Anand: Apple's A7 Cyclone Microarchitecture Detailed (2014-03-31)

[Rakehellion]

Not really. If you've read previous posts, you should've learned that Cyclone isn't anywhere near as efficient, so it consumes a lot more power. What Apple's doing isn't anything new. Intel and AMD have been building fast CPUs for many years now. Apple can go ahead and build their own, but for very TDP constraint devices like phones, efficiency (+ process node) is really all that matters. So the numbers show that Apple didn't succeed beating Intel in mobile processors at all, since it's still about a factor of 2 slower.

Reddit is a real credible source.

 

[jdubs03]

Apple is very aggressively hiring CPU/SoC guys. One of Silvermont's lead architects recently took a job with Apple...and I know for a fact that Apple is constantly trying to recruit Intel-based talent.

Good pickup on that info, thats a solid talent grab for apple, especially as knights landing is supposed to be based off of silvermont, which is strange because in 2015, goldmont is coming out: much, much better core.

 

[Khato]

and I know for a fact that Apple is constantly trying to recruit Intel-based talent.

I know that quite well too (I'm guessing you've noticed the fact that Apple is assembling a GPU design team by now, right?) But such doesn't mean that I know what their plans for world domination are just yet, haha. All I do know is that the kind of design team that they're massing is expensive and doesn't really make sense for simply continuing the status quo. They easily could be intending to make a play for a top to bottom SoC/range of SoCs to service all their products, but I doubt it - it's an awfully large expense/risk to take for not much gain in product margins. No, it really strikes me as a move to be able to design exactly what they need for future product lines fully in-house while letting others supply the chips for their devices in mature markets.

 

[Nothingness]

They easily could be intending to make a play for a top to bottom SoC/range of SoCs to service all their products, but I doubt it - it's an awfully large expense/risk to take for not much gain in product margins.

Yes, and existing Intel Core CPU would be very hard to beat as far as performance goes.

No, it really strikes me as a move to be able to design exactly what they need for future product lines fully in-house while letting others supply the chips for their devices in mature markets.

If by mature markets, you mean Mac, it has started to decline even for Apple (cf. page 27 of their 2013 annual report) so it indeed makes little sense to heavily invest in CPU development, especially given that Intel is doing a very good job here.

I wonder what new products will look like and if some will blur the line between lower end Mac and iPad as some think.

 

[NTMBK]

Good pickup on that info, thats a solid talent grab for apple, especially as knights landing is supposed to be based off of silvermont, which is strange because in 2015, goldmont is coming out: much, much better core.

KL's core will have as much similarity to Silvermont as Goldmont does, I suspect. They need to add 4-way SMT/hyperthreading, and a pair of massive 512-bit vector units, which implies some fairly significant replumbing of the core. Think of Silvermont as a starting point.

 

[NTMBK]

I know that quite well too (I'm guessing you've noticed the fact that Apple is assembling a GPU design team by now, right?) But such doesn't mean that I know what their plans for world domination are just yet, haha. All I do know is that the kind of design team that they're massing is expensive and doesn't really make sense for simply continuing the status quo. They easily could be intending to make a play for a top to bottom SoC/range of SoCs to service all their products, but I doubt it - it's an awfully large expense/risk to take for not much gain in product margins. No, it really strikes me as a move to be able to design exactly what they need for future product lines fully in-house while letting others supply the chips for their devices in mature markets.

Don't just consider the effect that Apple going in-house would have on Apple; consider the effect it would have on Intel:

-Intel is struggling to fill it's fabs as it is, utilization is going to be an even bigger problem
-Intel would lose a big chunk of revenue, which directly leads to reduced R&D, and hence less competitive chips for Apple's rivals
-Intel loses the "halo effect" from having its chips in the world's most prestigious computers; suddenly all those Windows PCs aren't powered by the same silicon as a Macbook Pro, they're powered by the silicon Apple dumped to replace with efficient new ARM cores.

It's not like it's going to kill Intel or anything, but it's certainly bad news.

 

[witeken]

Intel has had 30 years to design a decent mobile processor and they just can't. Even with their own fabs, the biggest engineering budget out there, and all their amazing "expertise" they just can't do it.

