No dedicated video RAM on ARM Macs

[JacobHarvey]

Apple's on die GPUs are still very much performance parts. Let's remember that the A12X GPU with its 7 cores already is at the same power level as the Xox One S. Then remember that not only is the Mac Apple Silicon going to be a new family but it can be run at much higher power levels as a Mac does not have the same thermal restrictions as, say, an iPad.

It's very impressive from the A12X's integrated GPU but remember that the Xbox One S effectively has the same GPU performance of the original Xbox One that was released way back in 2013. That original Xbox One used a custom GPU that was equivalent to the raw performance of a mid-range GPU of 2013 (an eternity ago tech-wise).

I think it will be a number of years before Apple will challenge or surpass AMD and nVidia at the high end GPU-wise (if they do end up devoting lots of resources to try and topple them in high end GPU performance).

Regardless, the GPUs that Apple will have ready for the Macs will likely offer plenty of performance and they've never gone for high end 'gamers' as a target market anyway.

Their focus on extremely good performance per watt will likely be a winning strategy for the vast majority of mac users and others who need fast portable devices.

By designing their own silicon they'll also be able to include more custom hardware acceleration of functions relevant to workflows in programs commonly used on macs (e.g. video editing software etc.) to rapidly speed things up without ramping up power consumption too greatly.

I think this will be far more exciting and important than just chasing raw GPU performance demanded by gamers etc.

 

[throAU]

I very much suspect you'll see HBM in the higher end models. Probably as a cache between the SOC and main memory much like AMDs HBCC in the Vega series (on PC Cards at least, not sure the Mac drivers expose that functionality).

It makes sense. It will work with the SOC in a small form factor device and provide the performance required.

 

[JacobHarvey]

User-upgradeable RAM is gone and not coming back, especially on portable machines. Pursuit of higher performance, lower memory consumption and reliability is not really compatible with modular RAM...

I'd have to disagree that modular memory by itself makes a device less reliable. Memory does not commonly go bad and if it does you can easily swap a new stick of RAM at any time to fix an issue (I think from that point of view it can actually help with the reliability and longevity of the machine as a whole).

With soldered ram it makes it extremely difficult (and from the point of an average user pretty much impossible) to fix bad memory. I agree that soldered ram allows for lower power consumption and other benefits which are very important for portable machines.

Another clear benefit (from Apple's point of view) is that having soldered RAM has allowed Apple to effectively lock you in to paying extra for RAM up front if you need it or think you will need it.

Apple has massive profit margins on those RAM upgrades based on the ridiculous prices they charge in Macs today (even the iMacs that you can get standard SODIMMs for). So it clearly benefits Apple if users must pay them upfront for any memory upgrade (instead of being able to use SODIMMs like in previous times).

 

[leman]

I'd have to disagree that modular memory by itself makes a device less reliable. Memory does not commonly go bad and if it does you can easily swap a new stick of RAM at any time to fix an issue (I think from that point of view it can actually help with the reliability and longevity of the machine as a whole).

The issue is that sockets themselves introduce new points of failure. And since modular designs have to be compatible with a range of different modules, certain tolerances are possible that might have a slightly detrimental effect on overall reliability. I don't have any data on this though, so it remains only a supposition on my part. My reason for this thinking is that in all my time working with computers I have never seen soldered RAM fail, but I have witnessed multiple modular RAM failures.

 

[jinnyman]

The issue is that sockets themselves introduce new points of failure. And since modular designs have to be compatible with a range of different modules, certain tolerances are possible that might have a slightly detrimental effect on overall reliability. I don't have any data on this though, so it remains only a supposition on my part. My reason for this thinking is that in all my time working with computers I have never seen soldered RAM fail, but I have witnessed multiple modular RAM failures.

I disagree. The chance you are describing is not substantial enough to justify onboard ram installation and possible detrimental effect it has on overall reliability of system in case of failure. The socket is so reliable enough for machinen’s life time that what you describe as a problem is really not the issue. Apple instroduced onboard ram only to charge more margin on upgrade and reduce the overall thickness of machine. Not for serviceability, reliability, and environment.

 

[leman]

I disagree. The chance you are describing is not substantial enough to justify onboard ram installation and possible detrimental effect it has on overall reliability of system in case of failure. The socket is so reliable enough for machinen’s life time that what you describe as a problem is really not the issue. Apple instroduced onboard ram only to charge more margin on upgrade and reduce the overall thickness of machine. Not for serviceability, reliability, and environment.

