The other aspects of a potential 'M1X'

[EntropyQ3]

For iMacs and Mac Pros, I would think it's unlikely Apple will go with higher bandwidth memory, e.g. HBM2, as it'll be too cost prohibitive to implement for consumer products.

I would argue that this isn’t necessarily true.

AMD didn’t believe so either for their higher end desktop graphics cards, they sold dual HBM stack cards at $399, and I bought mine for less. And that was the total cost to the end user. Silicon die at 500mm2, interposer, dual HBM stacks, PCB, power circuitry, cooler, profit margins...

The ballpark assumption of the additional cost to implement that memory system was roughly $50, subject to the usual business vagaries. It seems a bit high to me if we assume reasonably high volumes, but lets use it as a ballpark estimate for back of the envelope calculations.

The problem for AMD was that the benefits of HBM in terms of bandwidth and power draw didn’t provide enough value in their target desktop gaming market. The added bandwidth wasn’t enough to allow them to demonstrate higher performance than Nvidias competing offerings, in fact they had to clock the Vega chips to levels that also utterly obliterated any power advantage HBM brought to the table, in order to be competitive! At that point, the added cost of HBM didn’t help AMD at all.

But AMDs situation with their desktop gaming cards is not the same as Apples with their complete systems. Added cost is spread over the entire system, and is a much smaller part of the whole. There is no competitor that can sell competing Macs at some tens of dollars less, plus Apple targets a demographic that actually values cool running, quiet and svelte computers. The benefits of HBM can translate into value for Mac users, in a way it didn’t really for AMDs graphics cards. (Also, HBM is scaleable. A single stack on a small interposer to six (or even eight) stacks on a huge one. I can’t see much reason for Apple to go beyond at most two for even their most performant consumer systems.)

We’ve been this way before in our discussions, but the lack of any kind of input data makes it difficult to progress further. If we had a solid (hah!) rumour of a really wide GPU from Apple, well that would imply a redesigned memory subsystem. Or information from the manufacturing side, say if Hynix landed a big order for HBM that necessitated some expansion at some facility, which could be speculated to be due to Apple.

But we have none of that. Which I interpret to mean that if Apple has anything other than relatively trivial extensions of their mobile memory designs, they are probably not due for release in the very short term. Maybe.
Speculation is most interesting when there are tantalizing pieces of the puzzle available, and you need to fill in the blanks to connect them and/or figure out which pieces to discard. When the slate is just blank, well....

 

[quarkysg]

I would argue that this isn’t necessarily true.

AMD didn’t believe so either for their higher end desktop graphics cards, they sold dual HBM stack cards at $399, and I bought mine for less. And that was the total cost to the end user. Silicon die at 500mm2, interposer, dual HBM stacks, PCB, power circuitry, cooler, profit margins...

Well, AMD probably only needed smaller amount of VRAM storage for their graphic cards (at most 32 GB?), while for Apple in the Mac Pro’s instance, may need to support up to 1 TB or more. i would think this would be really costly. Also, I would imagine laying 1024 bit data bus traces to be a humongous challenge, if Mac Pro’s memory is to be user upgradable. Also, are HBM RAM module available for sale? I don’t seems to see such modules available for consumer sale.

Apple may want to go with wide data buses. So the question would be how wide is Apple going to go.

 

[leman]

But AMDs situation with their desktop gaming cards is not the same as Apples with their complete systems. Added cost is spread over the entire system, and is a much smaller part of the whole. There is no competitor that can sell competing Macs at some tens of dollars less, plus Apple targets a demographic that actually values cool running, quiet and svelte computers. The benefits of HBM can translate into value for Mac users, in a way it didn’t really for AMDs graphics cards. (Also, HBM is scaleable.

Case in point: Apple is the only company (I think) that offers HBM2 memory in a laptop solution, albeit at a significant premium — and people buy it. Savings from moving to their own silicon will likely allow them to spend more money on the memory subsystem without decreasing the profit margin — and make much better computers in the process.

Apple's RAM interfaces have been already described as "poor man's HBM", as they use a simple interposer-like structure to connect RAM to the SoC, putting everything on the same package. I can imagine this approach scaling.

Well, AMD probably only needed smaller amount of VRAM storage for their graphic cards (at most 32 GB?), while for Apple in the Mac Pro’s instance, may need to support up to 1 TB or more. i would think this would be really costly. Also, I would imagine laying 1024 bit data bus traces to be a humongous challenge, if Mac Pro’s memory is to be user upgradable. Also, are HBM RAM module available for sale? I don’t seems to see such modules available for consumer sale.

HBM and friends are by definition tightly integrated, so I can't imagine them being slotted. There is no slotted LPDDR RAM either... I suppose that Apple could use standard DDR5 modules — just with a lot of memory channels, but the mainboard complexity would be a nightmare and upgrading memory will be problematic as well (imagine a 16-channel Mac Pro RAM where each module has to be identical).

