Apple CPUs "just margins off" desktop CPUs - Anandtech

[cytg111]

If people though that Zen's IPC improvement over Dozer was great what about A12 IPC over Skylake? It's plain obvious that next gen cpu arch from Intel can't be x86 compatible anymore.

They tried this with itanium.. It is all about that user base..

 

[CatMerc]

x86 have lasted because racing cpu arch's don't have had any meaningful performance advantage. Now it seems that there will be twofold performance advantage for aarch64 so x86 can't hold it's place anymore and even Intel has to acknowledge that.

Nonsense.

What you see is the result of Intel's 10nm screwing the entire pipeline of new CPU's. There's a lot of 10nm IP waiting to be released that just doesn't find its way anywhere because they can't manufacture it. Skylake isn't the pinnacle of x86, not by a long shot. Even AMD with a fraction of either Intel/Apple's resources is going to beat Skylake (Based on what I've heard, nothing validated). It's 2015 tech, going up against A9X.

IPC means instructions per clock. For that ARMv64 starts to be class above x86, and there's probably nothing that x86 could do to overcome that deficit. It's ridiculous that cpu meant to phone starts to be faster than desktop-class x86. AMD did find that too, K12 was ARMv64 arch but it's side-project Zen prioritizes more because competition from Intel was so low level.

IPC comparison between ISA's is completely pointless, as one can have a lower "IPC" but still perform higher, because the instructions are denser. It's only a good tool when you're talking two identical ISA's (ARM vs ARM, x86 vs x86). You're better off comparing things in units of time.

As for K12, the competition in ARM servers was basically non existent, and K12 was still shelved. If it was better than Zen, you really think that would happen? I don't.

 

[naukkis]

IPC comparison between ISA's is completely pointless, as one can have a lower "IPC" but still perform higher, because the instructions are denser. It's only a good tool when you're talking two identical ISA's (ARM vs ARM, x86 vs x86). You're better off comparing things in units of time.

Even with same ISA different(or even same) cpu's could do same job with different instructions with different execution time and number of instructions. Spec test are for testing different configurations, give a problem and let compiler and cpu solve it as efficient as they could. ARM probably executes much more instructions than x86 for comparable performance but that won't matter, this performance "IPC" is about instruction level parallelism and for that A12 seems to be superior to Skylake. And that is mostly coming from more effective instruction set, A12 only decodes 6 ARM instructions per clock cycle - 192 bit where Skylake has to fetch 256 bits for 4 instructions - x86 has to burn much more power to find more instruction level parallelism than aarch64 with it's compact and simple 32 bit per instruction format.

As for K12, the competition in ARM servers was basically non existent, and K12 was still shelved. If it was better than Zen, you really think that would happen? I don't.

Rumors said that K12 was 10% faster than Zen. For such a low difference it was a no brainer to concentrate on x86 than try to change whole software base to different architechture - but as A12 shows there might be much more to extract from aarch64 what is possible from x86.

 

[Arkaign]

The 20,000lb elephant is OS/software and ability to seamlessly replace existing platforms without forcing the entire stack to be recoded at every level for a new architecture. Obviously an emulation/VM can help with that, but at what penalty?

I think Apple could conceivably replace x86 in Macs, but they'd probably have to ditch their Pro lineup, and non-native apps would be a difficulty to contend with, almost certainly be hit with serious performance penalties. That would be very very difficult, but at least conceivable, as many MacOS users don't use anything except Apple software and maybe 1-2 third party things (Adobe and MS most commonly).

A12X on PC, we're talking about reinventing a completely massive OS, software, and hardware matrix comprising so much more complexity and potential issues that it beggars belief that it could be done. Even if MS did a perfect job (lol), all of the software, especially at the B2B level where a lot of the money is would lag, with various levels of failure/success. An emulation layer necessary to support non-native x86 apps and devices in a preemptive multitasking environment would be a difficult challenge and sap a lot of performance potential and resources to accomplish. All to simply compete in an area where the vast majority of the market is already well served by systems $499 and under, with ~$100 ASP.

