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

[krumme]

Apple already has quite a lot of control on Intel's roadmap.

Besides full control they also wants all the money. Its very simple. Paying for laptop cpu is double stupid because it gives Intel money to enter their own territory.

 

[TreVader]

Sunspider 1.0
2x Cyclone @ 1.4GHz - 384ms
4x Haswell i7 @ 800Mhz - 387ms (Win 8.1, IE 11) And guess what, most of the work happens on one thread. (on the Win machine at least)
2x Haswell i7 @ 1400Mhz - 215ms (IE11 process affinity was restricted to CPU cores 0 & 2)

Where did these numbers come from? You type them out yourself and don't quote a source?



I think we should compare intels 1st processor (8088?) to apple's first (A7) and see which company did a better job

 

[krumme]

As a sidenote. Apple is hiring all day for the camera technology. In general i think their future strategy in to design all the critical technology themselves. Its also a way to use all their money and calm their voicing shareholders.

 

[coercitiv]

Where did these numbers come from? You type them out yourself and don't quote a source?

I'm benching as we speak, all you need is a PC and a browser, remember?

I think we should compare intels 1st processor (8088?) to apple's first (A7) and see which company did a better job

Interesting argument considering your previous comments.

Sunspider. Actually if you take A7 sun spider scores Apple is beating intel already on a clock/clock core/core basis.

 

[witeken]

In practical terms, Apple is already beating Intel in mobile processors. Maybe intel isn't all that amazing after all? Maybe the only thing it takes to design a CPU that beats Core is deep pockets (hence AMDs failure to beat intel).

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.

 

[Nothingness]

Nice straw man, since when does performance not matter anymore?

You definitely missed my point, but I'm not surprised, we never understand each other. And I hope that will stay this way :biggrin:

It's a browser benchmark. It's useless when doing cross platform comparisons.

It's even useless on the same platform: remember how Piednoel complained Anandtech should have used IE for Silvermont Sunspider scores.

Alas the state of benchmarking on mobile isn't improving.

 

[TreVader]

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.

How many mobile phones have shipped with intel processors?





Yeah, that's what I thought.

 

[Nothingness]

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.

Intel phone SoC is twice faster than Apple's one?

/me look at the date, OK got it, but it's not funny.

 

[witeken]

I think we should compare intels 1st processor (8088?) to apple's first (A7) and see which company did a better job

This is hilarious. You're comparing the biggest achievement of all time, building a microprocessor, to an Apple CPU made with modified off-the-shelf ARM IP which is made possible by multiple decades of innovation? And then I'm still ignoring the fact that today's chips are 350² times more transistor dense.

Without doubt Intel did a better job.

 

[witeken]

Intel phone SoC is twice faster than Apple's one?

/me look at the date, OK got it, but it's not funny.

Read carefully, I've said performance/watt. If you still don't understand, I can explain it to you: when you put fast core at high clock speeds in a phone, like Qualcomm's Snapdragon 800, the thing can't sustain it's highest clock speed for a meaningful time. Within a few seconds, the chip will start throttling, effectively reducing performance. When you replace that with a CPU that has double the performance per watt, the CPU won't throttle, effectively doubling CPU performance (at the same TDP).

Sure, you can note that Intel doesn't ship any phone SoCs yet, or that the BOM is to high, or that no one will buy them, or that you just want good battery life because you don't use your phone for intensive tasks, and that's all fine, but that's completely irrelevant to the point I make. Those remarks are separate discussions (which I'm fine to discuss, but you mustn't use them to disprove my comments).

 

[TreVader]

This is hilarious. You're comparing the biggest achievement of all time, building a microprocessor, to an Apple CPU made with modified off-the-shelf ARM IP which is made possible by multiple decades of innovation? And then I'm still ignoring the fact that today's chips are 350² times more transistor dense.

Without doubt Intel did a better job.

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.

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.

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.

 

[Homeles]

It's even useless on the same platform: remember how Piednoel complained Anandtech should have used IE for Silvermont Sunspider scores.

Alas the state of benchmarking on mobile isn't improving.

Yeeeeeeepppp. The validity of various benchmarks doesn't really matter anyway -- we've got a lot of people around here and elsewhere that believe that the only benchmark that counts is the "me" benchmark. Unfortunately, humans are horribly biased and are terrible instruments for measuring just about anything.

