Why Intel and AMD don't make chips like the M2 Max and M2 Ultra

[sack_peak]

Why Intel and AMD don't make chips like the M2 Max and M2 Ultra
By Matthew Connatser
Published 5 days ago

Intel, AMD, and Apple all have powerful CPU and GPU architectures, yet Apple's M2 stands alone in PCs. Here's why.

https://www.xda-developers.com/why-intel-and-amd-dont-make-chips-like-the-m2/

 

[leman]

It's not very clear what the question actually means. It could be one of these:

- Why don't Intel and AMD make huge SoCs (answer: cost and limited market interest; but: consoles, and Intel is starting to)
- Why don't other vendor make energy-efficient GPUs (answer: they do, but it's cheaper to run GPUs at higher frequency so outside of expensive specialised professional products Nvidia doesn't bother)
- Why don't Intel or AMD make fast low-power CPUs like Apple (answer: because they don't know how, at least not yet)

But yes, the article seems to cover the most important points.

 

[thebart]

Isn't the answer they have to keep compatibility with a legacy architecture?

 

[sack_peak]

Isn't the answer they have to keep compatibility with a legacy architecture?

Intel actually wants to free up silicon space by dropping some of that.

This will not adversely impact users of legacy architecture as Intel is likely to still offer the same old chips up for sale.

 

[deconstruct60]

Why Intel and AMD don't make chips like the M2 Max and M2 Ultra
By Matthew Connatser
Published 5 days ago

Intel, AMD, and Apple all have powerful CPU and GPU architectures, yet Apple's M2 stands alone in PCs. Here's why.

https://www.xda-developers.com/why-intel-and-amd-dont-make-chips-like-the-m2/

This part isn't quite correct.

" ... All that area reserved for the memory buses has a domino effect. It's less room for more CPU and GPU cores, ..."

The primary domino effect is on I/O ; not cores. The typical AMD/Intel SoC will have x16 (or more ) PCI-e lanes. The M-series typically has x4. Apple completely walks away from any kind of dGPU option. AMD/Intel do not. Apple uses (trades off) much of the die edge space that 'better' , more general I/O bandwidth for more memory specific I/O bandwidth.

" ... These are much, much smaller CPUs thanks to the fact that they aren't hobbled by a massive memory system required by a large integrated GPU. ..."

Apple really is not that far behind on CPU core count. ( if compare unified , monolithic dies ) . And Apple 'uncore' is bigger 'trade-off' issue ( does Intel/AMD have a SSD controller built into the main die? Nope. ). There is a matter of just how much diverse function Apple integrates onto the main die ( eschewing discrete options to implement functionality. Drives , security , etc. )

The overall tension this take on what Apple is doing is missing a whole lot of the holistic ( effectively anti-discrete function) viewpoint Apple takes. The memory subsystem is largely being driven by the GPU. GPU needs wide ... the whole chip gets wide. But to a large extent that is just fall out from the constraint that the GPU shall not be discrete.

Dropping dGPUs means dropped Nvidia. And most of the PC market system vendors don't want that. ( e.g., The Qualcomm Oryon (Nuvia core ) solution supposedly has substantive dGPU support built into the package. IMHO, a goofy trade-off that will assure that they lag behind Apple's leading edge, but that is what the buyers/system vendors asked for ( trying to make every system possible for everybody) ) . AMD also wants to sell AMD dGPUs. Intel wants to sell everything to everybody.

Apple has six classic PC form factor products. The M-series is in a couple of iPads too. Their SoC doesn't have to do everything for everybody.


The other issue is that Apple is doing it lower power consuming ( max Pref/Watt). Intel is going tiles/chiplets but that costs in Perf/Watt. AMD single dies don't have core counts or die sizes way higher than Apple's. Intel ... not really either ( or are seriously lagging on iGPU. )


The other problem with the article is that it ignores that Intel/AMD CPU designs have largely bubbled down from the server/desktop space into the mainstream products. So their cores have ended up on the 'large' side. That is changing. Intel is using 'E cores' (smaller than the super sized ones) and AMD has 'Cloud' cores ( again where they put a design constraint on core area footprint. )


Intel and AMD are drifting toward some of the same design trade-offs. Intel hinting....

https://cdn.mos.cms.futurecdn.net/CvcqX85m7TkTkbAbZGrsth-1200-80.png

If the SoC tile/chiplet there as wide as an Apple M Pro/Max? No. They are trading off faster data rates on fewer lanes ( LPDDR5X which Apple hasn't done yet. But when they do AMD/Intel will be left behind. )

 

[deconstruct60]

Isn't the answer they have to keep compatibility with a legacy architecture?

