Apple Introduces M1 Pro and M1 Max: the Most Powerful Chips Apple Has Ever Built

[R0H1T]

Yes, but that's precisely why pointing out that the M1P/M are monolithic allows for huge power savings as they don't need off-die interfaces for most of this. Keeping data on silicon is a massive power savings. Of course they're working with 10 (8+2) CPU cores and an 8-"core" GPU, not a 32-64-core CPU, so the interfaces can also be much, much simpler.

My point is/was that Apple has at least 3-4 things that make the Mxx chips so much more efficient, I won't count ARM in there so ~

• 5nm node
• LPDDR5
• monolithic die
• UMA for much more efficient use of memory

AMD or Intel need at least 3 of them to come close to some of the claims Apple are making today, IIRC AMD is planning at least 2 if not 3 of these for their zen4/5 APUs so I'm honestly expecting much higher efficiency numbers especially if they can bring true unified memory access quickly.

 

[apoklyps3]

oh, good old benchmarks, especially on a closed system.
everybody powers up their system of preference just to run benchmarks all day long
laughable, at best. the day we compare pears and knives is not far away.

 

[Valantar]

oh, good old benchmarks, especially on a closed system.
everybody powers up their system of preference just to run benchmarks all day long
laughable, at best. the day we compare pears and knives is not far away.

... so ... you're just demonstrating that you haven't really understood what you're bringing up. "Benchmarks" are never a problem. A benchmark is just a comparable way of measuring something. Synthetic benchmarks can be a problem, as they often purport to represent something which they can't really claim to represent - such as 3DMark being a gaming benchmark. A benchmark being synthetic means it uses custom-made tests rather than replicating actual workloads from actual real-world use cases. A good example would be CrystalDiskMark vs. AnandTech's Storage Bench SSD benchmark - CDM runs a fully synthetic test that is explicitly designed for peak performance ("advertising numbers" and is in no way representative of any real-world performance), while ATSB uses captured application traces (i.e. logs of how real-world software reads from and writes to storage while in use) and is thus both reliable and representative (as long as the applications used are the same, obviously). Single workload benchmarks used to represent the full range of performance of a system - such as how Cinebench is often used - is also a problem, but that's not down to it being "a benchmark". And single workload benchmarks are both valid and useful as long as they aren't used as a singular indicator of overall performance, but are viewed as what they are: a test of a specific use case.

SPEC? SPEC is not synthetic, nor is it a single workload. SPEC is a very wide-ranging benchmark suite, put together by an industry-spanning consortium, and uses sections of real-world software for its performance measurements. You can read about what each benchmark does and what software it is based on on the SPEC website. The tests range from video compression to code compilation to weather simulations and 3D modelling . Is this a perfectly representative set of tests? Of course not. That is literally impossible. But it's the best and most widely accepted set of performance tests in the computing industry. That does mean it has a bias: it's very much focused towards the types of workloads where heavy computation are useful for professional use. That means no gaming tests, no regular office workloads, etc. Still, it is a good representation of actual CPU performance across a wide range of heavy CPU-focused tasks.

Oh, and if you're still not happy with that, did you even bother to look at the other pages where they test other software? Sure, there isn't a lot of it, but that isn't all that strange given that this was essentially a day 1 test on a brand new hardware+software platform. When AT looks at the new chips, there will absolutely be other software tested.



Edit: for those of you speculating in GPU performance (@r9 @billEST @Ravenas): Apparently Apple's comparison GPU in their M1 Pro charts was a 3050 Ti in a Lenovo Legion 5 (82JW0012US). No word on the benchmarks used though. For the M1 Max they were comparing against a mobile 3080 in both a Razer Blade 15 (their "slim laptop" example, 100W GPU TDP) and an MSI GE76 Raider (160W GPU TDP). They're not claiming quite 3080-level performance looking at those (very vague) graphs, but it's pretty close. Of course real-world tests will be extremely dependent on software and driver support, as we can see in AT's M1 testing (where they use both synthetic benchmarks as well as RotTR, and the M1 comes close to a mobile 1650 in synthetics but falls well behind in actual gaming). Apple is undoubtedly most focused on GPGPU tasks like rendering, but this still looks mighty interesting overall. Might we see game developers start to target Apple devices now that some of them can deliver actually good GPU performance? Or will drivers and platform issues hold them back