You might be right. It would be nice to see some empirical data on this stuff (e.g. frequency of failures on RAM for pre 2012 MBP vs post 2012), but I doubt that there is anything publicly accessible...

By the way, speaking about environmental concerns, I believe the best way would be component-level repair (replacing the failed RAM module directly). If I remember correctly Apple did mention something like that long time ago, would be interesting to know whether they actually do component-level refurbishing on returned logic boards.

Modular RAM does incur significant resource overhead, and I won't be surprised if it is actually more environmentally friendly to replace the entire logic board on the rare occasion that RAM fails rather than have small overhead per DIMM board.

I hope that one day we can go fully integrated without losing the ability to repair individual components by using technologies that quickly allow one to identify the failed circuit and replace it without detrimental effects for the entire board.

 

[Danny82]

Nice discussion but all this are too technical for me.. would like to know, assuming the a12x is around gtx 1050 power/speed.. and that is a 2018 gpu design.. let's say the a14 has a 20% increment per year in speed.. plus if they make it into a laptop chip let's call it a14m, and giving it 8 power cpu core and 16 gpu core which is double the core court on a12z.. plus a HBM2 16gb.. more than this the price would go crazy.. would my expectation of it reaching at least desktop class RTX 2060 super be too much..? Because being Apple, they will make their gpu low power also, so if it does reach a desktop RTX 2060 super level at low power consumption.. I would be very happy already, especially if it is in the 14" mbp..

 

[leman]

Nice discussion but all this are too technical for me.. would like to know, assuming the a12x is around gtx 1050 power/speed.. and that is a 2018 gpu design.. let's say the a14 has a 20% increment per year in speed.. plus if they make it into a laptop chip let's call it a14m, and giving it 8 power cpu core and 16 gpu core which is double the core court on a12z.. plus a HBM2 16gb.. more than this the price would go crazy.. would my expectation of it reaching at least desktop class RTX 2060 super be too much..? Because being Apple, they will make their gpu low power also, so if it does reach a desktop RTX 2060 super level at low power consumption.. I would be very happy already, especially if it is in the 14" mbp..

With a 16 core Apple GPU and HBM2, performance levels of 2060 RTX at under 50 watts power consumption are realistic. Just my uneducated opinion of course

 

[Danny82]

Thanks, as believe u all know the technical expect of things better than I do, just wanted to understand what is realistic expectation or pure dreaming.. I do know the HBM2 is dreaming though.. haha.. but well.. 16 core gpu should be bare minimum we can ask from apple.. I do hope Apple surprise me with something over my simple expectation of lower apple gpu being as good as desktop class RTX 2060 super in the 14" mbp which I am aiming for.. but please dont surprise me with the price tag.. that's all I ask for..

 

[JacobHarvey]

Nice discussion but all this are too technical for me.. would like to know, assuming the a12x is around gtx 1050 power/speed.. and that is a 2018 gpu design.. let's say the a14 has a 20% increment per year in speed.. plus if they make it into a laptop chip let's call it a14m, and giving it 8 power cpu core and 16 gpu core which is double the core court on a12z.. plus a HBM2 16gb.. more than this the price would go crazy.. would my expectation of it reaching at least desktop class RTX 2060 super be too much..? Because being Apple, they will make their gpu low power also, so if it does reach a desktop RTX 2060 super level at low power consumption.. I would be very happy already, especially if it is in the 14" mbp..

It is truly difficult to figure out what Apple is able to achieve, Apple themselves claimed the A12X has the GPU performance of the Xbox One S (which had the raw GPU performance of a GTX 750 Ti or Radeon HD 7790 - mid range GPUs from 2013). Some others say that it is equivalent to a GTX 1050 based on the Geekbench score or TFLOPs figures.

But there is no easy way to compare Apple's iGPU performance in things like 3D games to GPUs on other platforms right now.

For people who just point to TFLOPs and say a similar figure makes 2 different GPUs equivalent (or higher TFLOPs making one better than another), that's frankly not a very good metric at all if you want to compare how GPUs will perform in 3D tasks like gaming etc. (rather than in scientific compute tasks).

As an example, a Radeon Vega 64 had quite decent floating point performance with 12.66 TFLOPs but it gets significantly outperformed by a nVidia RTX 2070 SUPER in 3D games with 'only' 9.1 TFLOPs or a Radeon 5700XT with its 9.7 TFLOPs. There are so many other things that are more important in contributing to how well a GPU will perform in 3D tasks.

