Intel Q313 Results

[krumme]

What did you expect, Krumme? They are changing Atom from a backburner project to a bleeding edge project. They have to sink a lot of money there, and we already know that the old Atom doesn't cut it. How could they *not* lose money on this transition?

If they aren't at least breaking even by Q214, then I would start to be worried, but until then, there isn't much Intel can do, is there?

No they can not do much. I would even give it until q4 14. But man thats a short time when you are in the middle of it.

 

[CHADBOGA]

And yet you don't realize that Apple will never ever use an Intel chip for an iPhone/iPad/iPod. That leaves out of their hands a great chunk of the mobile market.

People said the same thing about Apple ever using Intel chips in their Mac's.

It is not hard to imagine a scenario within 5 to 7 years which sees Apple switch to Intel, just as it is not hard to imagine a scenario where Apple stays with ARM for the next 20+ years.

 

[videogames101]

There is actually another (correct) reason that will become very apparent on Thursday. AMD will outperform Intel in PC's. No hints other than that, sorry.

It's Thursday.

 

[Nothingness]

I know Infineon uses ARM cores too and was acquired by Intel, but they use licensed designs, not custom CPUs. No reason for them to have ever had an architectural license.

You are wrong
http://www.infineon.com/cms/en/corporate/press/news/releases/2009/INFCCS200911-009.html
Can't say if that was part of what Intel bought.

 

[R0H1T]

People said the same thing about Apple ever using Intel chips in their Mac's.

The move to Intel was because x86 was overall better in terms of performance but the bigger reason was bootcamp so you can't possibly say that windows wasn't the big part of their reason for this transition.

It is not hard to imagine a scenario within 5 to 7 years which sees Apple switch to Intel, just as it is not hard to imagine a scenario where Apple stays with ARM for the next 20+ years.

If & only if the performance gains offset Apple's loss(in terms of profit margin) & that Intel chips give them more of a market share btw which is rapidly declining right now.

 

[liahos1]

You are wrong
http://www.infineon.com/cms/en/corporate/press/news/releases/2009/INFCCS200911-009.html
Can't say if that was part of what Intel bought.

This is for the secure element business (NFC, Smartcards, Govt ID, Passports, etc) not infineon wireless (what intel bought).

 

[Nothingness]

The one thing that you've ignored is that Cyclone is a relatively large CPU. Have you looked at the die photos for each core? If they're both ~100mm^2 die, then the CPU core for A7 looks like it's 2-3 times larger than the CPU core for Silvermont. If these two CPUs have anywhere close to the same performance, then it certainly does look like Intel's CPU is much more efficient. That x86 baggage isn't really much of an issue, maybe?

Or to say it another way, it looks like Bay Trail can fit an extra two cores in the extra space and provide similar single threaded performance.

Right but don't forget one chip is 22nm, while the other is 28nm.

Dual core 28nm Cyclone is 22mm^2: http://forums.anandtech.com/showpost.php?p=35516740&postcount=171
Dual core 22nm Silvermont is 8.x mm^2: http://pc.watch.impress.co.jp/img/pcw/docs/615/261/html/14.jpg.html

 

[liahos1]

from everyone i've spoken to the concept of x86 baggage as it relates to mobile appears to be weak esp at 100mm^2 die sizes. i guess as you get down to to quark level of product it becomes more of a hindrance. true or false?

 

[Nothingness]

from everyone i've spoken to the concept of x86 baggage as it relates to mobile appears to be weak esp at 100mm^2 die sizes. i guess as you get down to to quark level of product it becomes more of a hindrance. true or false?

That's what I believe too. Note that it's not really a question of area (which depends on process and on transistor density) but rather a question of micro-architecture and targeted performance: as soon as you try to extract more performance, the decoders and microcode ROM are getting less and less significant compared to other parts of your design (large register files for renaming, large TLB RAM, large branch predictors, etc.). But as soon as you're trying to use less transistors for your CPU then the decoders (and microcode ROM for complex x86 instructions) start taking a proportionally larger space.

 

[Imouto]

People said the same thing about Apple ever using Intel chips in their Mac's.

It is not hard to imagine a scenario within 5 to 7 years which sees Apple switch to Intel, just as it is not hard to imagine a scenario where Apple stays with ARM for the next 20+ years.

Yeah, because changing suppliers is the same as giving up on your own custom SoC and cores development where you have total control over them and throw away the work done in the last 8 years. And when you're there, using Intel chips for your mobile gadgets Intel could pull the ultrabook move again.

