I don't think anyone is ditching SoC, it has too many advantages. But there's no reason to not expand on it outside of the SoC... you're never going to get everything into one chip if you want the power of a true workstation or server.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
What to Expect From Apple's M4 Ultra Chip Next Year
- Thread starter MacRumors
- Start date
- Sort by reaction score
I don't think anyone is ditching SoC, it has too many advantages. But there's no reason to not expand on it outside of the SoC... you're never going to get everything into one chip if you want the power of a true workstation or server.
Everybody is going to have to do this eventually as process tech is getting far more expensive to scale, and dies are getting to the physical size limit of what can be produced at reasonable cost.
The first ultra was a symptom of that, the max series of chips is already a really big die. As AMD have shown/talked about, not everything on the chip scales effectively with process shrink either so inevitably it is going to be more cost effective and efficient to split the SOC up into more different dies and link them together via a high speed fabric.
Would not surprise me if the M4/M5 Ultra is a number of dies instead of just two max chips - maybe a switching fabric die (like the AMD IO die), GPU dies and CPU dies.
I'm talking about desktop and workstations. It doesn't really take advantages with SoC.
Number of dies will only hurt Apple. Max itself is already huge and expensive to mass produce. And Ultra chip has way more expensive, low yield, and massive die size. Connecting two Max chips isn't really simple but complicated.
They'll eventually need to find a way to reduce the manufacturing fee and increase the yield by not using SoC design. But we'll see after M5 series.
All the Mac desktops currently use SoCs and there are plenty of advantages for desktops and workstations. They're just too limited compared to the old way of doing things and they don't scale well. I think a hybrid system is needed.
It may be complicated, but it's easier than building dies 2x the size of a Max - if not now, it will be soon.
There are practical limits to die size, especially within a desired cost/yield.
A chiplet-based desktop/workstation should allow Apple to really ramp up the GPU power, exciting...!
What do you all mean by SoC? System on Chip is more of a concept than a thing. It is a continuation of the ongoing process since personal computers came into existence: putting more and more logic into the CPU itself. The only thing different now is:
1) the memory is also on the same package (for bandwidth and latency reasons).
2) we're now moving to break up the monolithic CPU chip (like Intel and AMD already are).
If you mean the lack of dGPU, that has nothing really to do with the SoC, and Apple could choose to support dGPUs if it wanted to... Note: Apple doesn't want to, especially for the display pipeline.
It should also be pointed out that Apple is already using advanced chiplet packaging (“3DFabric”) in the M1/M2 Ultra. Their chip-first approach isn’t rumored to have changed to chip-last, so it appears they have built on that foundation for M4/M5 Ultra. We’ll just have to wait and see if they have done something differently, and the Mac Studio Brava “variation” and Mac Pro rumors suggest something is up, but it won’t be a fundamental change. There is nothing to indicate that. It does mean it’s possible Apple has started to get creative with the approach they are already using. Indeed, it almost seems likely.
Edit to add: [1] I use a past tense for much of the above because these choices were made more than two years ago, and [2] there is nothing in the chip-first approach that precludes using different chiplets. They don’t have to be identical like they were in M1/M2 Ultra.
Last edited:
good point.
Ultra Fusion is definitely a “fabric”, and the Ultra package is kinda a “chiplet”.
Yes, it’s not semantics, the M1/M2 Max with the silicon interconnect is a true chiplet—just because it doesn’t require the silicon interconnect to function doesn’t diminish that status.
Note: I learned this from Leman and others in the M4+ Speculation thread, where every aspect of advanced packaging has been discussed in detail (although not without going down some pretty convoluted rabbit holes and other sorts of dead ends).
It's literally all in one that you have to make it without issues. One tiny issue means you have to trash it. That's the con for SoC especially for bigger dies.
CPU, GPU, NPU, and other chips aren't chiplet and connecting two chips via Ultra Fusion is such a painful process for high fees and low yield.
Time will tell what Apple's chiplets will look like. Apple doesn't mind things being expensive.
Oh, I disagree with that definition of SoC. It might by implied by the words, but I don't think "Monolithic die" vs "chiplet dies on interposer" has anything to do with whether something is an SoC.
Personally, what makes something an SoC is the integration of all the I/O into the singe package that runs the machine. It is getting rid of NorthBridge and SouthBridge chipsets that makes something an SoC, not the number of dies in the package.
Definitions are messy, so I'm sure some would disagree with me.