Apple will likely allow for greater optimization of games and other programs in macOS for its GPUs to eke out smoother performance. But it will ultimately be up to developers to bring their games along to the platform and make use of those potential optimizations if they choose to develop for the new Macs.

Apple will also obviously have more powerful iGPUs in their macs than the A12X and A12Z. But there is a limit to how large they can make the silicon die that holds both the GPU and CPU (if they want good chip yields). Plus potential thermal issues arising from having such high performance GPU and CPU cores placed on the same piece of silicon (or very close together) are another concern, especially if they're targetting upper mid-range nVidia GPU desktop performance. Both of these factors could present significant challenges in terms of the overall performance level they will achieve (at least initially).

But ultimately, until we've got actual hardware to play with and comparable games to run across platforms and GPUs, I think it's just too difficult to speculate what exact nVidia GPU 'performance tier' Apple will be matching in the new ARM macs (when it comes to 3D performance and the end user experience).

 

[leman]

But ultimately, until we've got actual hardware to play with and comparable games to run across platforms and GPUs, I think it's just too difficult to speculate what exact nVidia GPU 'performance tier' Apple will be matching in the new ARM macs (when it comes to 3D performance and the end user experience).

True, and what makes it even more difficult is the fact that these GPUs operate very differently. In general, Apple GPUs will need to do less work drawing a scene than Nvidia or AMD GPUs. But depending on how the rendering pipeline is set up, they might quickly hit the memory bandwidth bottleneck - their biggest weakness.

My estimates of Apple GPU performance come from game-like graphical benchmarks. Unfortunately, newer state of the art benchmarks are not available for iOS. So we can only guess.

 

[macsplusmacs]

Maybe Apple just starts making its own discrete GPUs for their higher-end models.

 

[deconstruct60]

Remember that Apple's "ARM" is a LOT different than other implementations. They have an ARM Architecture license so while they use ARM instruction sets the cores are Apple's design. Add in that Apple poached a LOT of design talent from Intel, Qualcomm and even AMD and Nvidia over the past few years, especially from the team that designed "Conroe" (better known as Core2Duo et al) and it isn't that surprising that their A series SOCs have become beasts. Note that at WWDC they were very careful to say that we haven't yet seen the SOC they will be using on Macs - and the breakout sessions gave some hints but not a lot of details.

Apple's ARM arch license has little to directly do with GPUs. Apple didn't use ARM's GPUs even before they got an architecture license. They always were on PowerVR (imagination tech). They skipped ARM's GPU from the start. So their license coverage really doesn't have much to do with it.

Apple's arch license has allow them to come up with their own CPU core to GPU core interconnect and caching ("system memory" ) . Don't necessarily need a ARM core license to do that.

With that said, I would not be a bit surprised if they are preparing a design that has direct channels to the RAM from both the CPU and GPU blocks

Eh? That is what is pragmatically there now ( given that you'd have to take cache control into account) . It is shared unified memory access. The surprising thing would be separate direct channels for each of the CPU and GPU block to separate RAM subsets. That really isn't "both" (tying them together).
[ Two channels to memory which don't have any cache coherecy isn't an option. That wouldn't be surprising because it doesn't work on general workloads. Pretty unlikely to get something that is bound to fail. It just won't make it out the door. ]

this large central system memory thing that Apple has/is doing is a dual edge sword. At their relatively smaller core count levels it works pretty good. Can get all the cores pretty close and it is a "oversize' pot for them to share. The latency to this system cache can be relatively low. An order ( or two ) more magnitude more 'cores' haning off of it and may not work as well as the cores get farther and farther away ( and MMU TLBs get bigger if coverage much larger RAM capacities. ) .
[automerge]1593897742[/automerge]

Maybe Apple just starts making its own discrete GPUs for their higher-end models.

Pretty doubtful. Unless Apple was contemplating a grow like gang busters eGPU market there isn't really volume to support that in the Mac ecosystem. Nor is it likely ever coming to the A-series. The A-series iPad Pro line up is likely to remain the primary driver of what shows up GPU wise in the Mac series in terms of iGPU. Scaled a bit bigger perhaps but same basic building blocks.

