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

[TreVader]

how about them apples in your signature? didn't it hurt deep inside to know they were essentially flawed by using intel silicon?

I think they would and will be better with Apple silicon. I'd hate to think of apple building an iPhone or iPad with an Atom SoC LOL!

 

[witeken]

Last time I checked my Galaxy S4 isn't using intel silicon, neither is any iPhone or android phone i've considered buying ever.

Intel didn't start heavily investing in mobile since about 2010. It can take 4 years to design a new architecture so you can do the math.

It's pretty apparent from the fact that intel kept recycling it's awful netburst architecture (because they suck at designing) for some 4 years past the time when it was obsolete. Last time I checked, qualcomm doesn't sit on 10 year old designs while it loses marketshare to AMD or whomever. It couldn't, mobile is too competitive. If intel didn't have those fabs and a gargantuan budget, it wouldn't exist.

An Intel engineer disagrees with you:

In my mind, Netburst, much as it's maligned, brought some very good things internally for Intel design teams. First, unbelievable circuit expertise (the FP logic was running at 8GHz in Prescott stock!). Next, the trace cache which you can see reimplemented in Sandy and Ivy Bridge. Also, SMT. Building a validation team that could validate the beast pre- and post-silicon. The power-perf thinking i.e. frequency through power savings. Finally, the development of tools and project management required to do that kind of extreme design. All of these learning continue to this day and it's a very large contributor to why in client and server CPUs Intel can sustain the roadmap we have.

 

[witeken]

I think they would and will be better with Apple silicon. I'd hate to think of apple building an iPhone or iPad with an Atom SoC LOL!

Maybe if they can use Intel as a foundry. Else it would be funny to see the process node going from 22nm Tri-Gate to 28nm instead of 14nm.

 

[scannall]

Intel stopping as IDM would be the most stupid they could do. At the current rate, Intel could make every other foundry obsolete as serious manufacturing choice within ~6 years. That is their business model.

Not unless they plan on making their own litho machines. Their R&D keeps them a little ahead, but they are dependent on ASML for their equipment. Just like every other foundry.

http://www.asml.com/asml/show.do?ctx=427

 

[witeken]

If ASML is able to keep up with Intel's 2 year cycle, there won't be a lot of problems. TSMC will gradually fall behind because their leading process isn't their biggest source of income, so the cycles will take longer as the costs keep getting higher.

 

[Ajay]

Not unless they plan on making their own litho machines. Their R&D keeps them a little ahead, but they are dependent on ASML for their equipment. Just like every other foundry.

http://www.asml.com/asml/show.do?ctx=427

Well, while ASML's capabilities are probably a limiting factor for Intel's node development progress, they are not for other fabs. I wish IDC was here, he'd likely have some better insight on this. People underestimate the depth and breadth of Intel's process development R&D team.

 

[Arachnotronic]

Well, while ASML's capabilities are probably a limiting factor for Intel's node development progress, they are not for other fabs. I wish IDC was here, he'd likely have some better insight on this. People underestimate the depth and breadth of Intel's process development R&D team.

Intel is cool to hate.

 

[scannall]

Well, while ASML's capabilities are probably a limiting factor for Intel's node development progress, they are not for other fabs. I wish IDC was here, he'd likely have some better insight on this. People underestimate the depth and breadth of Intel's process development R&D team.

I understand Intel has a lot of the best and brightest minds in their development. Certainly, I'm not intending to disparage them at all. Have a look at the craziness the litho machines have to do now, and what they will have to do down the road. Everyone's node shrinks are going to come slower. I'm not seeing any reason Intel won't stay in the lead, I'm just seeing that lead getting more costly and time consuming.