30 years ago, smartphones didn't exist. 10 years ago, mobile phones existed but with a CPU ASP of a few bucks, why would they have cared about those CPUs (back then Intel's ASP was even higher than today)? It's only since about 2010 that Intel started realizing that smartphones could give them a lot of money.

They can't? Have you been living under a rock for the past year?

Apple tries it the first time, instantly they have a machine that sells like hotcakes and performs better than any intel, nvidia, or AMD SoCs and arguably better (certainly more advanced) than Qualcomm.

Actually, Apple had already built about 5 SoCs for their previous phones. And if you then throw a lot of money at the problem and hire a lot of smart people, sure you can design something decent.

Again, I don't know why you add sales into the discussion. They could have put Cortex A7 cores into the iPhone 5(s) and it would have also sold millions of times, but that doesn't make Cortex A7 a fast CPU. Also, "better" depends on your definition. Do you mean pure speed? A7 is nothing compared to Core or any desktop/laptop CPU. Do you mean multithreaded performance? A7 loses against most quadcores. Single threaded performance? Cyclone again loses against a multitude of other architectures.

The fact that you don't see those faster designs in phones/desktops has multiple reasons, but the main reason is power consumption. Intel has now made an appropriate architecture for phones and tablets, called Silvermont, which has a much higher performance/watt than any other CPU. So it could potentially be faster than Cyclone, since phones and tablets are TDP constraint products, but instead Intel has chosen to keep power consumption sane with a well-balanced architecture that isn't only suitable for the highest-end devices.

Intel is not the same company it was when it first started inventing stuff, my point about the 8088 is that your "benchmarks" (which are validated by nobody and contradict the evidence available to everyone) are equally ridiculous.

Then give the evidence instead of making ridiculous claims or comparisons. Spoiler alert: you won't find it. Need a source for my claims? Read this.



But I think I might stop this discussion. I tried to give you some information and learn you something, but you choose not to change your opinion based on facts, as far as I can tell. Then there's nothing left to talk about since I've already told you everything to disprove your claims that Intel doesn't build good CPUs.

I'd somewhat expect that he understands and was simply extending the comparison a bit further than is justified... which is at least somewhat understandable considering the post that it's in response to, no?

You understand it... hyperbole.

I'll give credit where its due but I refuse to believe that the A7 dominates so completely until I see OS independent benchmarks.

Great post.

 

[MisterMac]

Apple is very aggressively hiring CPU/SoC guys. One of Silvermont's lead architects recently took a job with Apple...and I know for a fact that Apple is constantly trying to recruit Intel-based talent.

....and how do you wager that's actually true when there's the "techtopus" illegal cartel going on between all the silicon valley giants with lawsuits incoming as we speak?
It seems extremely unlikely intel would let any talent go to Apple of all companies.


People talk about how Intel will suffer without macbook sales.
But imagine Apple?

The ISA optimization on professional programs, that mass need for a HPM process (which everyone but intel seems to abandon trying to design for).


Apple needs Intel more than Intel needs Apple ATM.
Macbooks\MacPro would crumble without intel's performance.

 

[NTMBK]

30 years ago, smartphones didn't exist. 10 years ago, mobile phones existed but with a CPU ASP of a few bucks, why would they have cared about those CPUs (back then Intel's ASP was even higher than today)? It's only since about 2010 that Intel started realizing that smartphones could give them a lot of money.

They designed the original Atom for "UPMCs", a concept not that far from a modern phone. But when people started replacing "big core" laptops with "small core" Atoms their margins tanked, and they panicked, and stalled Atom development.

They can't? Have you been living under a rock for the past year?

Again, how many smartphones have Atom in them? How many non-subsidized tablets?

Actually, Apple had already built about 5 SoCs for their previous phones. And if you then throw a lot of money at the problem and hire a lot of smart people, sure you can design something decent.

They had built the uncore, but the A6 was their first CPU design.

