Intel may dismiss tick-tock after Haswell, no performance CPUs (LGA based) anymore?

[pablo87]

So what has the cost of 22nm been so far, $24B? At least that was the capex total for 2011, 2012 a two fold increase from historical. What's worse, they're already taking tools out of 32nm factories for 14nm -

It doesn't seem like it was worth the performance improvement to go to 22nm, putting IB on 32nm might have worked too.

In any event, Intel is spending too much capex money, they need to change the approach because the status quo doesnt' work bottom line wise. They might need to start producing more parts on trailing fabs in order to get their capex down.

Keep in mind, integrated graphics was on a separate chip in the past, on older,fully depreciated fabs. now its on the leading edge fab which costs a lot more yet Intel has not been able to get any more money for the parts.

Just to give you an idea, continuously spending 12B capex means eventually depreciation will be 12B as well which means, depreciation alone is $40 per processor

What do those smartphone and tablet parts sell for again?

 

[jpiniero]

And to make sure people continue to upgrade and buy new PCs, Intel will have to give them a compelling reason. So they cannot stall on performance, either. I expect decent progress on performance, and great leaps in performance per watt from Intel.

But that's precisely the problem: 80% of their customers don't need faster cpus. There are much better ways to improve the experience - ie SSDs, better displays, etc. Microsoft isn't helping them with the disaster that is Windows 8 either.

And tick-tock continues. Intels server roadmap also confirms it. (Note server versions are behind desktop/mobile.)

Of course roadmaps can change. Intel's probably spending most of their effort now reducing power consumption and/or cost, mainly to deal with the ARM threat.

 

[ShintaiDK]

Just to give you an idea, continuously spending 12B capex means eventually depreciation will be 12B as well which means, depreciation alone is $40 per processor

What do those smartphone and tablet parts sell for again?

And ASP is 121$ in average.

 

[mrmt]

So what has the cost of 22nm been so far, $24B? At least that was the capex total for 2011, 2012 a two fold increase from historical. What's worse, they're already taking tools out of 32nm factories for 14nm

Pablo, we've been through this before. You are taking Capex with non-recurring items such as the McAfee acquisition and the brand new Oregon factory and assuming that Intel would need exactly this amount of Capex to keep business running. It won't.

Keep in mind, integrated graphics was on a separate chip in the past, on older,fully depreciated fabs. now its on the leading edge fab which costs a lot more yet Intel has not been able to get any more money for the parts.

This isn't a problem. IGP does affect layout and validation costs, and those aren't small by any means, but regarding manufacturing costs, it isn't a deal breaker. In the end Sandy Bridge is smaller than its Conroe or even some Netburst counterparts, and this relation gets even better with IVB, as the 4C is almost the same size of P4 Northwood.

More important your suggestion would imply moving the IGP from the same die to the same package, a backward step in SoC integration.

 

[naonao]

This article says:

• Intel will not provide new products for Desktop and non-BGA laptop segments in Broadwell era
• They will provide higher clocked Haswell instead for those segments in 2014
• Broadwell is "more than tick", and it will include some technologies that were previously planned for Skylake
• This will because Intel needs to be more competitive in tablet market, and mean the end of Tick-Tock strategy

It mentions nothing about Skylake and later.

 

[boxleitnerb]

That makes more sense than my interpretation. I hate googlish

 

[cbn]

What do those smartphone and tablet parts sell for again?

Lets see how they do with Silvermont......and especially Airmont (I believe that will integrate Wifi on the SOC.)

Keep in mind, integrated graphics was on a separate chip in the past, on older,fully depreciated fabs. now its on the leading edge fab which costs a lot more yet Intel has not been able to get any more money for the parts.

How do you feel about integrating radios on leading edge fabs?

 

[pablo87]

Lets see how they do with Silvermont......and especially Airmont (I believe that will integrate Wifi on the SOC.)



How do you feel about integrating radios on leading edge fabs?

It might make a lot of sense.

It has to do with balancing old fabs and new fabs, it seems they could easily build the vast majority of PC processors on old fabs but they'll still need to strike a balance between old and new so its quite possible that if they went that route (of building say 80% of processors on old fabs), they wouldn't have enough product to build on the new fabs thus Airmont. It also leverages their fab prowess and genius to get into tablets and smartphones.