 

[Khato]

Not unless they plan on making their own litho machines. Their R&D keeps them a little ahead, but they are dependent on ASML for their equipment. Just like every other foundry.

http://www.asml.com/asml/show.do?ctx=427

Pssst, just because they're number one doesn't mean that ASML is the only supplier of lithography machines. Want to guess who the 'major device manufacturer' who ordered Nikon's 450mm ArF immersion scanners almost a year ago was? http://www.nikon.com/news/2013/0704_450mm_01.htm The Intel investment in ASML to accelerate EUV and 450mm equipment was quite well disclosed whereas about all you get on the Nikon side is confirmation that they are working together and Intel is paying for some percentage of the development costs. It makes sense that Intel would want two suppliers - ASML because then it can get the entire industry to switch and maximize the cost savings and then Nikon because of exactly what happened with ASML.

If ASML is able to keep up with Intel's 2 year cycle, there won't be a lot of problems. TSMC will gradually fall behind because their leading process isn't their biggest source of income, so the cycles will take longer as the costs keep getting higher.

I continue to be of the opinion that the largest risk to TSMC and really all of Intel's competition is Intel venturing into the foundry space. If Intel's actually successful with their 14nm customers and gets more on board for 10nm, well, every foundry customer that Intel wins is one that someone else loses. And losing leading-edge process technology customers makes it that much more difficult to finance the next node at the same pace.

 

[TreVader]

Intel is great at manufacturing, the best even. But they suck at designing new products. They had to be dragged, kicking and screaming, into 64 bit by AMD64 and it's pretty sad that with 100 billion dollars in assets they still can't compete with the likes of qualcomm on SoC design.


intel better be scared, because Apple has the money to build better fabs and the talent to make better chips. How many iterations has the Atom gone through without being competitive? Apple designs one chip, and they're at the top of the mobile game.

 

[TreVader]

Maybe if they can use Intel as a foundry. Else it would be funny to see the process node going from 22nm Tri-Gate to 28nm instead of 14nm.

Apple has already designed an SoC that outperforms intels Atom by a huge margin on 28nm. So I guess that's exactly what you're seeing right now?



Idk where you got 14nm from, last time I checked intel hasn't even released Haswell- E on 20nm LOL

 

[witeken]

I'm just seeing that lead getting more costly and time consuming.

The way I see it is that as long as you can afford it, it's definitely worth going to a lower node. Not only do you get a useful increase in power and speed, but mainly the cost a of transistor drops a lot. So you will have both faster, less power hungry chips and extra die space to make chips that are virtually impossible on older process nodes, which will give you more money in return and higher margins.

At the end of Moore's law, things may change and it seems that Intel is also focusing a lot on improving costs, as Intel advertises 14nm and 10nm as a bigger shrinks than usual, and 450mm should also give an improvement of about 1 node.

 

[Bolshoi Booze]

Apple has already designed an SoC that outperforms intels Atom by a huge margin on 28nm. So I guess that's exactly what you're seeing right now?



Idk where you got 14nm from, last time I checked intel hasn't even released Haswell- E on 20nm LOL

LOL

http://browser.primatelabs.com/geekbench2/compare/2307544?baseline=2395855
http://browser.primatelabs.com/geekbench3/compare/507271?baseline=497549

excuse me, what did you say? cause I dont see A7 outperforming bay trail by a huge margin ..
I like apple's approach of using 2 stronger cores for mobile but BT still wins on MT and likely uses less power..
intel has 22nm chips since 2012 and they will launch 14nm chips later this year, while apple will be stuck with 20nm TSCM for their A8

 

[jpiniero]

I continue to be of the opinion that the largest risk to TSMC and really all of Intel's competition is Intel venturing into the foundry space. If Intel's actually successful with their 14nm customers and gets more on board for 10nm, well, every foundry customer that Intel wins is one that someone else loses. And losing leading-edge process technology customers makes it that much more difficult to finance the next node at the same pace.

This is true, but Intel at this point hasn't been willing to fab an actual legitimate competitor to their x86 processors yet. There's probably not that much business the foundry can do without doing that.

