space

That's not Infinity Fabric. That's LPDDR5 memory bandwidth and nothing more.

I know it's not infinity fabric(see above). it's still 400Gbps of LPDDR5 memory bandwidth, that normally would be at 50~

I don't believe Apple will bother with DIMMs, at least not in the short term. Really, a tiny number of users will need terabytes of RAM..why Apple have to bother with?

it makes no sense for them not to. and taken right from their stor page for the Mac pro.

In future, if Apple have to add DIMMs to Mac Pro, I have an interesting idea on how they may add it. I haven't seen it discussed anywhere on Internet yet. Save it for another day.

Me to. simple DDR5 slots but with Quad memory channels or octagon channels(same as the Unified memory) but with more latency(tecnical reasons) as it's further away from the CPU

 

[goMac]

Makes sense.
What about CXl over PCI-e 5.0 and above ? Wouldn’t that mitigate discreet card bottle necks ?

I can understand Apple’s need to have UMA to help it’s GPU … but locking into an AIO Chip with zero internal expansion…errrrm.. we are back to the cube and cylinder days

Raw bandwidth you'd probably need around 1200 gigabytes a second, assuming the cores scale with bandwidth on a quad configuration.

But I think at a certain point latency becomes a problem too. I live in UMA-as-software and not UMA-as-hardware. But I would suspect everything needs to be really physically close together to get the performance you need. Which is probably why Apple is putting RAM on the chipset. Even moving memory off the chipset onto the board could lead to performance issues. And like I said, they seem to be betting everything on latency.

 

[kvic]

I know it's not infinity fabric(see above). it's still 400Gbps of LPDDR5 memory bandwidth, that normally would be at 50~

it makes no sense for them not to. and taken right from their stor page for the Mac pro. [GRAPHICS snipped]

That's Intel Mac Pro though.

I think Apple will have Intel Mac Pro available for a couple more years (if you follow my train of thoughts in previous posts)

Anyway, it's debatable if Apple will have DIMMs in the coming Smaller Mac Pro. People's guess will be as good as anyone else perhaps

Me to. simple DDR5 slots but with Quad memory channels or octagon channels(same as the Unified memory) but with more latency(tecnical reasons) as it's further away from the CPU

Sorry..I believe you don't see the problem based on what you've posted in this thread

EDIT:

I also don't agree with what goMac said about UMA in the past few messages. I chose to disengage.

 

[goMac]

Yes they can. They are doing it right now.
CPU with L1+L2 cash is one level
Unified memory Is the second level
And your Storage is the last level.

I'm not sure you're reading all of what I'm posting.

You're right, that at basics, it's a cache coherency problem. I even used the words cache coherency in one of my recent posts.

But the problem is if you have multiple "expansion cards" with multiple banks of RAM, you no longer have unified memory. You have different pools of memory.

And again, yes, that's a cache coherency problem. Except now it's much bigger. You have tens or hundreds of gigs of data that all need to be managed and kept coherent.

If you're going to have to solve that problem, then the current Apple Silicon architecture no longer makes sense. If you can synchronize multiple discrete banks of memory across multiple cards quickly, just make the GPU separate again.

If you can solve that problem, you've "solved" the same problem Apple Silicon is trying to fix in the first place.

That's why it doesn't make any sense. If Apple figures that out they can go back to a standard architecture with separate GPUs and CPUs. Why bother with any of this? If you get rid of the need for unified memory then _you don't need unified memory any more._ That's why I'm saying it doesn't make sense to combine a unified memory architecture with a discrete one. Either the discrete components will drag unified memory down, or you've designed discrete components that operate just as fast as unified memory, and then why bother with unified memory.

(My hunch again is that physics will probably be a problem here, and that one reason everything exists on the same package is to minimize the distance signals need to travel.)

how AMD does it is not even equivalent to how Apple does it. it's close to 4 times slower.( ithink it's about 70~Gbps9

Memory speed has nothing to do with the definition of UMA. At that speed, that would be faster than vanilla M1 anyway.

There are UMA Ryzens with fast memory. There are UMA Ryzens with slow memory. Just depends what you're willing to spend. Ryzen systems like the Xbox Series X use faster UMA memory, faster than even M1 Max AFAIK.

You can even build PC with a Ryzen 4700S. That has DDR6, which is faster than what Apple uses. (AMD only sells it as a kit because it's literally PS5 CPUs/boards that didn't pass binning. Technically it's a UMA chip, but unfortunately the GPU is what didn't pass binning on it. Lol.)