Without Intel, without X86 and in a way behind process they pulled an A7 versus Bay Trail that was supposed to "Conroe" the market.

I totally see Apple moving to Intel.

 

[Phynaz]

It's Thursday.

AMD will announce results after the market close.

I'm betting that what SW is all in a lather about is AMD may report growth in desktop CPU sales. Percentage wise it could be be a pretty big number, due to their starting from such a small number.
(i.e. AMD may only have to increase sales by 100,000 chips to see 10% growth, whereas Intel may need to increase sales by 8,000,000 to see 10% growth).

 

[mrmt]

I'm betting that what SW is all in a lather about is AMD may report growth in desktop CPU sales. Percentage wise it could be be a pretty big number, due to their starting from such a small number.
(i.e. AMD may only have to increase sales by 100,000 chips to see 10% growth, whereas Intel may need to increase sales by 8,000,000 to see 10% growth).

Actually they have to increase betweene 500.000-600.000 to see that 10% growth.

 

[Phynaz]

Actually they have to increase betweene 500.000-600.000 to see that 10% growth.

Yeah, I was just throwing numbers out there for an example.

 

[SlimFan]

Right but don't forget one chip is 22nm, while the other is 28nm.

Dual core 28nm Cyclone is 22mm^2: http://forums.anandtech.com/showpost.php?p=35516740&postcount=171
Dual core 22nm Silvermont is 8.x mm^2: http://pc.watch.impress.co.jp/img/pcw/docs/615/261/html/14.jpg.html

Yes, that's the process delta. And I don't think there's any evidence that says Intel's 22nm is 2-3x the density of a 28nm process.

 

[Exophase]

You are wrong
http://www.infineon.com/cms/en/corporate/press/news/releases/2009/INFCCS200911-009.html
Can't say if that was part of what Intel bought.

Alright, an ARMv6m and ARMv7m license. I'm hesitant to even mention that in this conversation. That's a world apart from the ARMv7a and v8a ISAs we're talking about here. Especially if they didn't go for the architecture extensions used by Cortex-M4 and from this alone I have no reason to believe they did. But you're right - I thought they were just using ARM11s or Cortex-A5s or something, not custom microcontroller-class CPUs.

I actually had no idea anyone even licensed these over just using Cortex-M cores.

The one thing that you've ignored is that Cyclone is a relatively large CPU. Have you looked at the die photos for each core? If they're both ~100mm^2 die, then the CPU core for A7 looks like it's 2-3 times larger than the CPU core for Silvermont. If these two CPUs have anywhere close to the same performance, then it certainly does look like Intel's CPU is much more efficient. That x86 baggage isn't really much of an issue, maybe?

Or to say it another way, it looks like Bay Trail can fit an extra two cores in the extra space and provide similar single threaded performance.

How can you mention efficiency in a discussion while only mentioning area and performance without mentioning power consumption? We don't know a lot about Apple's design but we do know it has 64KB+64KB L1 cache vs 32KB+24KB for BayTrail. Of course that's going to take more transistors. You make a move like that when you want to increase perf/MHz vs absolute MHz, which looks to be exactly what Apple's gunning for here. That's a clear cut perf/W move. I bet Cyclone made a lot of other decisions like this.

And I think they're making the right decisions. If CPU perf/W is as critical to users as everyone in the industry is making it out to be then it seems reasonable to make the whole SoC 5-10% larger. Maybe this isn't true if you're ARM and are making a core that you want to sell for use in very cheap applications, like quad A9s in a 25mm^2 Rockchip SoC, but in Apple's case big SoCs are their bread and butter. The other side of that is their adherence to dual core, where the overall impact from this core bloat is smaller.

 

[SlimFan]

How can you mention efficiency in a discussion while only mentioning area and performance without mentioning power consumption? We don't know a lot about Apple's design but we do know it has 64KB+64KB L1 cache vs 32KB+24KB for BayTrail. Of course that's going to take more transistors. You make a move like that when you want to increase perf/MHz vs absolute MHz, which looks to be exactly what Apple's gunning for here. That's a clear cut perf/W move. I bet Cyclone made a lot of other decisions like this.

Yes, Cyclone has much bigger L1 caches, that's definitely true. That's why I didn't talk transistors, but rather area. That is likely one of the things that can improve perf/MHz. But your assumption that this is a fundamental perf/W improvement isn't something I necessarily agree with.