 

[witeken]

Intel is great at manufacturing, the best even. But they suck at designing new products. They had to be dragged, kicking and screaming, into 64 bit by AMD64 and it's pretty sad that with 100 billion dollars in assets they still can't compete with the likes of qualcomm on SoC design.

For your information, Intel acquired Infineon in 2011. That's also the reason that AMD won't compete in mobile, even though they have Jaguar.

intel better be scared, because Apple has the money to build better fabs and the talent to make better chips. How many iterations has the Atom gone through without being competitive? Apple designs one chip, and they're at the top of the mobile game.

The first iteration of Atom, the Saltwell architecture was already competitive in performance/watt. But it suffered from issues like being built on an n-1 process node. Silvermont now it without doubt at the top in term of performance/watt and second behind Apple in term of IPC.

 

[TreVader]

For your information, Intel acquired Infineon in 2011. That's also the reason that AMD won't compete in mobile, even though they have Jaguar.

The first iteration of Atom, the Saltwell architecture was already competitive in performance/watt. But it suffered from issues like being built on an n-1 process node. Silvermont now it without doubt at the top in term of performance/watt and second behind Apple in term of IPC.

Silvermont is by no means "without a doubt" anything whatsoever. It can't even compete with the Adreno320 in graphics. It's a joke, and that's why nobody is using it.



To this day AMD still produces a better SoC in terms of all around performance, and Silvermont has nothing on S800 let alone S805 in perf/watt or graphics. It has higher IPC and higher compute under load, and it has lower idle but it still sucks power when doing anything other than idling.


How can you claim intel is anywhere in mobile when it hasn't got a single big design win at all? Tegra is all over intel and Tegra is a joke.

 

[Khato]

This is true, but Intel at this point hasn't been willing to fab an actual legitimate competitor to their x86 processors yet. There's probably not that much business the foundry can do without doing that.

Quite correct, but there are a few reasons for such. Even if Intel was allowed, say, Qualcomm to use their fabs it likely wouldn't happen right away. Moving foundries requires a decision pretty early in the design process and represents a large risk, especially since Intel's foundry business is pretty much unproven. Sure they can make products for their internal design teams, but there are quite a few differences between that and normal foundry work. Which is why my guess is that Intel's luring its current customers in with pretty good prices/other incentives to get the ball rolling. After all, it's a longer term play and it'll be quite interesting to see how it unfolds.

 

[TreVader]

LOL

http://browser.primatelabs.com/geekbench2/compare/2307544?baseline=2395855
http://browser.primatelabs.com/geekbench3/compare/507271?baseline=497549

excuse me, what did you say? cause I dont see A7 outperforming bay trail by a huge margin ..
I like apple's approach of using 2 stronger cores for mobile but BT still wins on MT and likely uses less power..
intel has 22nm chips since 2012 and they will launch 14nm chips later this year, while apple will be stuck with 20nm TSCM for their A8

So 14nm magically makes intel the leader in mobile SoC? I'm pretty sure they have to design a decent chip first, which bay trail isn't. Bay trail is already on a process node advantage and still gets it's a$$ kicked by the A7 (apple's FIRST EVER chip). What makes you think that an A8 on 20nm is going to magically force apple to lose to intel?


Process node advantage gets less and less with each node, so intels fabs mean less and less every generation.

 

[TreVader]

I think given 3 years intel could've been competitive in mobile, and by competitive I mean getting a lot of design wins. They need to stop continually recycling old architectures (like Bay Trail) and start from scratch.


The problem with intel is it's all about the long con and playing it safe. Intel wants to stay consistently profitable for the foreseeable future, they aren't trying to take over the world or to cure cancer. They're just grinding out earnings.

Apple wants to take over the world. Intel wants to be the best at what they do (making chips). Personally, I like Apple's ambition.

 

[Sweepr]

Silvermont is by no means "without a doubt" anything whatsoever. It can't even compete with the Adreno320 in graphics. It's a joke, and that's why nobody is using it.

