Apple Testing at Least Nine New Macs With Four Different M2 Chip Variants

[zephonic]

Allegedly, this is Apple’s CPU roadmap, Jade 2 C-die is the M1 Ultra. Jade 4 C-die would be what’s in the next MacPro.

https://forums.macrumors.com/proxy.php?image=https%3A%2F%2Fhypercritical.co%2F2021%2F05%2F21%2Fimages%2Fcity-of-chiplets.png&hash=b0439a31967c3b3c35e2814098a9279f

 

[chucker23n1]

Jade 4C-Die is probably canceled.

 

[EugW]

Allegedly, this is Apple’s CPU roadmap, Jade 2 C-die is the M1 Ultra. Jade 4 C-die would be what’s in the next MacPro.

https://forums.macrumors.com/proxy.php?image=https%3A%2F%2Fhypercritical.co%2F2021%2F05%2F21%2Fimages%2Fcity-of-chiplets.png&hash=b0439a31967c3b3c35e2814098a9279f

That was from Mark Gurman. So it would appear somebody made a nice pic with Mark Gurman’s leaked info.

Jade 4C-Die is probably canceled.

Yeah I was wondering about that. It will be interesting to see what the Mac Pro SoC turns out to be. The number of cores might be in the same ballpark but the design may be somewhat different for 4C than one would extrapolate from 2C.

 

[theorist9]

As others have noted, Apple doesn't discount product during it's active lifecycle and part of how they are able to do that is because nobody knows when a new version of that product is coming out so people are not encouraged to wait to buy the new model or get the existing model on closeout discount.

As John Gruber (who has been talking to Apple PR for a long time) says, they "don't like to explain themselves". There's maybe a little bit of Osborne Effect fear to it as well: if you know how soon a new model is coming, you might not bother getting the existing model. But if you don't, you're more likely to buy what's available today.

Sure, sure. And, practically, I agree with both of you that it's unlikely Apple will give timeline info. about the M2 MBP's at WWDC (which was the subject of my original post). But both of your points actually support the idea that they might. Here's what I had in mind, but didn't say explicitly: People are wondering if we'll get M2 MBP's this year or next. I think it will be next year (2023). So when I told the poster to whom I was responding that Apple might say something, I had mind it might be something like: "The M2 MBP's won't be released this year". Or "the M2 MBP's won't be coming until 2023." This effectively encourages people who are thinking about waiting, in the hope the M2 MBP might be released in 2022, to buy now.

I.e., when I was thinking Apple might provide some timeline info., I didn't have in mind that Apple would be saying when they would be available, but rather when they wouldn't (this year).

But, again, I agree with both of you that it's unlikely. I just thought it would be interesting to explore the possibility.

 

[theorist9]

I think, if feasible, it's in Apple's interest to have a yearly update cycle for the chips in their Macs. Let's assume chip speeds increase an average of 10%/year. It's better for sales, marketing, and branding if, in years 1, 2, 3, and 4, their chip speeds are x, 1.1 x, 1.21 x, and 1.33 x (yearly updates) instead of x, x, 1.21 x, and 1.21 x (bi-yearly updates). And, since the Mac and iPhone chips are based on the same architecture, if they can get the iPhone chips on a yearly update cycle that makes yearly Mac chip updates possible. But I don't know if Apple wants to invest the resources to update the chips yearly in all of its Mac products.

 

[Mcckoe]

Wonder if the chip will be referred to as “The Apple M Ducer“

 

[eddie_ducking]

Deleted

 

[eddie_ducking]

Sorry for the previous post, it was typed a while back while browsing in a different thread, no idea why it appeared in this thread.

I'm firmly of the option that the Mx cores will leapfrog the Ax designs. With the A13/M1 design it was deemed necessary or prudent to take 18 months to release augmented Mx chips/Macs based on the core architecture. Now that might change in the future, but for the next couple of years, especially if the supply constraints continue, the iPhone will continue to receive yearly updates, but the Mac, purely down to to the various flavours, will have to be staggered and that takes time ... you can't release in 2023 an M2 Ultra based on then A14 when the A15 has been "out" for 6 months.