Again, I don't know why you add sales into the discussion. They could have put Cortex A7 cores into the iPhone 5(s) and it would have also sold millions of times, but that doesn't make Cortex A7 a fast CPU. Also, "better" depends on your definition. Do you mean pure speed? A7 is nothing compared to Core or any desktop/laptop CPU. Do you mean multithreaded performance? A7 loses against most quadcores. Single threaded performance? Cyclone again loses against a multitude of other architectures.

Seriously? The iPhones 5S comes top of almost every single CPU performance bench that Anand throws at it: http://www.anandtech.com/show/7335/the-iphone-5s-review/5 The only device which beats it is a quad-core tablet Silvermont, with considerably higher TDPs.

The fact that you don't see those faster designs in phones/desktops has multiple reasons, but the main reason is power consumption. Intel has now made an appropriate architecture for phones and tablets, called Silvermont, which has a much higher performance/watt than any other CPU. So it could potentially be faster than Cyclone, since phones and tablets are TDP constraint products, but instead Intel has chosen to keep power consumption sane with a well-balanced architecture that isn't only suitable for the highest-end devices.

You keep saying that Silvermont has higher performance/W; prove it. And no, graphs using "platform power consumption" don't count, as oddly enough the insanely high-res display in the iPad Air uses a lot of power; considerably more than the SoC.

Then give the evidence instead of making ridiculous claims or comparisons. Spoiler alert: you won't find it. Need a source for my claims? Read this.

Seriously, you're citing Ashraf? You do know that he posts on these forums as "Intel17", right? And that he owns stock in Intel? Does that seriously strike you as a reliable, unbiased source?

 

[Eug]

I'll give credit where its due but I refuse to believe that the A7 dominates so completely until I see OS independent benchmarks.

That's going to be difficult, since no other OS runs on A7. Or do you have a specific set of benchmarks on iOS in mind?

This is exactly how I see it. People here are trying to compare silicon to silicon, but they're not accounting for the software layer on top of them that often is more important than the hardware itself.

Apple has a huge advantage with being able to leverage their operating systems and extract the most out of their hardware. Intel does what it can, but ultimately doesn't have the level of control that Apple does.

In a way, you may be looking at it backwards, at least partially. Apple also designs its hardware to best support its own OS and software.

In the real world, yes, Apple's vertical integration advantage does matter, and it matters a lot. But when we're trying to discuss Silvermont vs. Cyclone in a vacuum, it doesn't. People need to be careful to recognize where software is winning out, rather than the hardware.

I don't think we necessarily should separate the two so much, when we're talking future development destined for future shipping devices. Hell, Apple doesn't even provide detailed specs on its iDevices. The only reason we know is due to teardowns and such.

 

[witeken]

Again, how many smartphones have Atom in them? How many non-subsidized tablets?

A few, and probably none. But if you're not happy with my claims that Silvermont is a good architecture, then just wait until Broxton. It will have fixed every problem you have with Bay Trail / Merrifield. Not just great performance and power, but also good wireless, GPU and other SoC IP, no BOM problems, 1 chip for both tablets and phones, and easy to make fast follow-ups to react to the competition.

They had built the uncore, but the A6 was their first CPU design.

Yes, but they've been focusing on mobile (SoCs) for a much longer time than Intel, not just from the iPhone 5 on.

Seriously? The iPhones 5S comes top of almost every single CPU performance bench that Anand throws at it: http://www.anandtech.com/show/7335/the-iphone-5s-review/5 The only device which beats it is a quad-core tablet Silvermont, with considerably higher TDPs.

TreVader made some claims about how fast Cyclone is. I responded to this by asking where Cyclone is so fast because there isn't 1 metric where Cyclone is fastest (if desktop CPUs are also allowed). My point is that he's neglecting power consumption, which is just as important as performance (performance/watt), in tablets and phones.

You keep saying that Silvermont has higher performance/W; prove it. And no, graphs using "platform power consumption" don't count, as oddly enough the insanely high-res display in the iPad Air uses a lot of power; considerably more than the SoC.

It's funny that you say that the display of the iPad Air uses a lot of power, since the Air uses an IGZO display: source.
The power consumption and performance/watt of the A7 has already been estimated a few posts ago by Enigmoid, you should read it.
But this comparison of Ashraf basically gives you all the information you need: Intel Vindicated, Very Competitive With Apple's A7. Don't forget that 2.4GHz quadcore Silvermont has been measured at using 2.5W.