Apple currently gets absolutely monstrous margins on their 7nm production. Taking a serious amount aside and dedicating it to compete with McDonald's instead of Lexus doesn't particularly sound wise.

If they had a massive amount of available fab production in need of a use, that would be one thing.

I mean put a pin in this, come back and check fall 2020, I'd put odds at maybe 10% that we see a non-x86 Mac product by then, and maybe 1% or less a PC product featuring an Apple CPU. Microsoft already had a rough go of using ARM with Surface RT. It was a solution in search of a problem to be honest, based on the financial fundamentals at play.

 

[NTMBK]

Apple could reuse the chip for iPad Pro in laptops and iMacs, and an ARM chip for Mac Pro/iMac Pro could be reused for Apple's internal server farms.

 

[CZroe]

Huh? Intel and AMD will continue to make x86 products. Software demands it.

Apple switched from PowerPC to X86 easily enough. The App Store would only make this easier now that it is the standard way to distribute software for OSX. If Apple makes the switch then a lot of the Max ports can port back to Windows with ARM support.

Microsoft tried to do something similar with the Windows Store so that all Windows Store apps would be compatible with Windows RT, XBOX ONE, Windows Phone, and more, which would have totally freed them from X86 if the plan had worked out. If Apple made the switch then it is feasible that Microsoft could succeed if they tried again.

If Apple added a decent X86 core or pulled off code translation better than the Transmeta Crusoe did to ease the transition then I can totally see ARM taking over from X86.

 

[Ichinisan]

The 20,000lb elephant is OS/software and ability to seamlessly replace existing platforms without forcing the entire stack to be recoded at every level for a new architecture. Obviously an emulation/VM can help with that, but at what penalty?

I think Apple could conceivably replace x86 in Macs, but they'd probably have to ditch their Pro lineup, and non-native apps would be a difficulty to contend with, almost certainly be hit with serious performance penalties. That would be very very difficult, but at least conceivable, as many MacOS users don't use anything except Apple software and maybe 1-2 third party things (Adobe and MS most commonly).

A12X on PC, we're talking about reinventing a completely massive OS, software, and hardware matrix comprising so much more complexity and potential issues that it beggars belief that it could be done. Even if MS did a perfect job (lol), all of the software, especially at the B2B level where a lot of the money is would lag, with various levels of failure/success. An emulation layer necessary to support non-native x86 apps and devices in a preemptive multitasking environment would be a difficult challenge and sap a lot of performance potential and resources to accomplish. All to simply compete in an area where the vast majority of the market is already well served by systems $499 and under, with ~$100 ASP.

Apple currently gets absolutely monstrous margins on their 7nm production. Taking a serious amount aside and dedicating it to compete with McDonald's instead of Lexus doesn't particularly sound wise.

If they had a massive amount of available fab production in need of a use, that would be one thing.

I mean put a pin in this, come back and check fall 2020, I'd put odds at maybe 10% that we see a non-x86 Mac product by then, and maybe 1% or less a PC product featuring an Apple CPU. Microsoft already had a rough go of using ARM with Surface RT. It was a solution in search of a problem to be honest, based on the financial fundamentals at play.

Do you think Apple might be able to toss some x86 cores into the same physical package? Maybe that could be used just for legacy software and virtualization.

 

[Arkaign]

Do you think Apple might be able to toss some x86 cores into the same physical package? Maybe that could be used just for legacy software and virtualization.

It's absolutely possible if they can work out the licensing aspect. I wouldn't count them out, though it would be more useful if they could make a monstrous VM in terms of efficiency I would think. Otherwise you're talking a solid hit in transistor count.

 

[gdansk]

If it was any other company, this would be great news. But it's Apple. So it's practically useless for the vast majority of us. Even if they do put it in the next Macbooks it won't be backwards compatible for the PC crowd. And it will never be priced reasonably.

 

[mattiasnyc]

Apple switched from PowerPC to X86 easily enough.