 

[EugW]

I'm firmly of the option that the Mx cores will leapfrog the Ax designs. With the A13/M1 design it was deemed necessary or prudent to take 18 months to release augmented Mx chips/Macs based on the core architecture. Now that might change in the future, but for the next couple of years, especially if the supply constraints continue, the iPhone will continue to receive yearly updates, but the Mac, purely down to to the various flavours, will have to be staggered and that takes time ... you can't release in 2023 an M2 Ultra based on then A14 when the A15 has been "out" for 6 months.

M1 is based on A14. M2 is thought to be based on A15 or possibly A16.

 

[chucker23n1]

Sorry for the previous post, it was typed a while back while browsing in a different thread, no idea why it appeared in this thread.

I'm firmly of the option that the Mx cores will leapfrog the Ax designs. With the A13/M1 design it was deemed necessary or prudent to take 18 months to release augmented Mx chips/Macs based on the core architecture. Now that might change in the future, but for the next couple of years, especially if the supply constraints continue, the iPhone will continue to receive yearly updates, but the Mac, purely down to to the various flavours, will have to be staggered and that takes time ... you can't release in 2023 an M2 Ultra based on then A14 when the A15 has been "out" for 6 months.

You got your numbers slightly off, but besides that: if the M2 launches late this year, it’ll be a sibling to the A16, not the A15. In fact, that may explain why we haven’t see the M2 yet.

However, I don’t think it’ll leapfrog. There’s simply no reason for Apple to give their Mac SoCs* different core designs. Different additional features (e.g. Thunderbolt), yes, but the cores will be designed to scale up and down. They’ve benefitted hugely from that strategy. The Watch gets old e-cores. The Macs get higher-clocked p-cores.

 

[ric22]

Sorry for the previous post, it was typed a while back while browsing in a different thread, no idea why it appeared in this thread.

I'm firmly of the option that the Mx cores will leapfrog the Ax designs. With the A13/M1 design it was deemed necessary or prudent to take 18 months to release augmented Mx chips/Macs based on the core architecture. Now that might change in the future, but for the next couple of years, especially if the supply constraints continue, the iPhone will continue to receive yearly updates, but the Mac, purely down to to the various flavours, will have to be staggered and that takes time ... you can't release in 2023 an M2 Ultra based on then A14 when the A15 has been "out" for 6 months.

I think you might be right. They can't fall THAT far behind, surely. I was gonna say your numbers (A14/15/16) were a bit off, but someone just beat me to it.

 

[eddie_ducking]

DOH!, sorry, my bad, but the 1yr Ax vs 18 month Mx release schedule is going to come undone fairly quickly ?

 

[fourthtunz]

The rumor is at the next MacBook Air will be getting an even bigger screen, likely due to the notch.
The rumor is it’ll be going from 13.3 to 13.6 inches.
I was not talking about Apple going down to the 12 inch form factor, I was purely talking about in terms of thinness and weight.

Gotcha, the air is kinda heavy but not compared to the 15's.
It'll be interesting to see what the M2 version of the air brings, whenever it comes out.
I'll get that or the 14", Power wise the m1 air is all I really need but
I'd like more SSD, memory and ports for what will become my forever computer

 

[theorist9]

I'm wondering what the RAM configuration will be on the AS Mac Pro.

If, as predicted, it will be available in both Ultra and 2X-Ultra configurations, that would be 8 and 16 RAM sticks, respectively. I'm going to predict that, rather than using the LPDDR5 found in the AS MBP's and Studio, they will use desktop-class DDR5. This gives three advantages:

(1) While the LPDDR5 standard allows for 32 GB sticks, only 16 GB sticks are commercially available. This would limit limit the 2x-Ultra Mac Pro to only 256 GB (32 performance cores, so 8 GB/performance core), far below the 12 sticks x 128 GB/stick = 1.5 TB available for the 24- and 28-core Intel Mac Pro's (64 and 55 GB/core, respectively). Even with the upcoming 24 GB LPDDR5x sticks, that's still only a max of 384 GB. By contrast, the DDR5 spec allows for up to 128 GB/stick and, since manufacturers went up to that maximum value with DDR4, it seems likely they'd also offer it with DDR5. [Currently, they're up to 32 GB/stick with DDR5.]

(2) I believe DDR5 offers higher bandwidth than LPDDR5 (two 32-bit channels per DIMM for DDR5 vs. two 16-bit channels per DIMM for LPDDR5).