But they'll also need to raise ASPs on PCs and Servers; one way they could do so is by coming out with products that take the die space occupied by integrated graphics (which costs a lot but doesn't raise ASP) with more "CPU' features (more cores, more cache, why) for which there is probably demand for. Perhaps integrated on die gets primarily relegated to older fabs eventually (unfortunately, not the case with Haswell though we don't know the mix they plan to build; if its another "fastest ramp" in history then you'll know they made the wrong decision).

One thing either they don't realize or its not POR is that while the high end processors subsidizes the value ones, the old fabs subsidize the new ones and taking tools out of the old fabs because they don't have enough product to build there is plain dumb. The old fabs have always been a strategic cost advantage for Intel, just as much as having the most technologically advanced fabs gives them a huge advantage. Why? Because TSMC milks their old fabs for as much money as the market will bear.

In a nutshell:

leading edge fabs = high end PC processors and Server + tablet/smartphone with integrated radio

trailing fabs = mainstream and value PC processors

The key in all this is to maintain ASP at all cost which is going to be very difficult given that parts for tablets and smartphone sell for such low prices and are in such high volume (mix). Add higher priced parts, do whatever it takes to stop the PC bleeding (ultrabook, Haswell and Windows 8 ain't going to cut it) and follow the usual better they cry than we cry strategy (right now that seems to be the motherboard - its pretty high cost for the value its adding - mine cost $129, I don't get it).

Its going to take one heck of a new CEO to navigate through this; I guess he'll be getting a head start.

One thing for sure: the traditional TICK TOCK model should be irrevocably broken as its too simplistic for what lies ahead for Intel.

PS - crossing this chasm is going to be helped of course by AMD's exit from performance x86.

 

[Nemesis 1]

For Me Haswell will be my last water cooled pc . I am really thinking of using it graphics IGP and buying the second geneeration PHi . Maybe the 1st,


Introducing OpenCL* 1.2 for Intel® Xeon Phi™ coprocessor






Submitted by Arnon Peleg (Intel) on Mon, 11/12/2012 - 16:07





Intel is extending open standards support to include OpenCL* 1.2 for Intel® Xeon Phi™ coprocessors.

Today, at Supercomputing 2012, we announced the future availability of OpenCL* 1.2 support for the Intel® Xeon Phi™ coprocessor. OpenCL broadens the parallel programming options from Intel and allows developers to maximize parallel application performance on Intel Xeon Phi coprocessors. This support helps preserve the investment OpenCL developers have made on Intel® Xeon® processors, Intel® Core™ Processors, and Intel® HD Graphics. OpenCL code developed for these processors will run on Intel Xeon Phi coprocessor with minimum changes.





On April 2012 Intel introduced the Intel SDK for OpenCL Applications 2012 supporting OpenCL 1.1 on 3rd Generation Intel® Core™ Processors and Intel® Xeon® processors. On August 2012 Intel announced OpenCL 1.2 support for Intel CPUs in the form of a Beta. Today we are extending our OpenCL support to cover all our compute devices: CPUs, Intel HD Graphics, and the just announced Intel Xeon Phi coprocessor. That OpenCL code can now run on all devices available opens up endless programming possibilities for HPC and parallel programming in general

The new Beta release of the Intel® SDK for OpenCL Applications XE 2013 lets OpenCL developer in the High Performance Compute domain explore and develop OpenCL applications on the Intel® Xeon Phi™ coprocessor. Developing today with the Beta product will help programmers to preserve the investment they made on Intel® Xeon® processors, Intel® Core™ Processors, and Intel® HD Graphics. OpenCL code developed for these processors will run on Intel Xeon Phi coprocessor with minimum changes.

 

[cbn]

It might make a lot of sense.

It has to do with balancing old fabs and new fabs, it seems they could easily build the vast majority of PC processors on old fabs but they'll still need to strike a balance between old and new so its quite possible that if they went that route (of building say 80% of processors on old fabs), they wouldn't have enough product to build on the new fabs thus Airmont. It also leverages their fab prowess and genius to get into tablets and smartphones.