Apple is focused on folks optimizing for Apple GPU probably because

1. it will be present in all Mac SoC ( all the way up and down the line up).
2. The native iOS/iPad OS apps will run against that very familiar iGPU. Apple wants them to run well so that build a deeper moat around iOS/iPad OS. ( also gets more GPU optimizer skilled iOS/iPad OS developers too if that becomes a common Mac dev skill too. )


Once developers are fully sunk tons of time into that then they have 'oh by the way" dGPU support in a fewer set of Macs than they do now. [ Primarily going to shove Intel iGPUs , which is the dominate GPU by units shipped , out of the way. Along with a few other dGPUs in some MBP 16" and iMac 21-24" (maybe all of them in a few generations). That will leave an even smaller set of macs with dGPUs at all. ]

 

[Kostask]

Because Apple can use a lego block approach to their SOCs, they can do what they want, literally. I can see SOCs with 4GB/8GB of on SOC VRAM, while still using off SOC DDR4/5 CPU RAM. I can see some other SOCs being made with various amounts of HBM2 on SOC. The iDevice SOCs already have on SOC LPDDR3/4 RAM, I don't see it as much of a stretch to incorporate either VRAM or HBM2 onto an SOC.

 

[Pressure]

Because Apple can use a lego block approach to their SOCs, they can do what they want, literally. I can see SOCs with 4GB/8GB of on SOC VRAM, while still using off SOC DDR4/5 CPU RAM. I can see some other SOCs being made with various amounts of HBM2 on SOC. The iDevice SOCs already have on SOC LPDDR3/4 RAM, I don't see it as much of a stretch to incorporate either VRAM or HBM2 onto an SOC.

You can't have HBM on die but only implemented via an interposer on the substrate, like VEGA10 and VEGA20.

 

[Yebubbleman]

Something that Apple was stressing over and over again during WWDC is that Apple ARM Macs will be based on their integrated system on a chip. Among other, it means that CPU and GPU share system memory (you may say that this makes the GPU integrated since having own video RAM is the primary criterion of distinguishing between iGPUs and dGPUs). This approach does have its benefits - there is no need to transfer data between CPU and GPU for example, the GPU can take advantage of the virtual memory system, the power consumption cost is much lower and it does work well for Apple GPUs who because of their architecture need much lower memory bandwidth than current dGPUs. However, this won’t scale to high-performance applications. Even if Apple uses LPDDR5 in their new Macs (RAM bandwidth approaching 50 GB/s), they won’t be able to complete with modern GDDR6 etc. solutions that deliver bandwidth of 200 GB/s and higher.

This is where my speculation starts. What if Apple kept the unified memory approach and its many advantages, but used high bandwidth memory instead? They already have a lot of experience with HBM2 and if I understand it correctly, it’s latency is comparable to the latency of regular RAM, so it can be used as CPU RAM (unlike typical video RAM that trades latency for bandwidth). Combining an Apple SoC with 32GB of HBM2 will allow bandwidths of over 400GB/s, which compares to those of fast desktop GPU, while also potentially allowing speed ups on the CPU side.

There are reasons why I think Apple could potentially pull this off. First of all, this kind of system is going to be very expensive (interposers are complex and cost a lot of money). This is probably why we don’t see it much in everyday computing as companies prefer more conservative solutions that scale to different markets. But Apple doesn’t care about this. They don’t have to cater to different markets, they have their target pretty much locked in. The 16” MBP already costs a lot of money - and they might as well funnel the savings from using their own chips into a more expensive memory subsystem. This would also be advantageous to Apple, since nobody else would be even close to offering anything even remotely comparable. This would be a very power efficient design, potentially capable of very high performance, and at the same time it would be in some ways simpler than the traditional PC designs (no need for different types of memory, no need for a bus between CPU and GPU, power delivery system can be radically simpler). A single SoC at 80watt TDP could potentially deliver desktop-class CPU and GPU performance.

Note: Unified memory architectures are used by gaming consoles, I assume in order to simplify the design and optimize the memory transfer. But consoles use high latency memory, so programmers have to take this into account.

What do you think?

I think you're trying to look at the Apple SoC like it's a PowerPC or Intel based Mac with a more traditional desktop computing system architecture and not like it's an SoC. I think many people are, and that's understandable because, as far as the Mac is concerned, that's what we all know.

However, if you watch the WWDC 2020 video on the Apple Silicon Macs system architecture, they explain that the GPU and CPU sharing the same system memory on the SoC is not even remotely the same sort of thing as what Intel chips have been doing since the first generation Intel Core i3, i5, and i7.

And considering that Developer Transition Kit Mac mini got more performance out of Final Cut Pro than any Intel based Mac mini could sans an eGPU, I think that their graphics are going to be fine...that is, so long as everything knows to take advantage of Metal.