Seriously, you're citing Ashraf? You do know that he posts on these forums as "Intel17", right? And that he owns stock in Intel? Does that seriously strike you as a reliable, unbiased source?

Yes, since a few weeks when I was talking about him to himself without knowing it was him for some posts .

Yes, I think he's a reliable source. I'm not just basing this opinion on some words like "he's own Intel" but because I've read quite some articles of him and I know that he isn't (very) biased or incompetent. Also, from reading the article myself. Even if his conclusions were wrong, he still gives the objective measurements where you could base your own conclusions upon, although I don't think that is particularly necessary.

 

[ams23]

You keep saying that Silvermont has higher performance/W; prove it. And no, graphs using "platform power consumption" don't count, as oddly enough the insanely high-res display in the iPad Air uses a lot of power; considerably more than the SoC.

It is certainly conceivable that Silvermont CPU cores have better perf. per watt than Cyclone CPU cores, but we would need to see app-specific data to know for sure.

Note that in the Krakken benchmark, the Cyclone CPU cores (in iPad Air) actually have virtually identical perf. per watt compared to the Swift CPU cores (in iPad 4). The former CPU consumes ~ 3w while the latter CPU consumes ~ 1w (at 3x slower performance) in this benchmark.

Total platform power at idle is 2w lower for iPad Air vs. iPad 4, so the important thing to look at is the difference between idle and peak power.

 

[Eug]

The fact that people have to defend Intel's engineering skillz in the face of the Apple onslaught is a testament to Apple's abilities.

I think the bottom line here is that Cyclone is excellent, and Apple is a new leading power in CPU design, something we wouldn't have thought possible just 5 years ago.

 

[ams23]

The fact that people have to defend Intel's engineering skillz in the face of the Apple onslaught is a testament to Apple's abilities.

I think the bottom line here is that Cyclone is excellent, and Apple is a new leading power in CPU design, something we wouldn't have thought possible just 5 years ago.

There is no denying that Cyclone is a very impressive and ambitious CPU design (especially for an ultra mobile CPU), but do keep in mind that Cyclone appears to consume up to ~ 3x more power with CPU-intensive applications than Swift, so CPU perf-per-watt is not necessarily better than their previous effort.

 

[Enigmoid]

That's going to be difficult, since no other OS runs on A7. Or do you have a specific set of benchmarks on iOS in mind?

Throw out any browser benchmarks. The optimizations apple does are ridiculous and you can easily see optimizations in the differences between sunspider scores using the same CPU on different android phones.

Apple can get basically Mantle level optimizations as they design the SOC and the browser.

Not too sure about geekbench. Its very encoding/encryption driven and the A7 is insanely strong with encryption (for the lot of good it does in most phones). The problem with geekbench is that the A7's insanely wide core is going to do very well--disproportionately well--when compared to consumer apps. The power use is also going to be quite high making use of that very wide core. And lets not forget that geekbench was initially designed as an Apple benchmark and is compiled completely differently.

3d mark physics would be somewhat useful but the A7 has poor random memory performance (and this is why it shows no improvement vs the A6). Bullet arranges the objects randomly in memory and the A7 sucks at random reads. Arrange them sequentially and performance increases.

 

[witeken]

The fact that people have to defend Intel's engineering skillz in the face of the Apple onslaught is a testament to Apple's abilities.

Or as testament to the human psyche :hmm:?

 

[NTMBK]

Or as testament to the human psyche :hmm:?

The guy with an Intel avatar is lecturing us on fanboyism? :hmm:

 

[Nothingness]

It is certainly conceivable that Silvermont CPU cores have better perf. per watt than Cyclone CPU cores, but we would need to see app-specific data to know for sure.

Has there ever been such measurements done? IIRC all we've heard is that BT is using 2.5W, but did someone ever run some benchmark in a loop? Or tried a power virus as Anand did on Swift and Cyclone? Might be interesting

There is no denying that Cyclone is a very impressive and ambitious CPU design (especially for an ultra mobile CPU), but do keep in mind that Cyclone appears to consume up to ~ 3x more power with CPU-intensive applications than Swift, so CPU perf-per-watt is not necessarily better than their previous effort.