Architecture of the whole PC. ( compatibility with the "box with slots " notions ) . It doesn't really have much substantively to do with the CPU micro-architectural differences. It is more that some folks want a CPU (only) or minimally a CPU focused package. Apple is far more focused on building a more completely a system on a chip ( SoC). Apple is not after maximum permutations and combinations of discrete chips.

 

[bobcomer]

Intel actually wants to free up silicon space by dropping some of that.

Their market doesn't though...

 

[MacDaddyPanda]

Because Apple poached the engineers from Intel. Intel wasn't paying those people enough. Nor heading in a direction that would lead down this path.

 

[leman]

Because Apple poached the engineers from Intel. Intel wasn't paying those people enough. Nor heading in a direction that would lead down this path.

Please. Intel was never interested in that direction.

 

[MacDaddyPanda]

Please. Intel was never interested in that direction.

Yeah that's kinda what I said in my last sentence.

 

[Pressure]

I find AMDs approach fascinating with the Zen4 and Zen4C cores going in the Phoenix 2.

AMD don't believe in the big and little core approach but instead offer two cores sharing the same register-transfer level (RTL) template but optimise one of the cores for density (Zen4C). The result is both cores are exactly the same but Zen4C has a higher cell density (which makes it 35% smaller) and sacrifices clock speed potential compared to the less dense Zen4 core.

The Phoenix 2 will have four Zen4C cores and two Zen4 cores that are identical except for max clock speed. This makes a lot of sense in devices that are thermal limited like laptops, gamepads etc. In these cases having Zen4 cores would just be a waste of silicon area as they would be limited in speed due to the thermal design power.

 

[deconstruct60]

Their market doesn't though...

A narrow subset doesn't. Most of the market will take a trade off if that means better mainstream performance on future cores. (which it probably will.) . Windows has dumped 16-bit stuff over time. It is a far slower deprecation -> retirement process timeline than Apple would use but it has been done. 16-bit 'real mode' isn't what vast majority of folks are running at all.

 

[deconstruct60]

Please. Intel was never interested in that direction.

Intel was interested. They were not particularly interest in doing it 'right'. Intel pitched using x86 for the iPad. They wanted the business... they just didn't have anything that would be competitive.

Intel didn't 'have to' take StrongArm. Similar to how Apple is hammering Mac SoCs to look mainly look like A-series SoC ... similar problem with Intel trying to diversify away from just the x86 'hammer' ( when have a hammer everything looks like a nail. GPU ... do that with x86 cores. mobile SoC ... do that with x86 cores. etc. )

 

[leman]

I find AMDs approach fascinating with the Zen4 and Zen4C cores going in the Phoenix 2.

AMD don't believe in the big and little core approach but instead offer two cores sharing the same register-transfer level (RTL) template but optimise one of the cores for density (Zen4C). The result is both cores are exactly the same but Zen4C has a higher cell density (which makes it 35% smaller) and sacrifices clock speed potential compared to the less dense Zen4 core.

The Phoenix 2 will have four Zen4C cores and two Zen4 cores that are identical except for max clock speed. This makes a lot of sense in devices that are thermal limited like laptops, gamepads etc. In these cases having Zen4 cores would just be a waste of silicon area as they would be limited in speed due to the thermal design power.

Yes, that's a great strategy for AMD and it's working out well. I think it's particularly interesting since it betrays a bit about how these CPUs are designed. Them being able to make the core more compact and more energy efficient by constraining the operating frequency range tells us how costly it must be to pursue those high Ghz figures.