There is also a distinct possibility of custom RAM expansion modules, but how would that work? Just imagine an expansion module with 2000 pins

Anyway, curious to see what they will come up. I have a suspicion that upgradeable memory might be a history and a Mac Pro will instead use multiple "processing boards" each of it contains the SoC (CPU+GPU) and tightly integrated RAM.

Apple may want to go with wide data buses. So the question would be how wide is Apple going to go.

I think that if we are willing to accept some basic premises, this becomes an exercise in arithmetics. Let's say that Apple intends on keeping true unified memory and that they want to match the existing Mac Pro. Vega II Duo has 1TBs of RAM bandwidth. A single DDR5-6400 module (64-bit bus) offers around 50GB/s, so to reach 1TBs you will need 20 of those in parallel or a 1280-bit memory bus. At the same time, 20 DDR5 modules would take a ridiculous amount of space, so you'd want to do at least some sort of stacking with multiple memory challenge per stacked package. Which again leads us to HBM... but now I am in a realm of fantasy

 

[Boil]

Don't forget the rumors of 64-core & 128-core Apple GPUs...?!?

 

[EntropyQ3]

Well, AMD probably only needed smaller amount of VRAM storage for their graphic cards (at most 32 GB?), while for Apple in the Mac Pro’s instance, may need to support up to 1 TB or more. i would think this would be really costly. Also, I would imagine laying 1024 bit data bus traces to be a humongous challenge, if Mac Pro’s memory is to be user upgradable. Also, are HBM RAM module available for sale? I don’t seems to see such modules available for consumer sale.

Apple may want to go with wide data buses. So the question would be how wide is Apple going to go.

I think you put the finger on the conundrum Apple faces. IF you assume that GPUs will use the same memory as the rest of the system (which I do, but that doesn’t make it true), AND you assume that the GPU will perform ”up there”, then something needs to happen with the memory subsystem.
I’d argue that prioritising memory speed vs. memory size is the way to go. Memory speed benefits all users, memory size only those that need really large amounts of RAM, which in this day and age of 5GB/s secondary storage can’t be all that many.
Leman has proposed very wide DIMM-based solutions, which is a possibility if you want to retain the option of really large RAM pools. (Or is that RAM oceans? ;-))
For my brand of scientific computing speed is preferable to size, and that goes for everything that can be GPU accellerated today, including 3D-graphics.

Or they could ditch the unified memory model, or they could use HBM as an intermediate, cache like step in a tiered system that would hopefully be intelligently and semi transparently handled by the system. There are options. 2022 can’t come soon enough.

 

[dmccloud]

A lot of the speculation in this thread relies upon "traditional" notions of how the GPU and CPU interact with each other. Apple already took advantage of the fixed length of the ARM ISAs instruction set to go wide with their decoders, which is an approach AMD has already said wouldn't work on x86 because the variable length of the x86 instruction set creates much more complexity in the decoding pipeline. While AMD is still working on their version of unified memory, they are still beholden to the x86 architecture and how x86 systems access RAM and other system resources.

For Apple, I think the combination of UMA and going wide with the decoders is a better indicator of where they may go with future GPU solutions than anything AMD, nVidia, or Intel has done or is working on. The M1 is not only a departure from the x86 approach to system design, but a departure from the vast majority of ARM-based processors on the market, including those in Apple's own phones and tablets. (Yes, I'm aware that the M1 is based on A14 core designs, but the interconnects and internals of the SoC are very different from the A series).

If you haven't had a chance to do so, take a look at the iFixit teardowns of the M1 Macs. While the M1 Mini is the best example of just how much less space Apple Silicon logic boards can use, you can see just how much it could be shrunk in the MBA or MBP. In a larger machine such as the 16" MBP, iMac, or Mac Pro, Apple could actually go much larger with the SoC and logic board, possibly even running dual or quad SoCs (in the Mac Pro) and still have plenty of room left for cooling, peripherals, battery, etc. Apple built in the ability to scale up the M-series designs from the start, and these M1 machines are in some ways the proof of concept for Apple Silicon as a whole.

 

[leman]

A lot of the speculation in this thread relies upon "traditional" notions of how the GPU and CPU interact with each other. Apple already took advantage of the fixed length of the ARM ISAs instruction set to go wide with their decoders, which is an approach AMD has already said wouldn't work on x86 because the variable length of the x86 instruction set creates much more complexity in the decoding pipeline. While AMD is still working on their version of unified memory, they are still beholden to the x86 architecture and how x86 systems access RAM and other system resources.