(3) DDR5 doesn't need to be soldered (LPDDR5 generally does), which means RAM could be user-upgradeable. [This doesn't mean that they will make it user-upgradeable; but given that DDR5 RAM can be swappable, and given that Apple is trying to avoid further negativity from their professional customers, I'd be surprised if they didn't do this with the Mac Pro.] This would be especially valuable if Apple wants to release the Mac Pro this year, when only up to 32 GB DDR5 memory sticks are available, since users would know they'd be able to upgrade to more memory when the 64 GB and 128 GB sticks come out.

 

[CWallace]

I'm wondering what the RAM configuration will be on the AS Mac Pro.

Speculation I have seen is it will be a mix of on-package (like current M-series SoCs) along with off-package DIMM-based memory to allow for massive memory pools.

The off-package memory is said to be very slow (slower than the Intel Mac Pro), so I could see the on-package RAM being 256GB to try and keep as much as the OS and application in the "fast pool".

 

[theorist9]

Speculation I have seen is it will be a mix of on-package (like current M-series SoCs) along with off-package DIMM-based memory to allow for massive memory pools.

The off-package memory is said to be very slow (slower than the Intel Mac Pro), so I could see the on-package RAM being 256GB to try and keep as much as the OS and application in the "fast pool".

Is the idea, then, that they're going to stick with LPDDR for the on-package RAM, to avoid any redesign from what they already have with the current chip, and then supplement this with off-package RAM? And also that the off-package RAM is going to be intermediate in speed between on-package RAM and SSD?

 

[chucker23n1]

Is the idea, then, that they're going to stick with LPDDR for the on-package RAM, to avoid any redesign from what they already have with the current chip, and then supplement this with off-package RAM? And also that the off-package RAM is going to be intermediate in speed between on-package RAM and SSD?

Yes and yes. They’re not gonna do 1.5 TiB of LPDDR5, and they probably don’t want to design their M2 Extreme to have entirely different memory controllers. Plus, even if they did want to, that would come at a big latency cost. At that point, integrating the GPU might not even be a good idea any more.

So my speculation was that you get DDR slots that you can opt into per-process. You lose performance but gain capacity.

 

[theorist9]

Yes and yes. They’re not gonna do 1.5 TiB of LPDDR5, and they probably don’t want to design their M2 Extreme to have entirely different memory controllers. Plus, even if they did want to, that would come at a big latency cost. At that point, integrating the GPU might not even be a good idea any more.

So my speculation was that you get DDR slots that you can opt into per-process. You lose performance but gain capacity.

By "entirely different memory controllers" are you referring to what would be needed to switch from LPDDR5 to DDR5 (since that was the question at hand)? If so, why would DDR5 have higher latency than LPDDR5? I tried googling a latency comparison, but couldn't find one.

Also, if they're going to add off-package RAM, aren't they going to need different controllers anyways?

 

[ric22]

It probably would be good to separate the CPU, GPU and RAM in a Mac Pro, but that's a lot of work for what will probably remain an extremely niche machine. Plus it will upset the 'bandwidth' crowd.

 

[CWallace]

By "entirely different memory controllers" are you referring to what would be needed to switch from LPDDR5 to DDR5 (since that was the question at hand)?

I think it is more that you would have a memory controller for the on-package RAM and one for the off-package RAM.

If so, why would DDR5 have higher latency than LPDDR5?

I don't think it is the RAM type that causes the higher latency, but the fact that the data has to run along much longer traces on the systemboard to get to the off-package RAM compared to the on-package.

 

[theorist9]

I think it is more that you would have a memory controller for the on-package RAM and one for the off-package RAM.




I don't think it is the RAM type that causes the higher latency, but the fact that the data has to run along much longer traces on the systemboard to get to the off-package RAM compared to the on-package.

I thought @chucker23n1 was explaining why Apple wouldn't follow my prediction of using on-package DDR5 in place of on-package LPDDR5 to achieve the needed memory capacity (same number of chips, just more RAM/chip). If Apple did that, they would need just one type of memory controller.