My 500mm² guess was also a bit off, but even 432mm² is downright insanity for a monolithic laptop SoC. Especially on 5nm. Renoir is ~150mm², for reference. Plugging that into a wafer calculator gives us a total of 124 dice per 300mm wafer, so assuming a ~$10 000 wafer cost (which might be higher with 5nm, especially these days), that's $80 just in pure die costs for these chips assuming all dice can be used (which won't be the case), before packaging and any accounting for engineering/R&D costs. That is a very, very, very, very expensive chip. Though I guess that goes to show the benefits of vertical integration, as that R&D cost integrates into everything else the company is working on, rather than forming the basis for profit margins on something that's sold on to a second company.

 

[apoklyps3]

yeah, yeah, sure...apple is magic, the best in the world and the leader of inovation.

 

[Vya Domus]

And how is "trounce" extreme when we're talking about a >50% IPC advantage?

How did you come to that figure ?

 

[Deleted member 24505]

yeah, yeah, sure...apple is magic, the best in the world and the leader of inovation.

No, but in the mobile realm, pretty up there. Get back to your potatoes /s

 

[Valantar]

How did you come to that figure ?

By them matching or beating the ST performance of Intel and AMD desktop CPU performance (10900K and 5950X, 5.3GHz and 4.9GHz max boost) with a 3.2GHz SoC? Though you're right, 40% is more accurate. No less insane, but more accurate (I didn't actually look up the exact clock speeds or actually calculate anything before now - I obviously should have). The M1 falls slightly behind the 5950X in integer workloads but wins by a clear margin in FP workloads, making it overall faster despite only having 65% the clock speed (or the 5950X being clocked 53% higher, depending on which way you calculate). Saying they have a 40% IPC advantage is definitely no stretch, though.

The exact numbers (assuming both chips maintain peak boost during the entire workload, which they likely do for an ST test):
5950X:
SPECINT2017: 7.29, or 1.4878/GHz
SPECFP2017: 9.79, or 1,9980/GHz

M1:
SPECINT2017: 6.66, or 2,0813/GHz (39.9% higher IPC, or the 5950X has 28.5% lower IPC, depending which way you calculate)
SPECFP2017: 10.37, or 3,2406/GHz (62,2% higher IPC, or the 5950X has 38.3% lower IPC, depending which way you calculate)

As we're talking relative numbers here any percentage will always be somewhat misleading, but the advantage is undeniable and massive.

yeah, yeah, sure...apple is magic, the best in the world and the leader of inovation.

Magic? Not at all. Just an insanely rich company with near unlimited R&D budgets and the funds to hire whatever engineers they want, licence whatever patents they need, and buy oversized, ludicrously expensive silicon on cutting-edge nodes in a way nobody else can. Could Intel or AMD match them if they had the same resources and could sell these chips? Sure. But they don't, and they can't. The M1 Max is the size of Intel's largest monolithic server CPUs, but with a much lower core count, integrated GPU, NPU and a bunch of other stuff. Intel and AMD's laptop chips are mostly in the 150-200mm² range. Nobody has ever made a laptop SoC even remotely like this - in part because nobody has a guaranteed market for $4000+ laptops like Apple does.

 

[usiname]

Magic? Not at all. Just an insanely rich company with near unlimited R&D budgets and the funds to hire whatever engineers they want, licence whatever patents they need, and buy oversized, ludicrously expensive silicon on cutting-edge nodes in a way nobody else can. Could Intel or AMD match them if they had the same resources and could sell these chips? Sure. But they don't, and they can't. The M1 Max is the size of Intel's largest monolithic server CPUs, but with a much lower core count, integrated GPU, NPU and a bunch of other stuff. Intel and AMD's laptop chips are mostly in the 150-200mm² range. Nobody has ever made a laptop SoC even remotely like this - in part because nobody has a guaranteed market for $4000+ laptops like Apple does.

Also this is 430mm^2 on 5nm, almost 60 billions transisotrs. For comparison 5950x + rtx 3090 have 47 billions transisotrs in total

 

[Caring1]

This is why I love TPU.
So much intelligent discussion to learn from, and of course the odd numbnut thrown in for a laugh.