Not quite sure where you are going with this. One can argue that variable-length ISA limits peak CPU performance (but even that's not quite clear-cut), but what does it have to do with the GPU or UMA? Fetching code from system memory would be extremely slow, that's why instruction caches exist. Not to mention that code size is negligible in comparison to data size, you don't need much system memory bandwidth to keep the CPU instruction buffers fed.

By the way, Apple GPU uses a variable-length ISA...

For Apple, I think the combination of UMA and going wide with the decoders is a better indicator of where they may go with future GPU solutions than anything AMD, nVidia, or Intel has done or is working on. The M1 is not only a departure from the x86 approach to system design, but a departure from the vast majority of ARM-based processors on the market, including those in Apple's own phones and tablets. (Yes, I'm aware that the M1 is based on A14 core designs, but the interconnects and internals of the SoC are very different from the A series).

I can't say I agree. The fundamental architecture of M1 is the same as various players (Intel, AMD and Apple itself) have been using in SoC designs for a long while. You have your processors (CPU, GPU, whatever) connected to a shared SoC-level cache, which is fed by system memory controllers. M1 package technology is identical to what Apple has been using on the iPad for some years now. What sets M1 aside from some other microarchitectures is the sophistication of it's internal memory subsystem — it has very high memory level parallelism (making it more GPU-like in nature) and can sustain very high bandwidths to individual cores. But then again, that's also a property fo the Apple mobile A-series.

 

[JohnnyGo]

I feel like Apple will release a SOC with two RAM sizes: 32Gb or 64Gb.

If you guys are right (and believe me, all of you seem to be much more knowledgeable than me), Apple will use 4 modules (4x8Gb, or 4x16Gb) with should automatically double the current memory bandwidth.

Question: would 4x32Gb possible with available production memory chips today?

I also expect Apple to jump into LPDDR5 or DDR5 with the upcoming SOC. if this would be an M2 or M1X I don’t know.

Anyways, as most of you pointed out RAM speed / memory bandwidth improvements will allow for a huge jump in performance.

 

[Argon_]

I feel like Apple will release a SOC with two RAM sizes: 32Gb or 64Gb.

If you guys are right (and believe me, all of you seem to be much more knowledgeable than me), Apple will use 4 modules (4x8Gb, or 4x16Gb) with should automatically double the current memory bandwidth.

Question: would 4x32Gb possible with available production memory chips today?

I also expect Apple to jump into LPDDR5 or DDR5 with the upcoming SOC. if this would be an M2 or M1X I don’t know.

Anyways, as most of you pointed out RAM speed / memory bandwidth improvements will allow for a huge jump in performance.

I think they'd go up to 8 modules for a 64gb SoC, to better leverage the GPU.

 

[Boil]

Forthcoming (smaller) MacPro to have three RAM options...

64GB / 128GB / 256GB

All HBMnext (aka HBM3)...

 

[Varmann]

Forthcoming (smaller) MacPro to have three RAM options...

64GB / 128GB / 256GB

That was very optimistic!
It is Apple, if we are lucky they start at 16 GB.

 

[thenewperson]

That was very optimistic!
It is Apple, if we are lucky they start at 16 GB.

RAM already starts at 32 for the Mac Pro but sure

 

[quarkysg]

I did some digging on current Mac's design and this is what I found:

iMac 27" (2020): dual-channel DDR4 memory
iMac Pro: quad-channel DDR4 memory
Mac Pro (2019): six-channel DDR4 memory

Low end Apple Silicon Macs: non-upgradable dual-channel LPDDR4X (already released)

So I'm predicting the following :

High end mobile: non-upgradable quad-channel LPDDR4X
iMac 21.5": non-upgradable quad-channel LPDDR4X
iMac 27": upgradable quad-channel DDR5-4800
Mac Pro: upgradable six/eight-channel DDR5-6400

iMac Pro to be discontinued.

 

[leman]

High end mobile: non-upgradable quad-channel LPDDR4X
iMac 21.5": non-upgradable quad-channel LPDDR4X
iMac 27": upgradable quad-channel DDR5-4800
Mac Pro: upgradable six/eight-channel DDR5-6400

Unfortunately, with those configurations you can forget about good GPU performance. You will be topping out at 400GB/s on the Mac Pro — current Intel ones have up to 1 TB/s RAM bandwidth to the GPU.

 

[quarkysg]

current Intel ones have up to 1 TB/s RAM bandwidth to the GPU.

I think Apple's GPU architecture might not be as gluttony for bandwidth as existing AMD/nVidia GPUs. With a larger GPU cache and improved GPU driver, we may be surprised.

My thinking is also that although current top end Mac Pro have extremely fast GPUs with fancy VRAMs, it is still hobbled by PCIe3's 32 GB/s transfer rate between main memory and GPU VRAMs. Same is true for all dGPU Macs. In real world use, we may see Apple's UMA design shine thru. Apple may want to keep the ultra-wide data bus for future Macs.