Apple had a fraction of the market of MS and I would therefore guess far fewer 3rd party software manufacturers that needed to adapt to that.

I'd say that all we have to do is look at the vast amount of software that runs on some sort of x86 architecture and then ask ourselves what the incentives are for anyone to change that, at least in the short term. I have a hard time seeing it.

 

[beginner99]

People are throwing way too much dust in the air here. The IPC of the A12 is already vastly higher compared to the Skylake family, so much that SC performance of a 3-4W part is equal to 20W Intel parts on half the frequency. They might not hit 5 or even 4GHz on their theoretical ARM desktop Macs, but a 8-24Core CPU range @ just 3 GHz turbo would likely already outperform anything Intel has to offer. I'd expect them to hit at least 4.

A12 IPC seems to be well over 50% better in specint2006 than Skylake. It seems plain obviously that x86 time as performance cpu architecture starts to be over.

Maybe or maybe not. An important part you both are missing as was established on page 2 of this thread an A12 core on 7nm has roughly 3.5 the size of a skylake (coffeelake) core at 14 nm. Good luck making a 24-core monolith out of that....will be bigger than reticle limit even on 7 nm. Die size matters immensely for margins. For apple this isn't a huge problem as the cheapest device this sell in is like $800. Compare that to a skylake-based celeron/pentium that sell for like $40 in the <$400 segment.

Also Intel has multiple dies for different markets / performances target. Apple has 1 die, they can't make up a meaningful mac line with just 1 die, would need at least 3 or ditch some products at the high-end (mac pros) as sales of these are way too small to justify their own die.

What I'm saying is that no, apple will not enter the server market and no Apple will not replace intel without ditching multiple products from the mac line. That is entirely possible, but what are the actual benefits? Look at all the software devs using mac. Do you think they would keep using them if it wasn't x86 anymore? No, because they write software for x86 so best to own such a device.

 

[IntelUser2000]

But how do we know that the problem is x86 and not intel?

When intel hit a wall with the pentium 4 I am sure there was people also saying that x86 was dead back then. And then Intel showed the new core arch with huge gains in every metric.
The problem I see here is not x86. The problem I see now is that Intel doesn't have an Israel side project with a brand new arch like ten years ago they can use to save the situation. AMD with vastly less resources showed us that they can be up there with x86 if you execute well enough and I am sure that if apple poured the money they are pouring in their arm64 to a hypotetical x86 they could have achieved the same levels of performance and efficiency.

This.

I think the problem is we look at benchmarks that puts everything in equal ground when in the real world nothing is in a vacuum and is far away from a benchmark.

Success for x86 could have been simply because they had the largest volume, and most manpower and hours went into developing it, thus reaching the highest performance level. Now, very high volume is also in the ARM, no, more specifically Apple.

The factors that cause a company to be wildly successful(whether its Intel, or Apple) prevents it from being successful outside of its core market. We will unlikely see a desktop/laptop form factor taken over by Apple A-series chips in volume, because significant majority is sold as low-margin, low cost parts.

Agree, ICC provides better score than GCC for Xeon, but A12 still leads. Apple just makes 4-core A12 @3Ghz for MacBook Pro, Intel game over.

Comparing the A12 to a 165W Xeon or a 95W Desktop chip sounds otherworldly. However if you take a look at how the 15W or, even Core M chips do in single thread, it does bring it back down to earth.

Xeon Platinum 8180 @ 3.8GHz vs Core M3-6Y30 "Skylake" @ 2.2GHz max Turbo from Anand's testing:

46.4 vs 24.4 +90%
25 vs 15.3 +63.3%
31 vs 28.2 +10%
40.6 vs 38 +6.8%
35.6 vs 48.1 -26%
30.8 vs 19.3 +59.6%
86.2 vs 409 -78.9%
64.5 vs 37.3 +72.9%
37.9 vs 20.6 +84.0%

That's with the Xeon having 72.7% clock speed advantage. We have 5W Core M products that have the same Turbo clock as the Platinum 8180 now. Saying it performs like a 5W Core M as a headline doesn't look as impressive does it? However, that may be the reality, and yes that is still impressive.