 

[leman]

However, if you watch the WWDC 2020 video on the Apple Silicon Macs system architecture, they explain that the GPU and CPU sharing the same system memory on the SoC is not even remotely the same sort of thing as what Intel chips have been doing since the first generation Intel Core i3, i5, and i7.

Can you elaborate? Unified memory means unified memory. What other interpretation do you have?

 

[Waragainstsleep]

Can you elaborate? Unified memory means unified memory. What other interpretation do you have?

For a start, I think Intel CPUs with integrated graphics tended to 'rope off' some RAM from the system to be permanently reserved for the GPU (Thats certainly how Windows did it) while these new SoCs will dynamically allocate RAM back and forth as its required.

 

[leman]

For a start, I think Intel CPUs with integrated graphics tended to 'rope off' some RAM from the system to be permanently reserved for the GPU (Thats certainly how Windows did it) while these new SoCs will dynamically allocate RAM back and forth as its required.

If I remember Intels recent architecture papers correctly, the newer Intel GPU can access the same memory as the CPU. I don’t see a principal difference to the Apple SoC here. Maybe it was different back in the day.

P.S. The so called “CUDA unified memory” in turn is a misnomer. “True” unified memory means that CPU and GPU can access the same memory without any copies.

 

[PortoMavericks]

Unified memory doesn't have to be all equal. Take the Xbox Series X for instance, they're mixing three different GDDR6s with different bandwidth capabilities but it's still unified.

 

[sublunar]

For a start, I think Intel CPUs with integrated graphics tended to 'rope off' some RAM from the system to be permanently reserved for the GPU (Thats certainly how Windows did it) while these new SoCs will dynamically allocate RAM back and forth as its required.

I'm not overly familiar with how RAM is allocated on Intel iGPU these days but this discussion reminds me of the old days of Amiga Chip RAM vs Fast RAM. In a way, I am interpreting this to mean that all the system RAM will be available for video rendering or for CPU jobs without needing to copy graphics between SSD/RAM and VRAM.

The A12x/z in the iPad Pro 2018/20 is compared with the Xbox One S in terms of graphics capability but that old console used DDR3 RAM while the iPad Pro SoC uses LPDDR4X RAM (low power consumption variant).

While adding HBM2 (or GDDR6 RAM as the next generation consoles have done) may increase performance for gaming applications, there's a financial cost involved as well as potentially a heat penalty and a latency penalty which is where DDR4 and (later) DDR5 will win out because they manage much smaller chunks of data (unlike graphics cards which are constantly shifting the same big graphics). And there's potentially a scenario where all that HBM2 bandwidth is going to waste because there's no massive blocks of data being moved about all the time. That link suggests GDDR6 costs twice the price of DDR4 and runs hot. And DDR5 will be pricey to start with when it does appear.

I saw another link which suggests that HBM2 is pricier still - back when Vega was current graphics for them.

Apple are probably just happy using LPDDR4X RAM (for the laptops anyway) and can get a good price for it with their bulk buying. I would just make peace with the fact that next generation consoles are in their own stratosphere here with their custom memory design decision optimised for gaming jobs. It's probably not something that makes a whole bunch of sense for computer buyers otherwise elite PC gamers would have latched onto it and demanded motherboards that used GDDR6 or HBM2 as system memory by now.

This suggests that Apple have decided that 'adequate' GPU performance will do for now - something currently akin to the Intel Iris Graphics in 2019. They can buy common RAM at decent discounts as well as stable prices and pass the savings on (in their own way).

I'd like to see if Apple will offer desktop machines with upgradable RAM, nor have we seen what they can do with increased TDP budget in a bigger enclosure.

 

[leman]

While adding HBM2 (or GDDR6 RAM as the next generation consoles have done) may increase performance for gaming applications, there's a financial cost involved as well as potentially a heat penalty and a latency penalty which is where DDR4 and (later) DDR5 will win out because they manage much smaller chunks of data (unlike graphics cards which are constantly shifting the same big graphics). And there's potentially a scenario where all that HBM2 bandwidth is going to waste because there's no massive blocks of data being moved about all the time. That link suggests GDDR6 costs twice the price of DDR4 and runs hot. And DDR5 will be pricey to start with when it does appear.

I saw another link which suggests that HBM2 is pricier still - back when Vega was current graphics for them.

I was under impression that HBM2 does not have the latency penalty of GDDR, since it achieves the bandwidth by having a very wide interface instead of cranking the speed. Also, it’s power consumption is supposed to be very low. The only downside: it is indeed expensive.