In any case, the next Apple Silicon event cannot come fast enough.

 

[leman]

I think Apple's GPU architecture might not be as gluttony for bandwidth as existing AMD/nVidia GPUs. With a larger GPU cache and improved GPU driver, we may be surprised.

For gaming, yes, they need much less bandwidth. For compute work though, bandwidth is bandwidth. Apple might get away from offering less bandwidth on a consumer Mac, but they have to up the game on the professional ones.

And let's also not forget about the CPU cores... they also need memory bandwidth, and a Mac Pro has a lot of them.

My thinking is also that although current top end Mac Pro have extremely fast GPUs with fancy VRAMs, it is still hobbled by PCIe3's 32 GB/s transfer rate between main memory and GPU VRAMs. Same is true for all dGPU Macs. In real world use, we may see Apple's UMA design shine thru.

Depends on your application. For video editing, fast UMA is obviously worth gold.

 

[Varmann]

RAM already starts at 32 for the Mac Pro but sure

Yes, but they seem to be very greedy with RAM in the new M1 models. But let us hope

 

[leman]

Yes, but they seem to be very greedy with RAM in the new M1 models. But let us hope

Do they though? It's exactly the same limit as the Intel model and in return you get custom-packaged RAM with ridiculously low power consumption and low latency.

 

[Boil]

Pro desktops...
PCIe Gen4 / PCIe Gen5 / Proprietary Apple interconnect...
AMD Infinity Fabric / Architecture sort of deal to interconnect HBMnext memory pools...?
New Apple Mac Pro Cube modern version of old SGI O2 workstations...?
It's all about the vertical integration...?!? ;^p

 

[EntropyQ3]

While I’ve stated a preference for high speed, non-user expandable memory, as a consumer I have to say that Apples RAM prices gives me pause. They have a 1000% margin on the additional RAM for the M1 macs, (spotprice is $25.60, Apple obviously pay even less,) and it’s painful to even anticipate what they would charge for larger amounts.

Under those circumstances, being able to install third party RAM is quite valuable. Literally.

 

[thenewperson]

Do they though? It's exactly the same limit as the Intel model and in return you get custom-packaged RAM with ridiculously low power consumption and low latency.

Yeah, I don't think greedy is the right word. Stingy, maybe.

 

[crevalic]

I think Apple's GPU architecture might not be as gluttony for bandwidth as existing AMD/nVidia GPUs. With a larger GPU cache and improved GPU driver, we may be surprised.

My thinking is also that although current top end Mac Pro have extremely fast GPUs with fancy VRAMs, it is still hobbled by PCIe3's 32 GB/s transfer rate between main memory and GPU VRAMs. Same is true for all dGPU Macs. In real world use, we may see Apple's UMA design shine thru. Apple may want to keep the ultra-wide data bus for future Macs.

In any case, the next Apple Silicon event cannot come fast enough.

Apple silicon magic strikes again.

You want to load 16GB of data into RAM to analyze on an M1 Mac? Somehow takes only 8GB of RAM, right. How, you say? "Magic", "unified memory", "x86 bad bro".
You want to run some machine learning tasks on a couple hundred GBs worth of data using an AS GPU with a 1/10 the memory bandwidth of its predecessors? No worries, AS GPU architecture is "not as gluttony for bandwidth as NVidia/AMD" so it will only take like a hundredth of the time. How, you ask? "Magic", "data will teleport in the "larger" GPU cache bro", "GPU driver better", "x86 ba...I mean umm other GPU companies bad".

 

[Varmann]

Do they though? It's exactly the same limit as the Intel model and in return you get custom-packaged RAM with ridiculously low power consumption and low latency.

Yes, the low limit, but you have a max amount of 16GB. That limit has to be raised significantly for a MacPro model.

 

[crazy dave]

Yes, the low limit, but you have a max amount of 16GB. That limit has to be raised significantly for a MacPro model.

His point is that ram amount on the comparable units didn’t decrease with M1 chips. So, so far, there’s no reason to suspect that when Apple replaces models further up the chain that Apple will lower the available amount of memory you can have. However lots of questions for what they’ll do for iMac Pro and Mac Pro systems: how much ram they’ll put in and what type of ram it’ll be and how will it be installed? Apple themselves figuring these out are probably one reason (IO and big dies/chiplets being the other) why those machines aren’t slated to until the end of 2022 at the earliest. But yes of course they’ll need more than 16GB of ram at the top end for those systems, heck the M1X is reportedly coming with 32. And yes they’ll need more than that for the i/Mac Pro too.

============

On the subject at hand, I think LPDDR5 is unlikely for the M1X. My guess is if they were going to add it this generation they would’ve done it for the A14. Hopefully will be there for the A15/M2 series of chips.

 

[ArPe]

HUrRY uP IwAnT 16 inCh M2 nOW!!