M1 Max Quadro Mac Pro Leaks - The Alder Lake KILLER! 🤯

[throAU]

I don't believe making huge amounts of RAM (~1 TB) available to the GPU was ever was intended as one of the purposes (points) of UMA, since it would be a rare use case where the GPU would need that much RAM, and Apple doesn't design its architecture for rare use cases.

Why not?

Apple are positioning themselves as a Hollywood 3d production company and access to GPU memory is a massive constraint for high end modelling.

The Mac Pro in itself is aimed at a rare use case. Afterburner is aimed at a rare use case, and they created those cards for the prior Mac Pro and turned them into on-die processors in the M1 Pro/Max.

The only reasons we don't have GPUs with 1 TB of Ram today is due to impracticality and expense - not because it isn't desirable to have the GPU able to work on ANYTHING the CPU has access to as/when required without needing to do data copies around the place.

Even if the GPU itself doesn't need 1Tb, the whole point is that it can process stuff the CPU may need it to process, which could be anywhere inside the 1TB of memory the CPU has.

Apple's developer talks at WWDC explain the purpose of UMA was so the GPU and CPU could have access to the same RAM pool, obviating the need to copy data back and forth between GPU and CPU RAM

You even posted (quoting Apple which backs up my point) that the whole point is that the GPU can access the memory the CPU has. If the CPU has 1 TB (and that's not a massive amount in 2021) then the GPU needs access to that 1 TB for the whole unified model to work.

If you have a seperate island of memory (some fraction of the the 1TB for the CPU) that the GPU can't access in its entirety, you're back to copying data around which is inefficient.

 

[Analog Kid]

I don't believe making huge amounts of RAM (~1 TB) available to the GPU was ever was intended as one of the purposes (points) of UMA, since it would be a rare use case where the GPU would need that much RAM, and Apple doesn't design its architecture for rare use cases.

There may be a bit of a chicken and egg issue here. It may be rare that a GPU “needs” that much RAM because workloads are designed for the RAM that GPUs have and it’s rare that GPUs have access to that much RAM.

If you’re using a GPU as a graphics processing unit, then it’s probably not often you’d consume that much RAM. If you’re using a GPU as a massively parallel general purpose computing engine then more memory is more better, right?

Right now, huge GPGPU workloads are tiled and sent to the GPU a tile at a time based on memory constraints at the GPU. UMA removes that constraint and opens up GPGPU computing at both extremes: massive datasets requiring a lot of memory and much smaller datasets that ordinarily wouldn’t be worth the data transfer overhead to GPU.

 

[singhs.apps]

Apple are positioning themselves as a Hollywood 3d production company and access to GPU memory is a massive constraint for high end modelling

Most obvious case : CG rendering. If Apple can give access to GPUs the same amount of memory that CPUs have traditionally enjoyed, we can finally see CPUs bow out of that field (or be seen as the only option)
GPUs may have their issues but the massive speeds they bring to muster is sorely needed in those use cases.

 

[iPadified]

I think with the multi-SOC machines the memory won't be quite as "unified".

It will be package local.

So you'll have 2-4 packages (or more) complete with CPU cores, GPU cores and local RAM in a unified cluster with a slightly less high speed bus between the packages.

So it won't be entirely unified, but unified within the local package, which is then linked to other packages via a high speed (but not as high as the on-package bus, due to distance vs. physics) bus.

Think of an M1 Pro or M1 Max cluster as like an EPYC multi-core socket. The EPYC has multiple processor dies on the package, and multiple CCXs within each CPU die - but there's a slower motherboard bus between the sockets.

I can see each M1 based SOC package talking to a cache controller which then accesses resources on the other SOCs via the SOC to SOC bus.



edit:
this is all entirely speculation on my part, but based on what both intel and AMD have been doing, and industry trends/discussion for the past decade, it's how I see Apple scaling the M1 out further.

If so, the MPX modules might contain optional 4X Max. The MPX is 400W and would therefore fit. A master 4X Max where the Xeon is today and addition 2X4Max compute in the MPXs for highly parallel workloads. That is exactly the setup for multiple GPUs today. They built a beautiful and powerful box so why change it just because the SoC paradigm was introduced.

I headless trashcan Mac Pro/cube would be a nice addition to the lineup.