At any rate, this strategy is good for AMD since their IPC is low (comparative speaking at least — Zen4c at 3.1 Ghz is comparable to A13 at 2.65Ghz and probably uses more power), so they can tweak the floor plan to achieve different frequency/power targets. I wonder however if the same strategy will also work for a much wider CPU. And of course, it won't give you as drastic efficiency improvements as what we see with ARM E-cores. E.g. Apple's A17 E-cores consume 0.25-0.3 watts at 2.1Ghz while delivering IPC comparable to that of Zen4. Of course, that's a very different design goals. AMD's primary consideration is server market and their primary design goal is die area reduction. ARM's E-cores are there for background operation (although Apple seems to be increasingly going after throughout boost as well).

 

[bobcomer]

A narrow subset doesn't.

A large segment doesn't, business.

16-bit 'real mode' isn't what vast majority of folks are running at all.

Of course not, but there are still machines using it, we have several in the mill using it. (specialized testing machines, not user machines)

I disagree about it helping future core performance as well. I suspect it wouldn't make any difference at all to the end user. Maybe when we hit the wall on package size it will, but we aren't there yet. If you could tell me a real advantage, I might change my mind, but there isn't one yet. I already only use the used market for replacement parts, and I haven't personally used real mode in a very long time, but I like versatility.

 

[deconstruct60]

A large segment doesn't, business.

All of business is NOT in that subset.

Of course not, but there are still machines using it, we have several in the mill using it. (specialized testing machines, not user machines)

Right.... User machines do NOT need it. The vast majority of companies have more users than testing machines.

In the scope of systems that are connected to the Internet, 16-bit OS and apps is a thing of the past. Intel/AMD can still churn out 'boat anchored in the 90's " CPU packages for those narrow set of embedded controllers that still need them where the software is stuck in the 90's.

Even in the Arm embedded market 16-bit isn't really a 'thing' anymore.

Making the systems that are going to run 64-bit OS Windows/Linux for there ENTIRE service lifecycle have kludges for 1990's ( or 1980's ) vintage software is nuts . Just plain nuts. That software is never going to run on several 100's of millions of system going forward. Over a decade that will be into the billions. So that is pointless waste of a lot of silicon.

I disagree about it helping future core performance as well. I suspect it wouldn't make any difference at all to the end user. Maybe when we hit the wall on package size it will, but we aren't there yet.

Probably wishful thinking. removing stuff that relatively almost no one uses and applying those design and die area resources to stuff that does most likely will help. Decent chance contributes to less security defects also ( gratuitously complicated subsystems are substantive contributors to bugs. )


Even if just got one data register back from dropping this stuff on a 100-200 core system that would mean 100-200 more data registers. That is just that many more less trips to cache memory that can be skipped. Yes, that will have an impact.

If you could tell me a real advantage, I might change my mind, but there isn't one yet.

Isn't one that you haven't hand waved away. This is largely just misdirection. The core issue is that folks with software stuck in the far distant past don't want to be decoupled from the mainstream because they really don't have much overall market relevance. They increasing needs the overall masses to pay for the legacy costs of dragging around stuff that practically nobody is using.

I already only use the used market for replacement parts, and I haven't personally used real mode in a very long time, but I like versatility.

 

[bobcomer]

All of business is NOT in that subset.

No, it's not, but it's huge in manufacturing.

Anyway, I'll take more useful or less useful every single time.

Probably wishful thinking. removing stuff that relatively almost no one uses and applying those design and die area resources to stuff that does most likely will help. Decent chance contributes to less security defects also ( gratuitously complicated subsystems are substantive contributors to bugs. )

Got any proof it will help PC level hardware? I didn't think so.

Even if just got one data register back from dropping this stuff on a 100-200 core system that would mean 100-200 more data registers. That is just that many more less trips to cache memory that can be skipped. Yes, that will have an impact.

100 or 200 core systems?????????? That's a whole different thing than I'm talking about, that's server class hardware. Given that kind of power, one can emulate anything you need in software. Our midrange machine still runs software from before the 1990's, even though the architecture is quite different. (all without recompiling no less!)

I'm talking about PC's, not servers. But do notice that the makers of the server class hardware understand backwards compatibility too.

Isn't one that you haven't hand waved away.

Good. I hope they never do it then. No reason to.