Yea, 2.5GHz A12 chips perform like a 4GHz Skylake. But you can also put a 4GHz Skylake in a form factor not much bigger than the one the A12 is going into.

 

[Etain05]

Maybe or maybe not. An important part you both are missing as was established on page 2 of this thread an A12 core on 7nm has roughly 3.5 the size of a skylake (coffeelake) core at 14 nm. Good luck making a 24-core monolith out of that....will be bigger than reticle limit even on 7 nm. Die size matters immensely for margins. For apple this isn't a huge problem as the cheapest device this sell in is like $800. Compare that to a skylake-based celeron/pentium that sell for like $40 in the <$400 segment.

Also Intel has multiple dies for different markets / performances target. Apple has 1 die, they can't make up a meaningful mac line with just 1 die, would need at least 3 or ditch some products at the high-end (mac pros) as sales of these are way too small to justify their own die.

What I'm saying is that no, apple will not enter the server market and no Apple will not replace intel without ditching multiple products from the mac line. That is entirely possible, but what are the actual benefits? Look at all the software devs using mac. Do you think they would keep using them if it wasn't x86 anymore? No, because they write software for x86 so best to own such a device.

I don't think that's correct. I'm gonna use the image made by CatMerc some time ago:

For the A12 look at Andrei's article, the big core should be 2,07mm², compared to 8,73mm² for the Skylake core. So actually the situation would really be the opposite of what you describe. It's the Skylake core that is 4 times as big as the Vortex core.

Also I think Apple already has 2 dies, the one for their smartphones and the one for their iPads (the A12X that is coming). I don't think that it would be so impossible for Apple to have 2 more dies for their Mac line.

 

[IntelUser2000]

IPC is about same for every cpu with same arch. But spec-results from Intel compiler should be excluded, Intel compiler is hand optimized so well to spec-workload that it's pure cheating.

I really dislike and try to avoid using the term "IPC" as it can mean different things to different people. I just like to call it "performance per clock" because its more accurate and prevents people from arguing what is basically nothing more than semantics.

 

[Nothingness]

Maybe or maybe not. An important part you both are missing as was established on page 2 of this thread an A12 core on 7nm has roughly 3.5 the size of a skylake (coffeelake) core at 14 nm.

Nothing was established. I just saw 2 people picking numbers, forgetting about GPU and other IP area on a SoC and then making some deduction about CPU core area. Funny.

 

[IntelUser2000]

Etain05: beginner99 is talking about die size. Well, he should have clarified that but his comparisons are at the die level.

beginner99: That assume TSMC's 7nm is 4x the density of Intel's 14nm, which may not be true. The consensus seems to be Intel's 10nm is TSMC's 7nm. I also think that's too simplified.

The node where Intel became density-crazy was starting at the 14nm process. Broadwell's core is at 48.9% of Haswell's core, so looking at that only it was slightly greater than 2x in density compared to 22nm.

However, their Atom cores are different. With 14nm it has a 64% reduction, or nearly 2.8x increase in density. If you look at their small core then, they weren't misleading about the density increase with 14nm.

I'll add Atom core comparisons:
https://forums.anandtech.com/threads/the-intel-atom-thread.2341969/page-36#post-35517066
https://www.hardware.fr/news/14169/airmont-64-plus-petit-que-silvermont.html

I'm not content on adding the Bus Interface Unit on the Clover Trail core from Chip-architect's comparisons. Without the BIU and core part only, Clover Trail core is at ~5.1mm2.

Silvermont core size: 2.66mm2
Airmont core size: 0.84mm2
Goldmont core size: Roughly 1.2mm2

 

[Nothingness]

Yea, 2.5GHz A12 chips perform like a 4GHz Skylake. But you can also put a 4GHz Skylake in a form factor not much bigger than the one the A12 is going into.

Can't wait to have that Skylake in my pocket. Let's see how often it throttles to below 1 GHz.