PS. The 3D rendering complexity is limited by the respective GPU amount of RAM. No wonder that 64 Gb unified RAM in the Max chip is appealing. DS

 

[Boil]

If Apple uses a double-sided mobo (like in the 2019 Mac Pro), could the UMA RAM not be chips next to the SoC, but instead be (ECC) DIMMs on the back side of the mobo...? And since we are mounting stuff to the back side of the mobo, include some M.2 slots for NVMe SSDs...? It really is like the 2019 Mac Pro mobo...! ;^p

 

[Boil]

I headless trashcan Mac Pro/cube would be a nice addition to the lineup.

Thinking about a new Cube; 7.7" x 7.7" x 9.8", just a hair over 9.5 liters; seven (current 2018 Space Gray Intel) Mac minis tall...! So, the Mac mini has a 150W PSU in a 1.4" tall chassis; one would need only three of those to power a M1 Max Quad MCM, which means room for (optional) redundant PSUs...?

PSU(s) running up one side of the chassis, vertical mobo up the other; double-sided mobo (like 2019 Mac Pro), ECC DIMM slots & M.2 slots for NVMe SSDs on back side of mobo, vertical ports on back of chassis...

Heat sink (with vapor chamber?) filling interior volume between mobo & PSU(s); 2019 Mac Pro-style 3D venting front & rear, 180mm intake fan up front, 2021 MacBook Pro-style feet...

 

[UBS28]

I am not buying in these cube rumours.

The Mac Pro is an expandable MAC, so I think Apple might keep that.

Else, the Mac Pro is just a beefed up Mac Mini?

 

[thenewperson]

I am not buying in these cube rumours.

The Mac Pro is an expandable MAC, so I think Apple might keep that.

Else, the Mac Pro is just a beefed up Mac Mini?

The rumour-mongers themselves aren't exactly even sure if the smaller Mac Pro is a replacement or an addition. It'd definitely make sense both ways, but obviously be better received if it were in addition.

 

[iPadified]

I am not buying in these cube rumours.

The Mac Pro is an expandable MAC, so I think Apple might keep that.

Else, the Mac Pro is just a beefed up Mac Mini?

A cube and a tower can coexist as they fill different niches.

 

[theorist9]

Why not?

Apple are positioning themselves as a Hollywood 3d production company and access to GPU memory is a massive constraint for high end modelling.

The Mac Pro in itself is aimed at a rare use case. Afterburner is aimed at a rare use case, and they created those cards for the prior Mac Pro and turned them into on-die processors in the M1 Pro/Max.

The only reasons we don't have GPUs with 1 TB of Ram today is due to impracticality and expense - not because it isn't desirable to have the GPU able to work on ANYTHING the CPU has access to as/when required without needing to do data copies around the place.

Even if the GPU itself doesn't need 1Tb, the whole point is that it can process stuff the CPU may need it to process, which could be anywhere inside the 1TB of memory the CPU has.



You even posted (quoting Apple which backs up my point) that the whole point is that the GPU can access the memory the CPU has. If the CPU has 1 TB (and that's not a massive amount in 2021) then the GPU needs access to that 1 TB for the whole unified model to work.

If you have a seperate island of memory (some fraction of the the 1TB for the CPU) that the GPU can't access in its entirety, you're back to copying data around which is inefficient.

:
(1) I said "Apple doesn't design its architecture for rare use cases." Afterburner isn't a redesign of its entire architecture for a rare use case; it's an add-on module offered for a rare use case. So I stand by what I said. Apple isn't going to design its entire underlying architecture (UMA) just for rare use cases.

(2) You raise an interesting issue about shared access. My assumption is that, in typical cases were the GPU needs access to the same data the CPU is using, that doesn't necessarily mean it needs access to all of it. So just because the CPU may be actively working on a 500 GB dataset that's loaded into RAM, it wouldn't typically be the case that the GPU needed access to that entire 500 GB dataset at once (in the way the CPU does). But, I acknowledge that's my assumption.

(3) More generally: NVIDA's higest-end card rendering/graphics processing card is the A6000. It comes with 48 GB GDDR6 RAM, and should have about the same processing power as a 128-core AS GPU (= 4X Max). Let's call that processing power "1 GPU compute unit". NVIDIA knows Hollywood's computing needs extremely well, and (I assume) certainly would have the capability to produce a rendering-class GPU with far more than 48 GB VRAM, yet they don't.