 

[apoklyps3]

Keep writing nonsense, makes my day.
while you're at it don't forget tu buy apple's 25$ cleaning cloth. there's some "magic story" behind it too

 

[ppn]

They really maxed out the 5nm EUV with the MAX die with 420mm2, as wikichip said once the photo imprint the EUV machine makes or whatever can't be bigger than that. SO this is like big Ampere A100 as transistor count. but half size smaller.

 

[Valantar]

Keep writing nonsense, makes my day.
while you're at it don't forget tu buy apple's 25$ cleaning cloth. there's some "magic story" behind it too

Lol, I have literally never owned a single Apple product in my life, so... meh. Their current engineering prowess is undeniable though. But if you have anything to contradict this, I'd love to read. Got any sources?

Also this is 430mm^2 on 5nm, almost 60 billions transisotrs. For comparison 5950x + rtx 3090 have 47 billions transisotrs in total

Just goes to show how a 'wide and slow' design will always be more efficient than a 'narrow and fast' one as long as you can keep it fed. PC components are generally low margin and highly competitive, so they tend to minimize area and maximize clocks, which makes them inefficient. Of course that's a simplification (making a wide cpu core that performs well isn't easy!), but it works as a rule of thumb.

 

[apoklyps3]

engineering prowesss =)))
sir, I know a fan-boy when I see one.
no need to bury us in apple-fed promo material.

 

[Valantar]

engineering prowesss =)))
sir, I know a fan-boy when I see one.
no need to bury us in apple-fed promo material.

Hm, let's see, someone who had never bought a company's products, and generally doesn't advise others to do so (really not a fan of their closed ecosystems and monopolist tendencies) is a fanboy? That's interesting, for sure. Keep going, what else do you know about me that I don't?

 

[Vya Domus]

The exact numbers (assuming both chips maintain peak boost during the entire workload, which they likely do for an ST test):

That is indeed an assumption. Irrespective of that, I still don't think their core architecture is that much better, virtually all of SECS's tests are known to rely heavily on memory ops, favoring either huge caches or fast system memory.

 

[Richards]

By them matching or beating the ST performance of Intel and AMD desktop CPU performance (10900K and 5950X, 5.3GHz and 4.9GHz max boost) with a 3.2GHz SoC? Though you're right, 40% is more accurate. No less insane, but more accurate (I didn't actually look up the exact clock speeds or actually calculate anything before now - I obviously should have). The M1 falls slightly behind the 5950X in integer workloads but wins by a clear margin in FP workloads, making it overall faster despite only having 65% the clock speed (or the 5950X being clocked 53% higher, depending on which way you calculate). Saying they have a 40% IPC advantage is definitely no stretch, though.

The exact numbers (assuming both chips maintain peak boost during the entire workload, which they likely do for an ST test):
5950X:
SPECINT2017: 7.29, or 1.4878/GHz
SPECFP2017: 9.79, or 1,9980/GHz

M1:
SPECINT2017: 6.66, or 2,0813/GHz (39.9% higher IPC, or the 5950X has 28.5% lower IPC, depending which way you calculate)
SPECFP2017: 10.37, or 3,2406/GHz (62,2% higher IPC, or the 5950X has 38.3% lower IPC, depending which way you calculate)

As we're talking relative numbers here any percentage will always be somewhat misleading, but the advantage is undeniable and massive.

Magic? Not at all. Just an insanely rich company with near unlimited R&D budgets and the funds to hire whatever engineers they want, licence whatever patents they need, and buy oversized, ludicrously expensive silicon on cutting-edge nodes in a way nobody else can. Could Intel or AMD match them if they had the same resources and could sell these chips? Sure. But they don't, and they can't. The M1 Max is the size of Intel's largest monolithic server CPUs, but with a much lower core count, integrated GPU, NPU and a bunch of other stuff. Intel and AMD's laptop chips are mostly in the 150-200mm² range. Nobody has ever made a laptop SoC even remotely like this - in part because nobody has a guaranteed market for $4000+ laptops like Apple does.

