Is a system with 2 or more M1 chips possible?

[anthonymoody]

People have been referring to the next/higher-end spec AS machines as things like M1X, M1Z, M2, etc. implying a stouter, higher-end chip replacing the M1, or an enhancement to it.

But it got me wondering whether it could be much simpler to drop two (or more) M1s into, say, a revised iMac? If everything scaled linearly, you'd have support for twice as many external monitors, twice as many ports, twice as much RAM, etc.

Caveat: I'm not remotely technical in regards to this. Sorry if this is a dumb question.

 

[Superhai]

Rarely have two or more CPU's in a general purpose computer worked out any good, and there are various reason for that. Apple may have a solution for those hidden somewhere, but I doubt so, and you will lose the benefit of tightly integrated SoC. Connecting two Macs via TB3 will probably be a better solution instead of a motherboard with two processors.

 

[Mr. Awesome]

That’s one possibility. I remember seeing someone post about how you can run these chips in parallel. I think the more likely option (at least for things like the iMac, 16” MBP, etc.) is that there is an M1X chip that’s more geared towards high performance than the M1 is. Maybe in a future Apple Silicon Mac Pro they’ll have multiple MX chips running in parallel, and that will be the best of both worlds.

To summarize:
- M1 is the current high-efficiency chip used in the thin and light laptops and base model Mac mini.
- M1X will be more performance-focused for use in more powerful configurations of the 13” MBP and the Mac mini, in addition to the 16” MBP and the iMac.
- M2 and M2X will be the next iterative versions of the M1 and M1X. The likelihood of a new 16” MBP getting the M2 chip as people are saying is close to nil. Instead, it’ll probably get an M1X chip, and future versions of the MBA, 13” MBP, and Mac mini will get an M2, which is the next low-power chip to replace the M1.

 

[richard371]

It would be more efficient and maybe even cheaper just to combine the 2 m1s into 1 chip or just double the core counts to 16 etc.

 

[cal6n]

No one knows outside of Apple. I’ll be very surprised if they haven’t built a way of parallelising multiple SOCs into the architecture.

 

[anthonymoody]

Rarely have two or more CPU's in a general purpose computer worked out any good, and there are various reason for that. Apple may have a solution for those hidden somewhere, but I doubt so, and you will lose the benefit of tightly integrated SoC. Connecting two Macs via TB3 will probably be a better solution instead of a motherboard with two processors.

Hmm I'm not so sure about that. There have been a fair number of high-end Intel-based workstations wherein you could configure multiple Xeon chips.

It would be more efficient and maybe even cheaper just to combine the 2 m1s into 1 chip or just double the core counts to 16 etc.

Maybe we're saying the same thing (or maybe it's a distinction without a difference). And/or maybe that's the nature of this architecture. They can just grow the chip by adding more cores or whatever else makes sense for a given deployment.

 

[bobmans]

That's a waste of silicon and is never going to happen.

Why would you need 32 NE cores?
Why would you need 2 ISP's?
Why would you want to double the amount of efficient cores on a machine that's plugged into your wall?
Why would you need 2 IO controllers?
Why would you need 2 secure enclaves?
...

That would be a big waste of silicon and transistor budget.

 

[NewUsername]

Didn’t the Power Macs have two processors? I also don’t understand why two processors in one computer isn’t popular anymore?

 

[anthonymoody]

That's a waste of silicon and is never going to happen.

Why would you need 32 NE cores?
Why would you need 2 ISP's?
Why would you want to double the amount of efficient cores on a machine that's plugged into your wall?
Why would you need 2 IO controllers?
Why would you need 2 secure enclaves?
...

That would be a big waste of silicon and transistor budget.

All else equal when starting from scratch, sure. But we're not starting from scratch. The way to answer your question is: two M1s (a chip which already exists and is rolling of the lines in quantity) *may* be less expensive to produce *at this time* than a beefier chip that has more of only the things you want more of.

 

[bodonnell202]

More than likely there’s a chip coming with additional CPU and GPU cores, and somewhat higher clock rates that will be used in higher end MacBook Pros and iMacs (I’m guessing the entry level iMac will start with the M1, but that lineup will all be revamped at the same time with a new design).

 

[Joelist]

You have to stop calling M1 a CPU - it isn't. It is a SOC and this is exactly why you won't see multiple M1 in one machine. It serves no purpose.

What you WILL see is a new SOC with more Performance cores, more GPU cores and more cache. Probably also additional controllers and the ability to have higher RAM levels in the package.

 

[curmudgeonette]

But it got me wondering whether it could be much simpler to drop two (or more) M1s into, say, a revised iMac? If everything scaled linearly, you'd have support for twice as many external monitors, twice as many ports, twice as much RAM, etc.

That's exactly how AMD is assembling higher core count Zen "chips". Ryzen (1) was one silicon die, Threadripper had two, and Epyc had four. The multiple dies communicated between each other using Infinity Fabric. Sure, this duplicated lots of I/O, but it also created a place to use dies with defective I/O modules.

 

[bobmans]

All else equal when starting from scratch, sure. But we're not starting from scratch. The way to answer your question is: two M1s (a chip which already exists and is rolling of the lines in quantity) *may* be less expensive to produce *at this time* than a beefier chip that has more of only the things you want more of.

You need more silicon for 2 M1 SOCs compared to 1 bigger SOC with additional CPU/GPU cores so 2 M1 SOCs will always cost more.
They've already done the R&D for their current generation CPU/GPU cores so there won't be much additional R&D costs for a bigger chip with the same cores (but more of them).