I.e., if it really would be a general benefit to their customers' to offer, say, 480 GB VRAM/GPU compute unit, I would think they'd do it. The fact that they don't suggests it's not of general value, which is why I though it might provide benefit only in more specialized cases. [And those specialized cases might be quite interesting.]

 

[Boil]

I am not buying in these cube rumours.

The Mac Pro is an expandable MAC, so I think Apple might keep that.

Else, the Mac Pro is just a beefed up Mac Mini?

Why not both...?

A full Mac Pro Tower for those who need / want PCIe slots...

A Mac Pro Cube for those who don't...

 

[throAU]

Just admit it mate, you didn't think it through.

What I propose is not an architecture redesign any more than afterburner is

 

[Boil]

NVIDA's higest-end card rendering/graphics processing card is the A6000. It comes with 48 GB GDDR6 RAM, and should have about the same processing power as a 128-core AS GPU (= 4X Max). Let's call that processing power "1 GPU compute unit". NVIDIA knows Hollywood's computing needs extremely well, and (I assume) certainly would have the capability to produce a rendering-class GPU with far more than 48 GB VRAM, yet they don't.

I.e., if it really would be a general benefit to their customers' to offer, say, 480 GB VRAM/GPU compute unit, I would think they'd do it. The fact that they don't suggests it's not of general value, which is why I though it might provide benefit only in more specialized cases.

AMD had the Radeon Pro SSG GPU back in the Vega days...

I am wondering, as I have said above, if Apple will go the route of ECC DIMM slots for the UMA RAM (mounted on the back side of the mobo), rather than RAM chips adjacent to the SoCs...?

Sixteen ECC DIMM slots & eight M.2 NVMe SSD slots; 2TB UMA DDR5 ECC RAM & 32TB NVMe SSD storage...!

 

[theorist9]

Just admit it mate, you didn't think it through.

What I propose is not an architecture redesign any more than afterburner is

Ah, now we get personal again. I suppose I shouldn't be surprised, given my last interaction with you. Was hoping it would be better this time. Guess not. Anyways, I'm not the one who didn't think things through here. What's really going on is you lack the capability to address my points (likely because you can't understand them), and now you're just throwing up a smokescreen.

Do me a favor and please don't comment on my posts in the future. It just wastes my time.

 

[UBS28]

A cube and a tower can coexist as they fill different niches.

The Mac Pro line is not the money maker of Apple, so I don't think Apple will put alot of effort in updating these Mac Pro. It will just be 1 Mac Pro like how it currently is.

Apple didn't release a 15" MacBook Air next to the current MacBook Air, despite the MacBook Air being their best selling MAC.

 

[Melbourne Park]

On the memory - I've seen reviews of the Macbook Pro M1 Pro with 16 GB RAM, versus 32 GB ... not the Max version ... and the task performance was very similar. One reason mentioned was the speed of the SSD. It seems that for most tasks with a Macbook Pro 16 GB memory still does the job, with little time penalty ... sounds crazy but there are some arguments from work use tests that unified memory is different from the previous separate GPU environment.

It's possible that if the pipeline which includes using the SSD virtual memory to the CPU and the GPU, if that pipeline is not the bottleneck, lack of RAM memory will not slow things down .

 

[Appletoni]

He makes a very compelling case for the iMac Pro & Mac Pro's Apple Silicon chips.

I still feel bad for 2019 Mac Pro owners. That desktop should have debuted in 2017 instead of the iMac Pro so that owners enjoys over 5 years of use being phased out.

I need this chip inside the MacBook Pro 16-inch.

 

[throAU]

Ah, now we get personal again. I suppose I shouldn't be surprised, given my last interaction with you. Was hoping it would be better this time. Guess not. Anyways, I'm not the one who didn't think things through here. What's really going on is you lack the capability to address my points (likely because you can't understand them), and now you're just throwing up a smokescreen.

Lol.

I'm not making this personal, just implying that you posted something that refuted your own point (the bit I quoted) and you should just admit you're wrong rather than doubling down. I have no beef with you - just the point you're trying to argue.

You haven't addressed my point re: the requirement to copy if you have isolated memory pools. e.g., a portion of memory the CPU has that the GPU can not access.

Your other points fall down entirely (and that's why I didn't bother to explicitly address them) due to the requirement to perform copies with what you propose (non-unified memory) - which defeats the whole purpose of Apple's unified memory architecture's aim to eliminate copies.