 

[Waragainstsleep]

Apple are probably just happy using LPDDR4X RAM (for the laptops anyway) and can get a good price for it with their bulk buying.

I know the iPad Pro uses LPDDR4X, are we sure the DTK is using the same? I'm surprised no-one has at least leaked a pic of the unit with the base panel removed yet.

 

[awesomedeluxe]

What I think... is that Apple will definitely have a discrete GPU for most of its models, and that GPU will be made by Apple or AMD

Let's take the rumored APU at face value. 8x Firestorm cores, 4x Icestorm cores, some kind of iGPU. What is the purpose of the Icestorm cores? Of course the system is using the Icestorm cores to conserve power for less demanding tasks. But I think there is another reason for them being there.

That reason is to prevent the SoC from overheating when the iGPU is in use.

To get laptop-competitive performance from the A14 cores, Apple will want to clock them up as high as they can. Many people don't realize that Apple already pushes the limits quite a bit with the iPhone. I think they will be clocking them to the point where running all 8 Firestorm cores and the (presumably) 8 cores at the iGPU on the same time is just right out. That's why the Icestorm cores are there: they take over for four of the Firestorm cores when the iGPU is on. So I expect them to include a dGPU for pretty much all models down to the 14" in order to prevent the Firestorm cores from getting thermal constrained. Since the heat is now one inch that way it's a lot less likely to turn into a silicon puddle.

Now, could Apple pair 4 Firestorm cores to 8 GPU cores and two stacks of shared HBM2 memory on their 12"? Yeah, I think absolutely. But any other machine? No way. They will use LPDDR5 and have dedicated memory (quite possibly still HBM2E) on a separate graphics part.

That said I appreciate your thought process. I've thought about unified HBM2 too especially when Kaby G was around. I think on balance for most machines though it's just not worth giving up the ability to move all the heat generated by the GPU a little bit that way when things get real real.

 

[miscend]

Something that Apple was stressing over and over again during WWDC is that Apple ARM Macs will be based on their integrated system on a chip. Among other, it means that CPU and GPU share system memory (you may say that this makes the GPU integrated since having own video RAM is the primary criterion of distinguishing between iGPUs and dGPUs). This approach does have its benefits - there is no need to transfer data between CPU and GPU for example, the GPU can take advantage of the virtual memory system, the power consumption cost is much lower and it does work well for Apple GPUs who because of their architecture need much lower memory bandwidth than current dGPUs. However, this won’t scale to high-performance applications. Even if Apple uses LPDDR5 in their new Macs (RAM bandwidth approaching 50 GB/s), they won’t be able to complete with modern GDDR6 etc. solutions that deliver bandwidth of 200 GB/s and higher.

This is where my speculation starts. What if Apple kept the unified memory approach and its many advantages, but used high bandwidth memory instead? They already have a lot of experience with HBM2 and if I understand it correctly, it’s latency is comparable to the latency of regular RAM, so it can be used as CPU RAM (unlike typical video RAM that trades latency for bandwidth). Combining an Apple SoC with 32GB of HBM2 will allow bandwidths of over 400GB/s, which compares to those of fast desktop GPU, while also potentially allowing speed ups on the CPU side.

There are reasons why I think Apple could potentially pull this off. First of all, this kind of system is going to be very expensive (interposers are complex and cost a lot of money). This is probably why we don’t see it much in everyday computing as companies prefer more conservative solutions that scale to different markets. But Apple doesn’t care about this. They don’t have to cater to different markets, they have their target pretty much locked in. The 16” MBP already costs a lot of money - and they might as well funnel the savings from using their own chips into a more expensive memory subsystem. This would also be advantageous to Apple, since nobody else would be even close to offering anything even remotely comparable. This would be a very power efficient design, potentially capable of very high performance, and at the same time it would be in some ways simpler than the traditional PC designs (no need for different types of memory, no need for a bus between CPU and GPU, power delivery system can be radically simpler). A single SoC at 80watt TDP could potentially deliver desktop-class CPU and GPU performance.

Note: Unified memory architectures are used by gaming consoles, I assume in order to simplify the design and optimize the memory transfer. But consoles use high latency memory, so programmers have to take this into account.

What do you think?

Apple didn't announce that they will deprecate support for eGPUs in MacOS. So there is no reason to believe that Macs will no longer support AMD GPUs which always have their own dedicated VRAM.