 

[falainber]

It's not very clear what the question actually means. It could be one of these:

- Why don't Intel and AMD make huge SoCs (answer: cost and limited market interest; but: consoles, and Intel is starting to)
- Why don't other vendor make energy-efficient GPUs (answer: they do, but it's cheaper to run GPUs at higher frequency so outside of expensive specialised professional products Nvidia doesn't bother)
- Why don't Intel or AMD make fast low-power CPUs like Apple (answer: because they don't know how, at least not yet)

But yes, the article seems to cover the most important points.

The answer to the third question is an oversimplication. One should be aware that all these companies use the same architecture for different use cases: phones to tablets to laptops to desktops to workstations ro servers. When choosing architecture there are tradeoffs. AMD and Intel make most of their money in servers. Their architecture is optimized for this use case. Apple makes money on phones. That's why they use the architecture optimized for low-power. The tradeoff is they can't scale it to workstation/server type use cases. They are limited even in desktops (in memory size limits and GPU performance)

 

[leman]

The answer to the third question is an oversimplication. One should be aware that all these companies use the same architecture for different use cases: phones to tablets to laptops to desktops to workstations ro servers. When choosing architecture there are tradeoffs. AMD and Intel make most of their money in servers. Their architecture is optimized for this use case. Apple makes money on phones. That's why they use the architecture optimized for low-power. The tradeoff is they can't scale it to workstation/server type use cases. They are limited even in desktops (in memory size limits and GPU performance)

I fully agree that what I wrote is an oversimplification. I can’t really agree with the rest.

Apples P-cores have considerably higher throughput per watt and per mm2 compared to Intel server cores. Apple has advanced matrix coprocessors for matrix operations. Of Apple wanted, they could build a large server chip that would outperform anything else by a wide margin. Apple doesn’t make these chips because they are not interested in the server business. Workstation is a different matter, Apple is fundamentally limited by their low power consumption limits here.

Now, I would fully agree with you if you said that Intel optimizes their CPUs for high SIMD throughput (a decision they are slowly reversing as it proved being too costly), but that’s a different topic. Again, Apples solution to this is a specialized coprocessor.

 

[theorist9]

It's an odd article. It's supposed to be all about the tradeoffs between AS vs Intel/AMD/NVIDIA, yet there's not a single mention of efficiency. That's particularly surprising, given that he claims the key reason AS has lower perf/area is because "they aren't hobbled by a massive memory system required by a large integrated GPU."

This seems to entirely miss a key difference between AS chips and the others that contributes significantly to a lower perf/area: To get high efficiency, they are clocked much lower.

That cuts your perf/area in direct proportion to the reduction in clock speed. Granted, it might be the case that, to get higher clock speeds, you need to slightly reduce density, but that's a secondary effect.

 

[Joe Dohn]

- Why don't Intel or AMD make fast low-power CPUs like Apple (answer: because they don't know how, at least not yet)

They do. AMD's Ryzen Z1 Extreme (custom portable chip) can consume only 10 TPW while running Windows and running full-blown triple A games. The battery is still not good as Apple's but there are workarounds to that.

 

[floral]

Why Intel and AMD don't make chips like the M2 Max and M2 Ultra

Because they aren't Apple? :P

 

[sack_peak]

not good as Apple's but there are workarounds to that.

That's why it costs more.

No need to work around.

 

[leman]

They do. AMD's Ryzen Z1 Extreme (custom portable chip) can consume only 10 TPW while running Windows and running full-blown triple A games. The battery is still not good as Apple's but there are workarounds to that.

Oh please. Z1 single core is roughly comparable with Apple A15 and I can ensure you that the former is using substantially more power to get there.



For multi core, sure, AMD has more cores and they can clock them super low and still get decent combined throughput. It’s a very viable strategy, that for sure, and AMD had great success with it, but they need this strategy exactly because their single core starts using much more power at higher performance levels. That’s why we see a four-core Apple CPU competing in performance with an eight-core AMD CPU at a similar power consumption level.

 

[MRMSFC]

They do. AMD's Ryzen Z1 Extreme (custom portable chip) can consume only 10 TPW while running Windows and running full-blown triple A games. The battery is still not good as Apple's but there are workarounds to that.

Stuff I’m seeing is all at 1080p low settings, using dx12 (which needless to say is far more optimized for in games than Metal, or god forbid, opengl)

So sure, using far more optimized software on a lower resolution screen, and with much, much worse battery life.