If the pro max matches the 3080 in workloads it will be impressive... mark gurman said the desktop m1 max is 128 core gpu thats 3090 desktop performance

 

[dyonoctis]

That is indeed an assumption. Irrespective of that, I still don't think their core architecture is that much better, virtually all of SECS's tests are known to rely heavily on memory ops, favoring either huge caches or fast system memory.

But at the end of the day, what matter is the actual performance that people get from those chips on macOS. As time goes by i don't think that trying to see if their cpu core arch is factually better or not will matter that much, since the software will be optimised for their chip characteristics. For exemple, someone buying an AMD laptop who don't have ML hardware, won't get boosted performance when using apps who have ML task (photoshop is going in that direction).

Apple know their market well, most of the 6,8% buying their computers are doing graphic design, motion design, video editing, music editing, or just "office" like workload. And a lot of those tasks are being more and more accelerated by GPUs and other specialised hardware. I firmly believe that Apple isn't banking on having the most "versatile CPU" arch, but a versatile Soc

 

[Vya Domus]

But at the end of the day, what matter is the actual performance that people get from those chips on macOS. As time goes by i don't think that trying to see if their cpu core arch is factually better or not will matter that much, since the software will be optimised for their chip characteristics.

Sure, but if that's the case let's stop thinking that their chips are the greatest thing since sliced bread.

For exemple, someone buying an AMD laptop who don't have ML hardware, won't get boosted performance when using apps who have ML task (photoshop is going in that direction).

ML can run on GPUs just fine, that's actually still the preferred option. In fact Apple's own APIs most of the time wont even use the NPU.

Apple know their market well, most of the 6,8% buying their computers are doing graphic design, motion design, video editing, music editing, or just "office" like workload.

I don't think anyone buys 3000$+ laptops for office work, or if they do they're incredibly unintelligent. What Apple knows is that people want to use some of the professional software available for mac, not necessarily their hardware.

 

[Dredi]

That is indeed an assumption. Irrespective of that, I still don't think their core architecture is that much better, virtually all of SECS's tests are known to rely heavily on memory ops, favoring either huge caches or fast system memory.

I don’t think that valantar was talaking about core architecture, but instead of IPC. We’ll see what stupid amounts of L3 do for zen3 soon anyway, but I doubt that it does much for 1T SPEC.

 

[Vya Domus]

We’ll see what stupid amounts of L3 do for zen3 soon anyway, but I doubt that it does much for 1T SPEC.

L3 caches are sometimes too slow to provide meaningful speed ups, it's usually the L1 caches that get hammered.

I don’t think that valantar was talaking about core architecture, but instead of IPC.

Then it doesn't even make sense talking about IPC in that case, because any CPU will suddenly have higher IPC if it gets faster system memory or larger caches.

 

[Ferd]

Waiting for the funny cooling these new macs will have

 

[dyonoctis]

I don't think anyone buys 3000$+ laptops for office work, or if they do they're incredibly unintelligent. What Apple knows is that people want to use some of the professional software available for mac, not necessarily their hardware.

I did an internship in the communication dept of an university. The computers were either 27" imacs or 15" macbook pro, even for the people who weren't doing graphic work . It's stupid, but it's actually happening

 

[Dredi]

Then it doesn't even make sense talking about IPC in that case, because any CPU will suddenly have higher IPC if it gets faster system memory or larger caches.

IPC is always a system level metric, and always application specific.

why would anyone use faster memory, or add more cache, if it didn’t result in more instructions being executed per clock.

 

[Vya Domus]

IPC is always a system level metric, and always application specific.

IPC has never been a system level metric, it has always been processor specific. You can search for papers regarding measurements of IPC and you'll never come across a system level study of IPC because it doesn't make sense, the CPU is a constant, the system isn't. They always focus on isolating the characteristic of the CPU alone.

Plus to say that it's a system level thing implies that everything should be measured together, what do we do if we want to measure FLOPS throughput ? Do you count the GPU and all the various other accelerators in as well ? After all it's all on the same SoC, same system, right ?

why would anyone use faster memory, or add more cache, if it didn’t result in more instructions being executed per clock.

It results in more instruction been executed per clock some of the time, the upper and lower bounds of IPC and it's behavior remain exactly the same.

 

[billEST]

The M1 Max, at least on paper, makes every other CPU seem like a decade out of date... How can this be?

you put a 30 year prog on pc it works