It'll be interesting to see how they connect the GPU with the CPU while yet have a performant UMA if not in package, as this will increase latency. As I understand it, the low latency high bandwidth memory is contributing quite a bit to the performance of the M1. If the GPU is off package, it'll be interesting to see how they keep the GPU from being starved of memory bandwidth.
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.
Very Interesting M1 Analysis
- Thread starter leons
- Start date
- Sort by reaction score
Quite possibly - this would be closer to the familiar naming of Intel chips, e.g name or generation, plus number of cores and frequency (although I'm not sure whether Apple will advertise CPU frequencies).
What I am wondering is whether we will see "M1" variants with more cores around WWDC time, or "M2" variants in Sep-Nov? Or both?
As I see it, it's quite unlikely to have more than 4 efficiency cores as it looks like 4 is the magic number. Upcoming iterations of the M1 should see a 4+x scaling for the CPU cores, with x being the high performance cores.
Is the limiting factor to adding a powerful GPU on the same die, due to worse fabrication yields with larger dies, or the the thermal constraints of having the combined TDP of CPU and GPU cores concentrated in a smaller physical area?
I'm assuming that having some physical separation (i.e. separate CPU and GPU dies) reduces the temperature overall, or makes it easier for cooling solutions.
Last edited:
Issue is primarily yields. the silicon doesn’t make a very good thermal conductor, so the heat that is generated tends not to spread too much laterally; this is actually why hot spots are such a problem, and at least some intel processors used the technique of rotating through different cores, to give cores a chance to cool down.
Cooling of the package is a function of surface area - with a plain old set of fins and fan you can cool around 10 W/cm^2. With more modern air cooling techniques you can do better. But making the package bigger doesn’t hurt as long as the W/cm^2 being generated doesn’t increased. In other words, more heat is ok, as long as the package gets bigger to go along with it.
if Apple Silicon workstation doesn't perform faster in single core, this would be extremely disappointing, as most software only uses one core per task (1 for sound, 1 for graphics, one for ETC… this way After Effects beyond 4 cores is useless, Animate, Photoshop, Harmony, ETC….)
Some are very developed with multicore in mind, as Final Cut, but this would mind a 699USD Mac mini would be the best Apple device for most of the pro tasks, and this sounds very odd.
For most software the fastest single core, the best.
What would be very useful is (I don't know if it is the case alright) if the OS would be always available no matter how busy would be the "powerful" cores rendering something, thanks to the lower cores, this way you could keep doing simple tasks while rendering some video/3D (this is something sucks right now as all cores are "powerful" and Final Cut hangs systems even with 8 core processors)
Adding so many cores was "a patch" X86 came out in order to keep CPUs more powerful, due to the loss of the Moore law the last decade for single core, but I would prefer a 4 core 200x speed per core than a 16 core 50x speed per core.
In deed, this rule the far it goes, the more sense it makes: 2 core 5.000x Vs 5.000 cores 2x… In X86 that was a problem as the Xeon… but let's see if ARM has enough power/heat room for don't following this rule.
One of the best things Apple Silicon can revert in Macs is this nonsense offer of CPU/GPU where you can't have the best i9 CPU with the best GPU offer, due the MacPro starts with the Xeon, which in single core performance is far beyond the High end iMac, not too much difference BUT
the price is it!!.
This is the worst thing in Macs right now.
Last edited:
Apple products have a ~40% profit margin on average as well. For $699+ you get the whole M1 computer with software and services. To hit these price points Apple Silicone needs to be much cheaper than Threadripper and therefore needs to be on a smaller node. Of course it can be scaled, the M1 is the most efficient desktop chip with the least heat problems. But a faster chip needs much more cores, transistors and area. Especially when you have to build millions of them, every process node shrink is much welcomed at Apple HQ. They’ll wait for 4nm and we’ll have to wait for Pro Macs. Which in turn will be twice as fast as M1, with double the cores and double the RAM, twice as many Thunderbolt ports etc.
Apple's margin is around 40% overall, but I don't think the Mac Pro are at that level, considering that it's not a volume product. Mac Pros could be selling at 100% margin for all we know. The thing with Apple using Apple Silicon is that it doesn't have to incur the profit margin Intel and AMD is earning for the Mac Pro. The SoC is just cost (ignoring TSMC's margin) to Apple, and so provides Apple more flexibility in their cost management and pricing. This is why I think they put the M1 into the new iPad Pro.
IMHO, the SoC for the Mac Pro will still be on the 5nm process, as it has been field tested. Too risky to move the Mac Pros to a newer process and get bad reviews due to potential issues. I think Apple will move the Mac Pro to Apple Silicon by first half 2022 before WWDC, so a 4nm process will be too soon for this time frame. I don't think Apple will hit any issue scaling the M1 to Mac Pro level on the 5nm process.
If the Mac Pro is to transition at the end 2022 (which I think is unlikely), then it'll likely be based on the A15/M2 architecture on the 4nm process, as I'm sure the base Macs will move to the M2 early/mid 2022. Anyway, that's what my crystal ball is telling me
I believe at some Apple will have different number of cores or higher end GPU in the M family of chips as well as higher RAM configs (what most people call an Mx version).
