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

[Khato]

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.

I do wonder about this. The A7 in the iPhone 5s only scores 13,869 on 3Dmark Ice Storm compared to the 16,638 for a Baytrail based Celeron which Sweepr linked. Now on the A7 Apple is already using Imagination's PowerVR G6430, so unless Apple switches their graphics IP provider on the A8 they're limited to Imagination's PowerVR GX6650 at best. If Imagination's 50% performance improvement claim for Series 6XT over Series 6 is to believed and we assume perfect scaling that'd still only be a 2.25x performance improvement in graphics... Which isn't so great when you figure that Intel's both moving from IVB-based graphics to BDW-based and quadrupling the number of EUs... and they are already ahead anyway.

Really I suspect that the above is the reason why Apple started putting together their own GPU design team - Imagination isn't scaling up their designs as fast as Apple wants.

 

[NTMBK]

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.

That's certainly an... interesting way of looking at it. What you basically say is that TSMC has a healthy long tail on revenue from their process nodes, meaning that they have a ridiculously long time for them to turn a profit and make back their R&D costs. Intel, meanwhile, realistically has ~4 years to turn an ROI on their new nodes before they become completely useless.

It's a big problem for Intel, and its getting worse. They used to use their N-1 nodes for the northbridge and N-2 for the southbridge. But recently these external nodes are shrinking rapidly. On an old Core 2 system you would have a pretty beefy northbridge with both the memory controller and the integrated graphics, and a southbridge with IOs. Now Intel is down to just a single PCH on their Haswell line, and no external PCH at all on their Bay Trail SoCs. So what are they going to use to fill those older fabs now? On package memory is one potential avenue- they already tested the waters with Crystalwell, and Knight's Landing is going all in, apparently with up to 16GB on-package. But cost is going to be the problem for the consumer market.

 

[NTMBK]

I do wonder about this. The A7 in the iPhone 5s only scores 13,869 on 3Dmark Ice Storm compared to the 16,638 for a Baytrail based Celeron which Sweepr linked. Now on the A7 Apple is already using Imagination's PowerVR G6430, so unless Apple switches their graphics IP provider on the A8 they're limited to Imagination's PowerVR GX6650 at best. If Imagination's 50% performance improvement claim for Series 6XT over Series 6 is to believed and we assume perfect scaling that'd still only be a 2.25x performance improvement in graphics... Which isn't so great when you figure that Intel's both moving from IVB-based graphics to BDW-based and quadrupling the number of EUs... and they are already ahead anyway.

Really I suspect that the above is the reason why Apple started putting together their own GPU design team - Imagination isn't scaling up their designs as fast as Apple wants.

What was the TDP on that Celeron, and how much power do you think the iPhone 5S SoC was allowed to draw down? And the desktop Bay Trail platform has dual-channel memory support, but the tablet Bay Trail does not. Don't compare a desktop platform and a mobile phone and expect fair results.

 

[coercitiv]

What was the TDP on that Celeron, and how much power do you think the iPhone 5S SoC was allowed to draw down? And the desktop Bay Trail platform has dual-channel memory support, but the tablet Bay Trail does not. Don't compare a desktop platform and a mobile phone and expect fair results.

The Atom is dual channel too, although it has lower bandwidth due to memory clocks. It also sports lower GPU max clocks.

Atom Z3770 Icestorm 3D Mark Score - 13850
iPhone Cyclone - 13869

 

[jpiniero]

That's certainly an... interesting way of looking at it. What you basically say is that TSMC has a healthy long tail on revenue from their process nodes, meaning that they have a ridiculously long time for them to turn a profit and make back their R&D costs. Intel, meanwhile, realistically has ~4 years to turn an ROI on their new nodes before they become completely useless.

It's a big problem for Intel, and its getting worse. They used to use their N-1 nodes for the northbridge and N-2 for the southbridge. But recently these external nodes are shrinking rapidly. On an old Core 2 system you would have a pretty beefy northbridge with both the memory controller and the integrated graphics, and a southbridge with IOs. Now Intel is down to just a single PCH on their Haswell line, and no external PCH at all on their Bay Trail SoCs. So what are they going to use to fill those older fabs now? On package memory is one potential avenue- they already tested the waters with Crystalwell, and Knight's Landing is going all in, apparently with up to 16GB on-package. But cost is going to be the problem for the consumer market.