Nvidia engineers have stated that the copying around is such a problem that, paraphrasing - "compute is essentially free, memory access is expensive". I wish I could find the video again, it was from 2016 and I believe they were talking about advances in Pascal and what they were doing to reduce data transport requirements in the GPU.

This is why Apple have gone to a unified memory architecture (and why they are making such a big noise about it) - and they are going to be doing anything they can to preserve that paradigm in order to maintain its advantages.

Bolting on pools of memory for the CPU, that the GPU can't access defeats that.

Now - if you have such a problem with these discussions with me - put me on ignore. That's what the button is for.

I'm not going to go trying to remember every person I have had a disagreement with on this forum; I reply to posts as I see fit.

I have no memory of any prior contact with you.

 

[iPadified]

The Mac Pro line is not the money maker of Apple, so I don't think Apple will put alot of effort in updating these Mac Pro. It will just be 1 Mac Pro like how it currently is.

Apple didn't release a 15" MacBook Air next to the current MacBook Air, despite the MacBook Air being their best selling MAC.

Very possible that it will be one large Mac Pro due to economics but then I do not see the need for a smaller MP as the video suggested. There was some nice engineering done with the current tower and to ditch that after one generation seems odd.

Apple should provide 14 and 16 inch airs for those who need the screen real estate but not the compute performance. It should be room for that in the laptop market. Similarly they should have a low-cost 12.9 inch "iPad". Performance and screen real estate are two completely different parameters.

 

[Adarna]

I need this chip inside the MacBook Pro 16-inch.

Don't we all?

 

[quarkysg]

I need this chip inside the MacBook Pro 16-inch.

Maybe not the 16”, but maybe an 18” instead? Would be 3” thick tho. and weigh 5kg. Haha.

 

[Melbourne Park]

Very possible that it will be one large Mac Pro due to economics but then I do not see the need for a smaller MP as the video suggested. There was some nice engineering done with the current tower and to ditch that after one generation seems odd.

Apple should provide 14 and 16 inch airs for those who need the screen real estate but not the compute performance. It should be room for that in the laptop market. Similarly they should have a low-cost 12.9 inch "iPad". Performance and screen real estate are two completely different parameters.

Just a Pro version of the Mac Mini - that is inevitable IMO. And why not a duel CPU bigger Pro version ... with a PCI slot I guess. And Apple don't make displays ... LG and Sharp supply Apple displays for the ipads and Macbooks including the new mini LED screens. Apple could make a larger margin by not selling a screen ... as they no doubt do with the mini. But with Pro processors, Apple's margin would jump up ... and even more with duel Pro chips. If the cost of GPUs stays high - and they might still be high when the Mac Pros are introduced - Apple could increase their market share as well.

 

[singhs.apps]

(3) More generally: NVIDA's higest-end card rendering/graphics processing card is the A6000. It comes with 48 GB GDDR6 RAM, and should have about the same processing power as a 128-core AS GPU (= 4X Max). Let's call that processing power "1 GPU compute unit". NVIDIA knows Hollywood's computing needs extremely well, and (I assume) certainly would have the capability to produce a rendering-class GPU with far more than 48 GB VRAM, yet they don't.

At present they allow 2 way interlink (NVlink) between GPUs so total Ram capacity can reach 96GB with 48 Gb per card.
With pressure from the likes of AMD, Intel..and perhaps Apple, expect Nvidia to raise the vram on their pro GPUs but perhaps, as has been discussed here quite a few times already, Intel’s CXL may come to the rescue wherein for offline renders, a shared memeory access may become a reality.
The only real reason onboard memory will continue to exist for nvidia’s discreet GPUs is for viewport performance and..games.

I don’t have access to w6800x duos, but do they not also come with something like nvlink ? If so, do renderers see 128 GB ram ? Can someone shed light on this ?

 

[Serban55]

I need this chip inside the MacBook Pro 16-inch.

no space into that laptop(not even in the next bigger imac)...into a mac mini we could saw it..since the mac mini doesnt need batteries inside
an 16" with this chip...

 

[GuruZac]

• Large scale 3D rendering. Complex models can consume a lot of RAM.
• Scientific modeling with large datasets.

I’m sure there are others

So essentially tasks that less than 5% of MacPro customers will ever do. I think for the vast majority laptop and desktop users Apple Silicon offers a tremendous number of advantages. But that doesn’t mean there isn’t a place for Intel, especially with some of their new chips. I’m glad they appear to have woken up