It's nonetheless a nice achievement they could keep the same perf/W while tripling the performance. But I'd be surprised if they could achieve that again, while Intel have the process advantage that will help them.

 

[PPB]

Or as testament to the human psyche :hmm:?

There there, if you keep staying in denial, the harder it will be to go over the other 6 stages of grief :biggrin:

 

[witeken]

The guy with an Intel avatar is lecturing us on fanboyism? :hmm:

My avatar has nothing to do with this. I just like the Intel logo; it's clean and simple.

 

[ams23]

It's nonetheless a nice achievement they could keep the same perf/W while tripling the performance.

Yes, but do note that Cyclone also has the benefit of a more advanced [28nm] fab. process compared to Swift (which used a 32nm fab. process). The A7 SoC in total has a tad over 1 billion transistors in total IIRC!

 

[jdubs03]

This probably won't be a great comparison between the A7 and Bay Trail but, I figured I would at least introduce it.

http://www.notebookcheck.net/Review-Apple-iPad-Air-Tablet.105976.0.html

http://www.notebookcheck.net/Review-Asus-Transformer-Book-T100TA-C1-GR-Convertible.106219.0.html

 

[Exophase]

Total platform power at idle is 2w lower for iPad Air vs. iPad 4, so the important thing to look at is the difference between idle and peak power.

Still not a great way to estimate CPU power consumption in isolation. At the very least it doesn't take into consideration power consumed by memory or power regulators. And if you have anything going on the screen, which is the case for a lot of benchmarks reporting textual information, something is going to need to consume power rendering it. Even the LCD will probably have low power modes where it's in retention. And if the wireless went out of sleep to start downloading something for a benchmark it could still be on, or worse, if it's still communicating crap during.

Meanwhile, Intel is sure to get reviewers to rig up Atom systems measuring CPU rails directly, for comparison. Which is where the 1-2.5W numbers come from. Look at some idle to peak numbers at the wall for Baytrail devices and it's not so low.

 

[Nothingness]

This probably won't be a great comparison between the A7 and Bay Trail but, I figured I would at least introduce it.

http://www.notebookcheck.net/Review-Apple-iPad-Air-Tablet.105976.0.html

http://www.notebookcheck.net/Review-Asus-Transformer-Book-T100TA-C1-GR-Convertible.106219.0.html

If you pick a tablet with higher resolution, things are starting to look less good for BT:
http://www.notebookcheck.net/Review-HP-Omni-10-5600eg-F4W59EA-Tablet.108702.0.html
http://www.notebookcheck.net/Review-Lenovo-ThinkPad-8-Tablet.112867.0.html

I'm afraid that's still not enough to draw any conclusion.

 

[Khato]

Meanwhile, Intel is sure to get reviewers to rig up Atom systems measuring CPU rails directly, for comparison. Which is where the 1-2.5W numbers come from. Look at some idle to peak numbers at the wall for Baytrail devices and it's not so low.

And instead of applauding Intel for providing reviewers a capability that none of the other players in this market offer what response does Intel get? Yeah...

While I continue to appreciate Notebookcheck's efforts with respect to measuring power consumption it's pretty clear that their methodology is flawed. Exhibit A being the 3.3W delta for the iPad air between maximum idle and maximum load when we have contradictory evidence from Anandtech's review. My suspicion of such being that their 'load' test for Android, iOS, and maybe even OSX needs some work/isn't comparable to what they run in x86 Windows.

Even if you do take the Notebookcheck delta numbers for Baytrail at face value and use them for comparison to Anandtech's iPad Air numbers you're left with ~8W for Baytrail and, oh yeah, 8W for A7. Key difference being that that's a maximum of 2W per core for Silvermont and 4W per core for Cyclone. (Note that I'd love it if Apple provided the same level of detail/review assistance as Intel, but it's more likely that Intel would provide a wired-up iPad Air and measurement equipment for reviewers to use than Apple...)