 

[IntelUser2000]

Can't wait to have that Skylake in my pocket. Let's see how often it throttles to below 1 GHz.

It doesn't throttle running single thread, which is the focus of the Spec comparisons. The Skylake Core M3 with the 2.2GHz Turbo clock is fairly comparable to the 2.1GHz Xeon E7 4809 v4, which supports my argument.

Also, I did say Apple is quite ahead, just not otherworldly. Being able to fit that in a phone is an amazing feat, no doubt enabled by far better execution and planning.

 

[Nothingness]

It doesn't throttle running single thread, which is the focus of the Spec comparisons. The Skylake Core M3 with the 2.2GHz Turbo clock is fairly comparable to the 2.1GHz Xeon E7 4809 v4, which supports my argument.

Do you mean a phone with a Skylake CPU is available and has been shown not to throttle?

 

[Etain05]

Etain05: beginner99 is talking about die size. Well, he should have clarified that but his comparisons are at the die level.

I don't think that's true. He specifically said:

An important part you both are missing as was established on page 2 of this thread an A12 core on 7nm has roughly 3.5 the size of a skylake (coffeelake) core at 14 nm.

Plus it wouldn't make any sense to compare full dies. And he was mentioning this in regard to a supposed 28-core server part from Apple.

 

[IntelUser2000]

Do you mean a phone with a Skylake CPU is available and has been shown not to throttle?

I've updated my post. Look at that too.

I don't think that's true. He specifically said:

Go back further, and read post #44. It's about die size.

 

[Etain05]

Go back further, and read post #44. It's about die size.

Not really. Post #44 says:

If you data is correct an A12 core is roughly 3.5 times bigger compared to Intel Core while being on a lower / denser node. One would expect all that area to have some effect on performance. And who knows how apple cheats in benchmarking with power limits off or other stuff...

Again talking about cores. USER8000 was talking about dies, but he also forgot that the A12 actually also has 4 smaller cores, a NPU, and many other IP blocks which are not guaranteed to be present on the Intel die.

 

[Nothingness]

I've updated my post. Look at that too.

Sorry but you're still claiming that a Skylake could be put in a form factor similar to A12. And I still think this would throttle.

Go back further, and read post #44. It's about die size.

Yeah and then the swtich is made to "core" in post #111 to say it can't scale to 24-core. Silly.

 

[IntelUser2000]

Sorry but you're still claiming that a Skylake could be put in a form factor similar to A12. And I still think this would throttle.

Perhaps it did sound like that. But I never said that explicitely. What I meant was the differences are not great as comparing against a 165W Server chip that has 28 cores with all the I/O, memory controller and AVX units powered up fully, or even a 65/95W desktop part. That's why I continue to emphasize its not an otherworldly difference, but still an amazing one.

TDP isn't the biggest issue for Intel's Core platforms. It's that its not integrated enough, and the platform idle power is still too high, even for Tablets. Once you get the two fixed, a 4.5W chip can go into lot more devices.

And you know they have been at a standstill for 3 years now. No progress since Skylake. Arguably the problems go even further back and they may have been crawling for 2 more years prior to that. Literally everyone executes better than them nowadays. If we are talking about "could" and "if" then having a Tigerlake chip by now would have changed the competitive landscape enormously. Their constant fumbling clouds proper analysis of what's theoretically capable.

 

[CatMerc]

Food for thought: The fabric and uncore alone on Skylake cores eat about as much power as the TDP of an entire A12 SoC. They have different goals in terms of flexibility and scalability in both the architectural sense and user sense.

Combine that with the different nodes, and how Skylake is a 2015 architecture, with its successor being held up by manufacturing issues, making this an ARM Vs x86 comparison is silly.

Here's a funny thing. From what I've heard, Tiger Lake is supposed to bring idle power down to 6mW. That made Apple executives take a step back and re-evaluate their plans. Now consider that Tiger Lake was supposed to ALREADY be out, and it's quite clear to me that it's a problem of execution, not of architecture.