I'm familiar with that thread, since I've made several posts there, e.g.:
I was referring to this thread:
In this thread, we intend to share all relevant details, resources, and processes involved with upgrading soldered storage on Apple Silicon Macs. Notes All M1 systems utilize BGA110 "S5E" NANDs Some M2 systems use BGA110, while some use BGA315 Table of supported NAND configurations SSD...
forums.macrumors.com
But yes I am aware you are aware as you posted in the other thread on this subject.
I was pointing out its relationship to what you wrote.
More broadly, you start by saying "Only if you buy from Apple", as if you were going to explain why you disagree with what I wrote about secondary RAM, but then get into a long digression about aftermarket SSD upgrades. So I'm afraid I'm not following what specific point you're trying to make about what I wrote.
Ah sorry if my response was confusing. I was responding specifically to the below:
After all, that's exactly what they did with the MP's upgradeable NAND--you have to buy the Apple-branded product, and pay Apple prices.
Sure Apple offers expensive upgrade kits but they don't block PCIe SSDs. If you've already bought the Mac Pro, you don't have to pay Apple prices if you don't want to and you still get internal storage. Further 3rd party upgrade kits for even Apple's slots could feasibly be made. It's just hard to make a compelling business case out of it as it only applies to the Mac Pro and the Mac Pro can already add as much storage as you want easily. You wrote that even Apple's MP upgradeable NAND supports the idea that they would ensure there would no 3rd party RAM if that was upgradeable and I argue that it shows the opposite.
1) I think CXL could be more useful than how you portray it, for someone that needs to do an operation on a block of data too large to fit in their primary RAM (whether than be UMA or traditional). Swapping to the secondary RAM would be much faster than swapping to the SSD. Indeed, if it didn't provide such a benefit, there would be no market for it, and Samsung wouldn't have spent the substantial developent dollars on it.
That's
exactly how I think it'll be used. I'm just saying that's not a replacement for someone needing lots of UM - after all we already see solutions like Grace Hopper where huge amounts of on-chip VRAM are needed even though it offers cache coherent memory across an even faster bus than PCIe 5. Just as there is a market and use case for CXL there would still be a market for large pools of high bandwidth GPU memory. Perhaps I downplayed it too much, but that's what I was trying to say.
2) If Apple offered secondary RAM, I'm sure Apple would want to stop people from using aftermarket solutions, since (unlike the case with PCIe storage) I imagine there would be significant development costs incurred in modifying their UMA to accommodate secondary RAM, and they would want to recoup those.
So this all theoretical and seems a little silly to disagree about something that doesn't and may never exist for the Mac. But both solutions, CXL and RAM sticks, discussed here would ultimately be similar to PCIe storage. Basically there would be 3rd party solutions and I disagree that the costs would be large enough that Apple would block them unless Apple went and developed a completely novel and bespoke system and even then there would likely be a market for 3rd party solutions if that market were big enough.
*At the same time, I don't see anything about UMA that would preclude it. Currently, they have a mechanism to swap to the SSD when there's not enough RAM. With secondary RAM, they could do the same thing, just with an extra layer: They would swap to the secondary RAM instead of the SSD—and only swap to the SSD when the secondary RAM wasn't enough. Or even if this isn't how they would implement it (e.g., for some operations, it may be more efficient to have the most-frequently accessed data be resident in the primary RAM, and the remainder be in the secondary RAM, and give the CPU and GPU direct access to both, rather than having them accessed hierarchically), I don't think UMA would provide a particular barrier to implementing secondary RAM.
Very much agreed: it would have to function like super-swap so it would be transparent to the user and not fragment the user base which as
@bcortens pointed out is a potential issue. At that point I agree I don't think it would be much of a challenge to add to UMA. But there may be wrinkles I'm not seeing.
I know the UMA argument very well, as I used SGI O2 workstations professionally, and own one now.
Now there's a blast from the past!
The primary trade-off is limited memory size in exchange for performance.
Ehhh ... only at the very highest end of memory size and even then technically Apple can and likely will increase RAM availability at the higher end. At most product tiers it's more of an exchange of cost vs performance.
They are also, in modern form, unexpandable.
For RAM? Currently true and admittedly likely to continue to be true. Having said that CXL and other related solutions may ameliorate that in the future.
I’m also quite familiar with the SoC concept, and know the origin with microcontrollers.
The performance and size advantages are clear, but they are by nature unexpandable.
It depends. Beyond RAM, they are less modular in the since you cannot pair a low end CPU with a high end GPU (chiplets may help with this) or replace just a component of the SOC at a time (but would not help with this) but in terms of expandability, the ability to add more stuff on top, then in theory that's quite feasible for many common PCIe applications. Apple's PCIe lanes are limited but that's more a function of the laptop-focused design of the Max and the lack of an Extreme SOC. Those aren't intrinsic to the idea of using an SOC. In fact truthfully most of the industry is moving towards SOC design but they aren't necessarily losing dGPUs
yet. You could even allow motherboard upgrades which would save the chassis and allow people to upgrade all the components at once - a higher one time price vs amortized over time, but functionally equivalent. Of course Apple doesn't do this, but there isn't any reason why they couldn't. Theoretically you could even offer a socketed SOC, Apple almost certainly wouldn't but in theory you could.
That’s not great for my personal use-case, but it also creates ewaste, which is why I am such a strong believer in modular design, in servers, desktops and laptops.
- Alex
I feel the e-waste argument is a little overblown unless you're simply referring to yourself and not wanting to add to the problem which is commendable rather than the market's behavior at large. I don't disagree that in the very tail end of the distribution of computer users a non-upgradeable chassis might have a shorter life, but the median/mean? All the market research points in the same direction that very little changes. Perhaps it shouldn't be that way but it is - though not all of it is for bad reasons: SSDs, modern RAM, and SOCs etc ... last longer than most people change computers and theoretically they, even on the used market, simply don't need upgrades over reasonable lifetimes for a lot of these devices. Obviously focusing on workstations is a little different but even then ...
As I said above, even with the SOC you could in theory offer upgrades like MB upgrades (and even more theoretically socketed SOCs) which would be functionally equivalent in this regard, just bigger individual jumps than smaller ones. Another unfortunate trend is that even in the PC space motherboards don't last as long as they used to (both in terms of their quality which has taken a rather disturbing hit in recent years and also AMD/Intel telling people that the new generation of CPU is backwards compatible with last generation's sockets is now a big deal and met with great elation as opposed to being expected). That's not an excuse not to do better of course.
Basically Apple could build the following kind of workstation:
chipset based SOC with UMA on a upgradeable motherboard
lots of PCIe lanes
CXL or related solution for a large pool of RAM which was transparent to the end user
that would recover 90+% of the modularity/upgradeability while retaining almost all of of the advantages of Apple Silicon. Sadly I'm not sure if Apple will ever build such a device, but a good first step would simply be getting an Extreme SOC!