 

[Bolshoi Booze]

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.

you better be joking, seriously

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.

I wonder why you have 3 intel-based laptops then :awe:

Sunspider 1.0
2x Cyclone @ 1.4GHz - 384ms
4x Haswell i7 @ 800Mhz - 387ms (Win 8.1, IE 11) And guess what, most of the work happens on one thread. (on the Win machine at least)
2x Haswell i7 @ 1.4Ghz - 215ms (IE11 process affinity was restricted to CPU cores 0 & 2)
1x Haswell i7 @ 1.4Ghz - 270ms +/-10% (running the benchmark on one core only seems to affect confidence)

ouch, not even close..
I hope we'll see some core-based tablets starting with broadwell... there are rumours about a 4.5w part

 

[Exophase]

This is hilarious. You're comparing the biggest achievement of all time, building a microprocessor, to an Apple CPU made with modified off-the-shelf ARM IP which is made possible by multiple decades of innovation?

If you don't understand that Cyclone isn't modified off-the-shelf ARM IP then you probably shouldn't comment about it. You'd may as well claim that AMD's processors use modified off-the-shelf Intel IP.

 

[Arkaign]

If you don't understand that Cyclone isn't modified off-the-shelf ARM IP then you probably shouldn't comment about it. You'd may as well claim that AMD's processors use modified off-the-shelf Intel IP.

But that's exactly what AMD's processors are

Look at the history, x86, AMD's inception, and the x86 licensing. AMD eventually branched out a bit, but the basic x86 design hasn't actually changed all that dramatically.

ARM is ARM, x86 is x86, you just have different added features/emphasis over time.

 

[Khato]

If you don't understand that Cyclone isn't modified off-the-shelf ARM IP then you probably shouldn't comment about it. You'd may as well claim that AMD's processors use modified off-the-shelf Intel IP.

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?

On another note, the state of cross-platform mobile benchmarking reminds me quite a bit of the marketing fluff that Apple used to come up with back in the PowerPC G3-G5 era. Which is to say that there's no platform-independent method for measuring performance so we're left with results which are skewed in one direction or the other and no good reference point. As many have unintentionally pointed out, the best measure of how good Apple's CPU design team actually is would be how soon they stop using Intel in their products.

 

[North01]

Intel has had 30 years to design a decent mobile processor and they just can't.

I disagree, Bay Trail-T is a solid SoC. It finally enabled manufacturers to produce low cost, light weight, high efficiency, Windows 8.1 (x86) devices that actually have respectable performance.

 

[Arkaign]

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?

On another note, the state of cross-platform mobile benchmarking reminds me quite a bit of the marketing fluff that Apple used to come up with back in the PowerPC G3-G5 era. Which is to say that there's no platform-independent method for measuring performance so we're left with results which are skewed in one direction or the other and no good reference point. As many have unintentionally pointed out, the best measure of how good Apple's CPU design team actually is would be how soon they stop using Intel in their products.

This is very true, and will be interesting to follow.

Apple risks repeating the same situation that led to their G5 coffin nail for Motorola if Intel continues to massively improve their mobile offerings. If OTOH Intel doesn't continue to massively improve, then it could be a sage business decision.

I'm thinking Apple is hedging their bets here. It's not a bad idea to have a solid in-house processor team to fill in blanks and build things specifically to their needs. And at the same time, they can't go all-in unless their performance/power/heat/feature set remains within at least a comparable scale to non-Apple competition. If 2016 sees 10nm Intel products that are 50% faster with 35% lower heat/power usage, and similar/equal feature sets, then it would be market suicide to not use Intel products in their high-ASP lines.

Betting against Intel on this one (mobile) seems .. risky. But there is definitely room in the marketplace for a huge multitude of processors in various devices, so even if Intel extends their dominance down into areas like midrange tablets, there are plenty of other places to sell chips.

 

[Khato]

I'm thinking Apple is hedging their bets here. It's not a bad idea to have a solid in-house processor team to fill in blanks and build things specifically to their needs.