I think the answer is that a larger portion of chip sales is going to be on the previous node, ala Haswell Refresh.

 

[witeken]

That's certainly an... interesting way of looking at it. What you basically say is that TSMC has a healthy long tail on revenue from their process nodes, meaning that they have a ridiculously long time for them to turn a profit and make back their R&D costs. Intel, meanwhile, realistically has ~4 years to turn an ROI on their new nodes before they become completely useless.

It's a big problem for Intel, and its getting worse. They used to use their N-1 nodes for the northbridge and N-2 for the southbridge. But recently these external nodes are shrinking rapidly. On an old Core 2 system you would have a pretty beefy northbridge with both the memory controller and the integrated graphics, and a southbridge with IOs. Now Intel is down to just a single PCH on their Haswell line, and no external PCH at all on their Bay Trail SoCs. So what are they going to use to fill those older fabs now? On package memory is one potential avenue- they already tested the waters with Crystalwell, and Knight's Landing is going all in, apparently with up to 16GB on-package. But cost is going to be the problem for the consumer market.

This article explains it: Intel And TSMC: The Unwinnable Struggle

So for every dollar of, let's call it operating capital spending, Intel spends, it generates $2.55 in revenue and $1.95 in gross profit. On the other hand, TSMC returns a $1.02 in revenue for every $1 it spends on bleeding edge technology, and only $.51 of gross profit.

Obviously, this is an unsustainable situation for TSMC; the company HAS to be losing big time on every 28nm wafer it ships. The other 70% of the business is subsidizing the new process efforts.

Thus, $1 of TSMC cost, at 50% gross margin, sells for $2. $1 of Intel cost, at 77% gross margin, sells for $4.35. That is another number that TSMC will never achieve, largely because the fabless business model can't afford it.

The unavoidable conclusion here is that TSMC simply has to stop spending money trying to catch Intel; it is at an impossible disadvantage.

Intel's leading edge fabs generate ~90% of the company's income, while TSMC needs older process nodes to fund hardly profitable n-1 process nodes. The desktop market gives Intel a very solid amount of cash that it can spend to step into new markets such as smartphones and tablets.

 

[NTMBK]

This article explains it: Intel And TSMC: The Unwinnable Struggle





Intel's leading edge fabs generate ~90% of the company's income, while TSMC needs older process nodes to fund hardly profitable n-1 process nodes. The desktop market gives Intel a very solid amount of cash that it can spend to step into new markets such as smartphones and tablets.

That piece of analysis is pretty heavily flawed.

-It compares value of chips sold to end user (Intel) with value of chips shipped to manufacturers (TSMC). This means it is excluding costs such as sales, marketing, warehousing, and so on and so forth which TSMC does not have to deal with.
-It compares only a single year's revenues from only the bleeding edge- which as the article points out is only 30% of TSMC's revenues. This is ignoring the exact point that we were just talking about. TSMC's nodes remain commercially useful for many, many years, providing a very long tail for them to repay their R&D. Their long lifespan makes investing in the R&D up front more palatable, not less. Most of their revenues today are a result of massive investments made many years ago, and today's R&D costs are paying for the next 5+ years of revenue. Ignoring this is fundamentally misunderstanding TSMC's business model.

These are very basic errors to make, which render his conclusions basically worthless.

 

[witeken]

I've got another article for you, then: Intel: The End Of Moore's Law. And a picture:

You can also look at the hard facts: Intel will start volume production of 10nm in Q4 2015, less than 2 years after 14nm. TSMC will start shipping 20nm FinET at the end of 2015 or 2016. Then it will take another 2 years before TSMC has 10nm, which is the same as what Intel calls 14nm, so let's say in H2 2017.

Intel 14nm will be available in H2 2013, so TSMC will be 3 years behind. he only good thing is that FinFET will only be 3.5 years behind (2012->2016) instead of 5.