What we will likely see in the future tho is the M1 being replaced by binned higher-tier chips with the malfunctional cores disabled. To cut down on chip costs that's what you want to achieve. You want to produce one/multiple high-end chips (because raw costs aren't that high compared to R&D costs and there's hardly any difference in raw costs between different chips) and bin the ones with malfunctional / non-optimal performing cores disabled so you don't need to throw them away.

 

[leman]

Hmm I'm not so sure about that. There have been a fair number of high-end Intel-based workstations wherein you could configure multiple Xeon chips.

Long time ago, yes, before the industry figured out how to pack more cores on a single die. Today multiprocessor systems are basically delegated to server use only. Workstations like Mac Pro require a different performance profile.

 

[Strangedream]

I think the upgrade from the M1 will simply be an M1X with more performance cores (maybe from currently 4 to 8?). However, there's no need for a single machine to have multiple M1 - unless you want a workstation. Also I doubt Apple would be willing to create some software to support a multi-M1 configuration.

 

[dogslobber]

You could combine 2 Apple ARM chips together but it introduces bottlenecks. There is latency for ARM CPU A accessing the memory for ARM CPU B. If you don't design it right then there will be wait cycles for a memory access across CPU boundaries to be realized. This could make both CPUs slow as the Apple ARM chip gets its speed from memory on the die, and while an access is being realized across CPUs, the local CPU is being kept in waiting.

 

[anthonymoody]

Thanks all for the thoughts and explanations - very helpful and interesting!

 

[bill-p]

But it got me wondering whether it could be much simpler to drop two (or more) M1s into, say, a revised iMac?

No. As stated, M1 is an SoC (System on a Chip).

This is not like 1 + 1 = 2.

The M1 as an SoC is more like 1 + 2 + 3 = 6.

You're suggesting 2 of them, which means (1 + 2 + 3) * 2 = 12.

But if we're just doubling up on CPU cores and GPU cores, it can be (1 + 1) + (2 + 2) + 3 = 9.

And 9 is less than 12.

That's the simplest way I can explain it. Think about the next Apple Silicon chip as having more components than M1, but certain things stay the same. If you use 2 M1 chips, you're doubling up on stuffs that don't need to be doubled. Some of these are WIFI controller, Bluetooth controller, USB/Thunderbolt controller, audio chip, etc...

 

[Superhai]

Hmm I'm not so sure about that. There have been a fair number of high-end Intel-based workstations wherein you could configure multiple Xeon chips.

Yes they exist and for special purposes you can benefit from using them, but it is essentially smashing together two computers and some shared components.

 

[UltimateSyn]

Very interesting discussion, everyone! Thanks for making this thread, OP. I had been wondering about the pros/cons of that sort of implementation myself.

 

[MevetS]

Rarely have two or more CPU's in a general purpose computer worked out any good, and there are various reason for that. Apple may have a solution for those hidden somewhere, but I doubt so, and you will lose the benefit of tightly integrated SoC. Connecting two Macs via TB3 will probably be a better solution instead of a motherboard with two processors.

Hmmm ... my PowerMac Dual G5 was a very nice machine. As were the dual processor 2010 MacPros so I understand.

 

[Fomalhaut]

Didn’t the Power Macs have two processors? I also don’t understand why two processors in one computer isn’t popular anymore?

Yes, there have been dual-socket Mac Pros (PowerPC and Intel), and this is still quite common in Intel Xeon workstations, and *very* common in server hardware, with Intel supporting up to 8-socket systems, and of course many custom machines running hundreds or thousands of CPUs/GPGPUs.

The original reason for it was simply because CPUs weren't very powerful, often only having 2-8 cores. Having multiple sockets was the only way to increase power. These days we have up to 64-core/128-thread CPUs from AMD (I think Intel tops out at 28-core/56-thread), so single-socket machines are "good enough" for most workstation tasks. For a lot of single-user tasks having more than about 32 cores doesn't improve performance. Lots of cores are great for servers though, where you have thousands of users requiring high concurrency.

Having to communicate between multiple CPUs and share memory space adds cost, complexity and bottlenecks to the system. Techniques like NUMA (Non-Uniform Memory Access - https://en.wikipedia.org/wiki/Non-uniform_memory_access) attempt to has fast memory associated with each CPU that is also shared with other CPUs, but it adds overhead and gets complicated.

One of the advantages of the Apple M1 is that everything is so integrated on the SoC package, and importantly, physically close together. The physical distance that data has to travel becomes a limiting factor in bandwidth due the transmission limits (the speed of light, capacitative effects etc.). If you have to build interconnects between CPUs, you reduce those advantages.

We are much more likely to see higher core counts on Apple Silicon than multi-socket configurations. Bear in mind that there are already 128-core ARM CPUs (e.g. https://www.crn.com/news/components...core-altra-cpu-targets-intel-amd-in-the-cloud) so there is nothing inherently stopping Apple scaling up the Mx SoCs to match Intel/AMD workstation chips.

 

[JohnnyGo]

I have to side with the bigger core count crowd.

There maybe an M2X SOC with 8 performance cores that allows for Apple to sell an M2 with 4-6 performance cores (binned parts or just w/ cores turned off).

I expect a doubling (16/32) or quadrupling of RAM (32/64) in the package.

My guess is that Apple will have computers being sold with M1, M2 and M2X at the same time

 

[Joelist]

I would assume that the next iteration (i.e. M1 to M2) will release in parallel with the next core revision. (i.e. A14 to A15).

 

[Fomalhaut]

I would assume that the next iteration (i.e. M1 to M2) will release in parallel with the next core revision. (i.e. A14 to A15).

That would make sense. I think the next SoC we'll see will be an "M1X" or "M1-Pro" or some such modifier, to have the same underlying technology design but scaled out to have more CPU & GPU cores.