Intel Celeron J1850 - note that this is not the fastest BT chip graphics wise.
3DMark Ice Storm @ Android 4.1: 16638

http://www.futuremark.com/hardware/mobile/Acer+Aspire+AA3-600/review

Now Qualcomm's S600 - Adreno 320 (ouch), easily competitive and after many months it seems like Intel improved Bay Trail's graphics performance, especially under Android. Also, it might be true that BT didnt show up in many Android devices, but there are lots of smaller than 11'' Windows convertibles, 2-in-1s, notebooks and tablets based on it - they might not be as popular as iPads but they're usually much cheaper with fairly good battery life and much better performance than older Atom designs.

To this day AMD still produces a better SoC in terms of all around performance

Hm, no, definitely not for tablets. Beema/Mullins are still MIA and AMD's only x86 tablet CPU right now - A4-1200 is a joke performance wise, Bay Trail-T is almost 3x faster in MT workloads, 50% faster in ST workloads. Even the iGPU (usually AMD's strongest part) didnt deliver better gaming performance (in a <12'' tablet) according to Notebookcheck. The only <12'' tablet based on it - MSI W20 - barely delivers 5 hours battery life with clockspeeds capped @ 600MHz and a bigger battery than the ASUS T100 (Bay Trail-T based tablet with ~10 hours WiFi battery life).

http://forums.anandtech.com/showthread.php?t=2341969&page=78

 

[witeken]

Silvermont is by no means "without a doubt" anything whatsoever. It can't even compete with the Adreno320 in graphics. It's a joke, and that's why nobody is using it.

I didn't claim anything about the GPU. Cherry Trail will be the SoC to take GPU performance lead (BT is between Adreno 320 and 330, and CT will have a major redesigned gen8 GPU with 4x as much EUs). BTW, is there any Android game BT can't run?

Calculations show that Silvermont has the highest performance/watt, which Homeles already explained very simple: it has a process node advantage + Tri-Gate.

 

[witeken]

So 14nm magically makes intel the leader in mobile SoC? I'm pretty sure they have to design a decent chip first, which bay trail isn't. Bay trail is already on a process node advantage and still gets it's a$$ kicked by the A7 (apple's FIRST EVER chip). What makes you think that an A8 on 20nm is going to magically force apple to lose to intel?


Process node advantage gets less and less with each node, so intels fabs mean less and less every generation.

Why do you think Bay Trail isn't a decent chip? Apple will lose to Intel because it is 1 process node behind. Process node advantages are indeed not so high because of leakage, but that's why Intel's got Tri-Gate. You can't beat a 14nm Tri-Gate with 20nm planar...

 

[TreVader]

I didn't claim anything about the GPU. Cherry Trail will be the SoC to take GPU performance lead (BT is between Adreno 320 and 330, and CT will have a major redesigned gen8 GPU with 4x as much EUs). BTW, is there any Android game BT can't run?

Calculations show that Silvermont has the highest performance/watt, which Homeles already explained very simple: it has a process node advantage + Tri-Gate.

Intel can be depended on to disappoint on graphics. I pretty much guarantee you cherry trail isn't gonna touch the A8 in that arena.



Bay Trail is a lot like Tegra, it sorta works but it's not exceptional in any way.

 

[TreVader]

Why do you think Bay Trail isn't a decent chip? Apple will lose to Intel because it is 1 process node behind. Process node advantages are indeed not so high because of leakage, but that's why Intel's got Tri-Gate. You can't beat a 14nm Tri-Gate with 20nm planar...

I think Bay Trail is a decent chip, but I expect more from a company that has nearly infinite R&D funds, the best fabs in the world, and 30 years experience to make the best chip. Bay trail is not the best mobile chip by a long shot.

 

[Homeles]

I think Bay Trail is a decent chip, but I expect more from a company that has nearly infinite R&D funds, the best fabs in the world, and 30 years experience to make the best chip. Bay trail is not the best mobile chip by a long shot.

You know Apple's got somewhere around 10x what Intel does at least, right?