I suppose if the DDR5 chips are larger than the the LPDDR5 chips, then the package would need to be bigger, which means longer traces. However, Anandtech measured DDR5 RAM latency as ~14 ns and, relative to that, trace latency seems insignificant. E.g., suppose trace length increases by 1" as a result of going to DDR5 RAM. That adds only 0.2 ns to signal propagation time (using the rule of thumb that electrical signals travel at ~1/2 c). Is there something I'm missing?

 

[chucker23n1]

By "entirely different memory controllers" are you referring to what would be needed to switch from LPDDR5 to DDR5 (since that was the question at hand)?

In my hypothetical scenario, for starters, you need the memory controller to handle two different kinds of RAM. You also probably need to handle that at the OS level. Honestly, the UX kind of sucks; you'd be back to Mac OS Classic-like hittin Get Info on an application and asking it to use "the other" RAM.

But even if all RAM were DDR, not LPDDR, and if all of it were in slots, I don't think Apple's current approach would work. It heavily relies on parallelization. The way they achieve ~800 GiB/s on the Ultra, and presumably two or four times that on the M2 Extremageddon, is because they simply run every LPDDR chip in parallel. Unless they mandate that you not only equip all slots at all times but also with the same size each, that doesn't really work.

If so, why would DDR5 have higher latency than LPDDR5?

For multiple reasons, including simple physical realities. LPDDR chips are literally right next to the SoC. DDR slots are not.

Also, if they're going to add off-package RAM, aren't they going to need different controllers anyways?

Yes. Honestly, who knows what they'll do? None of the scenarios seem quite right to me.

 

[theorist9]

But even if all RAM were DDR, not LPDDR, and if all of it were in slots, I don't think Apple's current approach would work. It heavily relies on parallelization. The way they achieve ~800 GiB/s on the Ultra, and presumably two or four times that on the M2 Extremageddon, is because they simply run every LPDDR chip in parallel. Unless they mandate that you not only equip all slots at all times but also with the same size each, that doesn't really work.

So with 16 sticks for the 2X-Ultra, and options of 8/16/32/64/128 GB LPDDR5, the steps would be 128/256/512/1024/2048 GB RAM. Given the alternative is no upgradeability at all, I would think the pros wouldn't mind so much the option of purchasing the Pro with one of those five RAM options, knowing they have the ability to do aftermarket upgrades later if their needs expand, even if those upgrades require them to swap out all their RAM.

For multiple reasons, including simple physical realities. LPDDR chips are literally right next to the SoC. DDR slots are not.

Still not sure about latency. See my comment about the significance of latency due to trace length above. Also, real-world latency is loaded rather than unloaded latency, and that's affected by bandwidth, and doesn't DDR5 offer more bandwidth than LPDDR5?

 

[sublunar]

Speculation I have seen is it will be a mix of on-package (like current M-series SoCs) along with off-package DIMM-based memory to allow for massive memory pools.

The off-package memory is said to be very slow (slower than the Intel Mac Pro), so I could see the on-package RAM being 256GB to try and keep as much as the OS and application in the "fast pool".

That's the understanding that I am seeing, the package includes the RAM and the way its built suggests that 256Gb could be the spec minimum - any extra RAM added afterwards by the user would be traditional DIMMS and would be slower.

It's reminiscent of the old Amiga days where the whole custom chipset would have direct access to so called 'Chip Ram' with any extra ram being (ironically?) called Fast Ram which was accessible by the CPU only (CPU could access Chip RAM too).

Given the alternative is no upgradeability at all, I would think the pros wouldn't mind so much the option of purchasing the Pro with one of those five RAM options, knowing they have the ability to do aftermarket upgrades later if their needs expand, even if those upgrades require them to swap out all their RAM.

I think Pros would be fine with the ability to add extra RAM as required beyond the on package minimum. It remains to be seen what the performance penalty would be for adding all the extra RAM using a memory controller.

 

[chucker23n1]

I think Pros would be fine with the ability to add extra RAM as required beyond the on package minimum. It remains to be seen what the performance penalty would be for adding all the extra RAM using a memory controller.

That's assuming we see that at all. (Again, this would come with odd behavior, such as maybe having to choose the RAM per-process.)

We might also see a Mac Pro that's "simply" two to four M2 Ultras in a trenchcoat, and the ability to add PCIe cards, and that's it.