But what if Apple tries to extend the shelf life for each M version ? Like this:
- M1 was released Nov 2020 but will be used until October 2022)
- M2 will be released this October 2021 and will be used through September 2023), in the first year for Pro models (MacBook Pro, iMac Pro), and in the second tear for consumer versions (Mac Mini, MacBook Air, iMac)
- when the M3 is released next year, it will replace the M2 SOC on the PRO version Macs
Likely? Possible? What you guys think?
But what if Apple tries to extend the shelf life for each M version ? Like this:
- M1 was released Nov 2020 but will be used until October 2022)
- M2 will be released this October 2021 and will be used through September 2023), in the first year for Pro models (MacBook Pro, iMac Pro), and in the second tear for consumer versions (Mac Mini, MacBook Air, iMac)
- when the M3 is released next year, it will replace the M2 SOC on the PRO version Macs
Likely? Possible? What you guys think?
I'm actually the opposite, I'd want 16 core x50 over 4 core x200. I'm never in one app for any length of time, I jump around a lot, keeping the other apps in the background, some of them even doing processing. I've noticed the M1 MBA of mine feels sluggish to me using the same usage pattern, yet almost everyone here says it feels faster.
I guess my usage pattern isn't typical, so I guess I'm just saying that usage patterns make all the difference in the world as to which works best.
Your usage pattern does not even classify as heavy multitasking and you don't need multiple cores to run multiple apps simultaneously. Multitasking was a thing long before multi-core CPUs became commonplace and even with modern multicore systems each core will switch the app it is running dozens or hundreds times per second.
More slower cores only really make sense for server or HPC-like processing where you have a large amount (hundreds or thousands) tasks running independently and simultaneously. Regular "desktop" experience has completely different requirements: here you have very intense bursts of computation (drawing, input processing etc.) with extremely long periods of inactivity. Prosumer desktop needs a bit of both as is it often used to perform HPC-like workloads (image processing, rendering etc.), but even then less more powerful cores (if you could get them at the same power consumptions) are almost always a better choice.
You're very wrong about that. I notice the speed difference with multiple core, so by definition, it makes a difference.
Of course. I've been using computers since the early 1970's and have a computer science degree from the 90's, I know how things progressed and how they work.
So you think.
The current FireStorm/Icestorm cores are already incredibly fast and process at least twice the number of instructions at a time of any other core. The reason they would add some more Firestorms and some more GPU (and of course more cache and controllers) to an "M1X" is to lift the ceiling even higher for multithreaded workloads.
I wish I had one available, but i don't. Biggest processor I have is an i9-10900 with 10 cores + hyperthreading. I really do much prefer to use it than my MBA as it's much faster for my usage, even though it's a downgrade single core-wise from my MBA. Even my 3 year old laptop with 4 core +hyperthreading doesn't feel slower than my MBA. (though I never use those side by side like I do with the desktop and my MBA, so it's hard to tell)Well, try running your desktop apps on an Xeon Gold and see how well it goes
That said, it's possible the other attributes of my machine make a difference too. More RAM for one thing...
I'm envisioning a system in a package like this (please forgive my terrible art skills):
Ah, yes, but an i9-10900 is still a consumer-oriented CPU. It still counts as "relatively few rather fast cores". Those enthusiast-level CPUs are designed to be a great fit for prosumer workloads, as they support both high burst speed favored by desktop applications and multi-core sustained thoughtput required for HPC. And I have little doubt that an 10900 will be faster than M1 in almost everything you throw at it (unless its something that does not scale beyond a single core, where M1 will have a very tiny lead).
Anyway, it sounds to me like you simply need more performance than M1 offers. Which is not too surprising, as it's an entry-level device that is not really targeted on prosumers. I also find it insufficient for my daily work, even though it is faster than my i9 in many practical tasks. I am waiting for a 8-core machine which will twice as fast.
Yes, very much so. Just mid level consumer hardware, though maybe more business than just plain consumer.
It still counts as "relatively few rather fast cores".
Compared to server class or clusters, yep. Totally different job.
I'm really just talking about personal usage machines, it's been a very long time since I used anything in the server class of hardware that you're talking about.
Yeah, I will be interested in what comes next! At first I thought the low power cores would make more of a difference than they do, but I suppose those are just for power savings. I'd have wanted it to be 6+2 instead, since I rarely use my laptops on battery.
I mean the system-wide last-level cache (the one all processors share). I think it makes a lot of sense to have it on a separate die in such a system. This die would also contain the memory controllers. It's essentially the same topology as M1, just with functional dies "pulled apart" to improve yields and enable more configurations.
Then I think we agree, sorry, I kind of jumped on your previous post without really considering the context.
Low power cores are interesting as I think they are absolutely relevant to performance — just indirectly. Having low power cores means that low-priority tasks can run using those cores, which means giving more room for the high power cores to do their job. Icestorm consumes so little power that it doesn't even enter the equation. Also, these cores are very small — for of them take just a little bit more space than a single Firestorm core.
No, it is not. Firestorm is clearly the fastest core anywhere near its power envelope, but it is most definitely not the fastest of all cores on the market. If it's too close or application dependent when compared with x86 you just need to take a glance at POWER9 to see that.