While I'll freely admit that there's a good chance I'll be wrong, I still suspect that there's a decent chance of Apple's next major CPU architecture targeting a completely different market with the next iteration or possibly two of the smartphone/tablet SoC simply being tweaks of Cyclone on new processes. Why? Because Apple's success/major profits comes from inventing new markets. Continuing to design custom SoCs for their smartphone/tablet is a waste of their design team soon as a commodity part will do the job.

 

[Enigmoid]

Sunspider. Actually if you take A7 sun spider scores Apple is beating intel already on a clock/clock core/core basis.
And they're just getting started.

Sunspider is pretty much the worst test you can use its so browser and OS dependent.

Look at any browser test. The majority of the gains from the A6 to A7 are from software.

Look at the score increase from the iphone 5 launch to 1 year later with IOS 7. If you compared vs. the launch scores you would think the A7 was a much much bigger jump.

From 1672 to 2859 points. 71% increase.

When you actually look at Kraken score vs power usage.

(AT doesn't have ipad 4 benchmarks with IOS7 so I'm going to approximate the that chips are similar enough to do this. ex) the different in power consumption on the ipads is directly proportional to the difference in performance on the phones. Performance seems to differ by less than 10% so this seems reasonable)

Looking at the steady state power.

A6X - 1.2W delta
A7 - 3W delta

Performance increase = 14,000/5,000 = 237%
Power increase = 250%
Not seeing any better perf/W.

People need to realize that Apple designs the software and the hardware for these phones and can get crazy amounts of efficiency.

A7 vs S800 is like comparing a FX-4300 + Mantle to a i4-4770k + DX.

Go back in time and you can easily see this.

A single core A8 at 800 mhz is just thrashing those dual core 1.2 ghz A9 Android phones despite essentially the same CPU architecture. In fact that looks a lot like the playing field today with A7 vs. S800 but with quite obviously a significantly weaker phone.

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

 

[Exophase]

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?

Not really. It's not an exaggeration, it's just plain wrong. A lot of people think that Qualcomm and Apple are taking ARM Cortex-A series cores and changing pieces of them to get their cores, which isn't what's happening at all.

But that's exactly what AMD's processors are

That's what they were 30 years ago. I meant AMD processors today.

Look at the history, x86, AMD's inception, and the x86 licensing. AMD eventually branched out a bit, but the basic x86 design hasn't actually changed all that dramatically.

This mentality is way off. Only relatively small parts of a CPU are dependent on the instruction set. The rest of the CPU, the parts that really matter, have changed a lot.

That, and the x86 instruction set has changed tremendously, and for that matter so has ARM.

 

[Homeles]

People need to realize that Apple designs the software and the hardware for these phones and can get crazy amounts of efficiency.

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

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 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.

 

[pw257008]

Related question: Would Apple be able to as effectively design its software around Intel's silicon as it does around its own? Because if not, then at least in Apple products, Apple's chips are superior.

 

[Arkaign]

Related question: Would Apple be able to as effectively design its software around Intel's silicon as it does around its own? Because if not, then at least in Apple products, Apple's chips are superior.

Look at Intel-powered OSX Macbooks, iMacs, and Mac Pros. If anything, they're better than ever with Intel CPUs.

For mobile, it gets much murkier. If the competition doesn't improve much, then Apple has every reason to use their own in-house Arm-based derivatives. Their software already works with them, so from an ease of production perspective it's better to go status quo and use what they have (with future iterations each gen of course).

The wild card is the competition (premium Android devices). If Intel gets to 14nm/10nm and starts drastically beating arm products in package size, perf/watt, etc, then Apple will be hard pressed to stick with their own products. This is of course assuming that Intel prices them at an acceptable market-bearing level. Then it becomes a no-brainer for Samsung to bring out the Galaxy X1 to replace the S-series, if it means that their new flagship is faster, thinner, and has better battery life than the competition.

As far as we've come with mobile processors, there is still very, very far to go.

 

[Arachnotronic]

While I'll freely admit that there's a good chance I'll be wrong, I still suspect that there's a decent chance of Apple's next major CPU architecture targeting a completely different market with the next iteration or possibly two of the smartphone/tablet SoC simply being tweaks of Cyclone on new processes. Why? Because Apple's success/major profits comes from inventing new markets. Continuing to design custom SoCs for their smartphone/tablet is a waste of their design team soon as a commodity part will do the job.

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.