Edit: Another nice picture:

 

[nonworkingrich]

OK Russ, we got it - Ashraf Eassa is not the only seekingalpha writer that operates here under a different name. Good luck with your investment & happy pumping!

 

[Homeles]

TSMC's revenue is now up to over 20 billion annually.

 

[Nothingness]

These are very basic errors to make, which render his conclusions basically worthless.

Russ Fisher sometimes makes interesting points/speculations (like Intel and memory), but his uninterrupted flow of article about how great Intel is has become boring.

At least Ahsraf Eassa has shown some varying point of views and when he's wrong he just admits it. I have yet to see Russ Fisher doing so

Of course that doesn't make him always wrong, I just have a hard time trusting people always showing the same side of the coin.

/end of rant :biggrin:

 

[NTMBK]

Russ Fisher sometimes makes interesting points/speculations (like Intel and memory), but his uninterrupted flow of article about how great Intel is has become boring.

At least Ahsraf Eassa has shown some varying point of views and when he's wrong he just admits it. I have yet to see Russ Fisher doing so

Of course that doesn't make him always wrong, I just have a hard time trusting people always showing the same side of the coin.

/end of rant :biggrin:

His profile states that he's retired after 35 years in the semiconductor industry; I wonder which company he worked for? Perhaps the one he is also invested in?

 

[TreVader]

I do wonder about this. The A7 in the iPhone 5s only scores 13,869 on 3Dmark Ice Storm compared to the 16,638 for a Baytrail based Celeron which Sweepr linked. Now on the A7 Apple is already using Imagination's PowerVR G6430, so unless Apple switches their graphics IP provider on the A8 they're limited to Imagination's PowerVR GX6650 at best. If Imagination's 50% performance improvement claim for Series 6XT over Series 6 is to believed and we assume perfect scaling that'd still only be a 2.25x performance improvement in graphics... Which isn't so great when you figure that Intel's both moving from IVB-based graphics to BDW-based and quadrupling the number of EUs... and they are already ahead anyway.

Really I suspect that the above is the reason why Apple started putting together their own GPU design team - Imagination isn't scaling up their designs as fast as Apple wants.

http://anandtech.com/bench/product/1065?vs=1053

I see apples A7 scoring 15,000 to Bay Trails 12,000 in Ice Storm.


The Graphics test is the only one where Bay Trail wins, and it loses even worse in Physics. I think Apple will use Power VR for their next chip, and it's possible they might even lose overall in Ice Storm to Cherry Trail, but not by much and again I would depend on Intel to disappoint and Apple to impress in any new design, especially graphics considering Intel's history there.

 

[TreVader]

I've got another article for you, then: Intel: The End Of Moore's Law. And a picture:

You can also look at the hard facts: Intel will start volume production of 10nm in Q4 2015, less than 2 years after 14nm. TSMC will start shipping 20nm FinET at the end of 2015 or 2016. Then it will take another 2 years before TSMC has 10nm, which is the same as what Intel calls 14nm, so let's say in H2 2017.

Intel 14nm will be available in H2 2013, so TSMC will be 3 years behind. he only good thing is that FinFET will only be 3.5 years behind (2012->2016) instead of 5.

Edit: Another nice picture:

Intel 14nm will be available H2 2013? It's already Q2 2014!

You'll be lucky to see 14nm before Q2 2015 and by then TSMC will have been on 20nm for 6-9 months. So they're basically a year behind. If 14nm performs only 10% better than 20nm, then intel will have to actually improve their architecture which is something they hate to do and won't unless they start to lose revenue.

 

[Homeles]

Intel 14nm will be available H2 2013? It's already Q2 2014!

You'll be lucky to see 14nm before Q2 2015 and by then TSMC will have been on 20nm for 6-9 months. So they're basically a year behind. If 14nm performs only 10% better than 20nm, then intel will have to actually improve their architecture which is something they hate to do and won't unless they start to lose revenue.

It was obviously a typo.

Q2 2015? You're the forum's new SiliconWars.

 

[Nothingness]

His profile states that he's retired after 35 years in the semiconductor industry; I wonder which company he worked for? Perhaps the one he is also invested in?