And I think they're making the right decisions. If CPU perf/W is as critical to users as everyone in the industry is making it out to be then it seems reasonable to make the whole SoC 5-10% larger. Maybe this isn't true if you're ARM and are making a core that you want to sell for use in very cheap applications, like quad A9s in a 25mm^2 Rockchip SoC, but in Apple's case big SoCs are their bread and butter. The other side of that is their adherence to dual core, where the overall impact from this core bloat is smaller.

I didn't mention perf/W because nobody has any comparable power data between the platforms that I've seen. Once people have products in hand and do power measurement teardowns (and let's hope that happens at some point), I think it'll be really interesting.

This is classic speed-demon vs braniac. It's really interesting that Intel, the company with the highest IPC core available, ends up being the speed demon in mobile.

If it turns out that a 2.4GHz Silvermont is the same power and performance as a 1.3GHz Cyclone (arguable), and is 1/3 the area (different processes, definitely) ... then which core would you say is the "better" core? Obviously this is a technical argument, and not one that matters in the end product. It's still fun to consider, though.

The quad core phone thing is pretty stupid, so I don't really care about the fact that I can fit twice as many in the same are. That's definitely an Apple advantage. But then again, I don't really care about die size at all. That's a problem for the people who purchase the SOCs itself to care about.

 

[Exophase]

Yes, Cyclone has much bigger L1 caches, that's definitely true. That's why I didn't talk transistors, but rather area. That is likely one of the things that can improve perf/MHz. But your assumption that this is a fundamental perf/W improvement isn't something I necessarily agree with.

I doubt Apple did anything to improve perf/MHz that wasn't also a perf/W win. Otherwise they would have probably designed for higher MHz. Power consumption scales badly with frequency so you can optimize a lot in perf/MHz while still being a win in perf/W. That's why Haswell added a pretty large amount of extra resources for a pretty modest typical performance improvement. They're probably a lot closer to scraping the best IPC vs clock speed product they'll get than Apple is.

I didn't mention perf/W because nobody has any comparable power data between the platforms that I've seen. Once people have products in hand and do power measurement teardowns (and let's hope that happens at some point), I think it'll be really interesting.

You're right, the data's not out. So you shouldn't have said anything, not just dropped the most important metric from the comparison entirely

If it turns out that a 2.4GHz Silvermont is the same power and performance as a 1.3GHz Cyclone (arguable), and is 1/3 the area (different processes, definitely) ... then which core would you say is the "better" core?

Area is 18mm^2 vs 8mm^2, I trust Hans de Vries' numbers >_>

Given Intel's substantial manufacturing advantage it'd be pretty bad if they didn't have a better core. I'm more interested in whether or not Intel would have had the better core had they designed it for the same process. Something that's impossible to really tell.

 

[Arachnotronic]

It's really interesting that Intel, the company with the highest IPC core available, ends up being the speed demon in mobile.

If you've got a process advantage - use it to save money. Why go wide when you can stay narrow, clock higher than everyone else at the same power, and get better perf/watt? Smaller core, saves $.

 

[SlimFan]

You're right, the data's not out. So you shouldn't have said anything, not just dropped the most important metric from the comparison entirely

Well... it's likely that an M3 core has the best perf/W when focusing on just the CPU power... so I think it's okay to say *something* :sneaky:

Area is 18mm^2 vs 8mm^2, I trust Hans de Vries' numbers >_>

Given Intel's substantial manufacturing advantage it'd be pretty bad if they didn't have a better core. I'm more interested in whether or not Intel would have had the better core had they designed it for the same process. Something that's impossible to really tell.

I believe those numbers are core + L2, and both designs have the same L2 area. Did Hans say how the L2 cache sizes compare? That might give a hint as to the process comparisons, at least for SRAMs. I think that if you subtract the L2 areas, you're in the 2-3x range. That's pretty substantial, and I don't think process alone explains it.

It just seems like some people are excited about Cyclone's impressive IPC without analyzing at the entire picture (power, area, performance).

 

[CHADBOGA]

Yeah, because changing suppliers is the same as giving up on your own custom SoC and cores development where you have total control over them and throw away the work done in the last 8 years. And when you're there, using Intel chips for your mobile gadgets Intel could pull the ultrabook move again.

Without Intel, without X86 and in a way behind process they pulled an A7 versus Bay Trail that was supposed to "Conroe" the market.

I totally see Apple moving to Intel.

You need to look at the bigger picture and use two eyes instead of one. :awe:

 

[Nothingness]

If you've got a process advantage - use it to save money. Why go wide when you can stay narrow, clock higher than everyone else at the same power, and get better perf/watt? Smaller core, saves $.

Agreed, but as some point you have to change your strategy or you end up with a remake of Pentium 4.