He worked for Xilinx, but not as an engineer, IIRC. But working/having worked for a company does not imply you like everything that company does. I know people working at Intel who criticize some of Intel decisions

 

[North01]

http://anandtech.com/bench/product/1065?vs=1053

I see apples A7 scoring 15,000 to Bay Trails 12,000 in Ice Storm.


The Graphics test is the only one where Bay Trail wins, and it loses even worse in Physics. I think Apple will use Power VR for their next chip, and it's possible they might even lose overall in Ice Storm to Cherry Trail, but not by much and again I would depend on Intel to disappoint and Apple to impress in any new design, especially graphics considering Intel's history there.

3DMark Ice Storm Unlimited from Futuremark's top list (April) - Source

Dell Venue 11 Pro (Z3770) - Source

Average score: 15,545
Graphic score: 14,948
Physics score: 20,075


iPad Air (A7) - Source

Average score: 15,086
Graphic score: 19,123
Physics score: 8,652

It is also worth noting that Ice Storm is based on OpenGL ES 2.0 and Direct3D feature level 9.

 

[Arachnotronic]

3DMark Ice Storm Unlimited from Futuremark's top list (April) - Source

Dell Venue 11 Pro (Z3770) - Source

Average score: 15,545
Graphic score: 14,948
Physics score: 20,075


iPad Air (A7) - Source

Average score: 15,086
Graphic score: 19,123
Physics score: 8,652

It is also worth noting that Ice Storm is based on OpenGL ES 2.0 and Direct3D feature level 9.

It's pretty clear that Gen 7 is a pretty lousy architecture that offered poor performance/watt and performance/mm^2 normalized for process, but this is to be expected given that it first launched in the 2012 Ivy Bridge. Gen 8 LP should be a strong architecture, however, and I expect that internally, Intel realizes the importance of the GPU.

Gen 9, which shows up in Broxton during mid-2015, should be even better. It will be interesting to see how they do.

 

[TreVader]

3DMark Ice Storm Unlimited from Futuremark's top list (April) - Source

Dell Venue 11 Pro (Z3770) - Source

Average score: 15,545
Graphic score: 14,948
Physics score: 20,075


iPad Air (A7) - Source

Average score: 15,086
Graphic score: 19,123
Physics score: 8,652

It is also worth noting that Ice Storm is based on OpenGL ES 2.0 and Direct3D feature level 9.

I looked at the AT bench that I linked to in my last post and I had mixed up physics and graphics. However, I'm not sure if the T100 is clocked lower (Bay Trail) or what because it loses by 3000 points vs the iPad Air. So I guess in some form factor bay trail is a competent SoC.


I would like to see a top shelf device running an Atom SoC but I don't see it happening at the moment. And I think intel is making ground with Bay Trail but apple is also improving on their designs and I have a lot of confidence in Apple.

 

[witeken]

However, I'm not sure if the T100 is clocked lower (Bay Trail) or what because it loses by 3000 points vs the iPad Air. So I guess in some form factor bay trail is a competent SoC.

T100 uses a Z3740, with a clock speed of 1.86GHz.

I would like to see a top shelf device running an Atom SoC but I don't see it happening at the moment. And I think intel is making ground with Bay Trail but apple is also improving on their designs and I have a lot of confidence in Apple.

I've seen some rumors that the Nexus 8 will have Bay Trail.

 

[joshhedge]

I am curious about the possibility of Apple using a big.LITTLE combination between a 64 bit variant of the Swift Cores and enchanted Cyclone Cores, what is the general consensus on that? Swift is in no means slow, but significantly less power-hungry than Cyclone, but would this work in practice?

 

[NTMBK]

I am curious about the possibility of Apple using a big.LITTLE combination between a 64 bit variant of the Swift Cores and enchanted Cyclone Cores, what is the general consensus on that? Swift is in no means slow, but significantly less power-hungry than Cyclone, but would this work in practice?

Better power management for a big core is preferable, to be honest. big.LITTLE has penalties related to migrating workloads between core clusters, as well as the obvious size penalty. It's not like the iPhone 5S has battery life problems on light workloads.