But they'll also need to raise ASPs on PCs and Servers; one way they could do so is by coming out with products that take the die space occupied by integrated graphics (which costs a lot but doesn't raise ASP) with more "CPU' features (more cores, more cache, why) for which there is probably demand for. Perhaps integrated on die gets primarily relegated to older fabs eventually (unfortunately, not the case with Haswell though we don't know the mix they plan to build; if its another "fastest ramp" in history then you'll know they made the wrong decision).

One thing either they don't realize or its not POR is that while the high end processors subsidizes the value ones, the old fabs subsidize the new ones and taking tools out of the old fabs because they don't have enough product to build there is plain dumb. The old fabs have always been a strategic cost advantage for Intel, just as much as having the most technologically advanced fabs gives them a huge advantage. Why? Because TSMC milks their old fabs for as much money as the market will bear.

In a nutshell:

leading edge fabs = high end PC processors and Server + tablet/smartphone with integrated radio

trailing fabs = mainstream and value PC processors

The key in all this is to maintain ASP at all cost which is going to be very difficult given that parts for tablets and smartphone sell for such low prices and are in such high volume (mix). Add higher priced parts, do whatever it takes to stop the PC bleeding (ultrabook, Haswell and Windows 8 ain't going to cut it) and follow the usual better they cry than we cry strategy (right now that seems to be the motherboard - its pretty high cost for the value its adding - mine cost $129, I don't get it).

Its going to take one heck of a new CEO to navigate through this; I guess he'll be getting a head start.

One thing for sure: the traditional TICK TOCK model should be irrevocably broken as its too simplistic for what lies ahead for Intel.

PS - crossing this chasm is going to be helped of course by AMD's exit from performance x86.

Pablo,

Instead of Intel selling value processors on old fabs.......I wonder how much price premium Intel could charge fabless companies for use of its "old fabs"?

For example in middle of 2016, Intel should be on 10nm (with a new xtor design that is not FinFET...see interview with Mark Bohr here). That should leave some of its 14nm node (2nd generation FinFET) open to compete with TSMC's first gneration FinFET on 16nm (which I believe comes end of 2015).

P.S. Things should get even more interesting when Intel's 10nm becomes an "old fab". I have read that TSMC and others still plan to use the FinFET for that node...while Intel would be able to offer a more advanced xtor design for customer's using its old fabs (and maybe even an advantage in lithography too.)

 

[HURRIC4NE]

i think we should just dump all the processors and go outside and read a book on silicon and then come back and restart the whole processor business.... hey intel, focus.

 

[pablo87]

Pablo,

Instead of Intel selling value processors on old fabs.......I wonder how much price premium Intel could charge fabless companies for use of its "old fabs"?

For example in middle of 2016, Intel should be on 10nm (with a new xtor design that is not FinFET). That should leave some of its 14nm node (2nd generation FinFET) open to compete with TSMC's first gneration FinFET on 16nm (which I believe comes end of 2015).

P.S. Things should get even more interesting when Intel's 10nm becomes an "old fab". I have read that TSMC and others still plan to use the FinFET for that node...while Intel would be able to offer a more advanced xtor design for customer's using its old fabs (and maybe even an advantage in lithography too.)

That's another good strategy, even more so if Intel invested in said company..

However, x86 is so ingrained inside Intel, that many of their businesses went nowhere because of it eg. StrongARM/Xscale.

I believe they had agreed to fab for some companies already (fpga house?), it would be interesting to know how well that is going from customer's perspective...

It is very important for Intel otherwise current capex and eventual depreciation/ factory charge is not compatible with current ASP's/profit levels.

 

[cbn]

That's another good strategy, even more so if Intel invested in said company..

However, x86 is so ingrained inside Intel, that many of their businesses went nowhere because of it eg. StrongARM/Xscale.

Yes, the alliances formed would be interesting. (Competitors cooperating with each other? Well, it is happening with Samsung and Apple....but maybe that is bad example. )

So who could be the possible clients for a second tier Intel fab? (Assuming Intel's second tier is better than a first tier fab from TSMC?) Anyone want to venture a guess?

 

[mrmt]

That's another good strategy, even more so if Intel invested in said company..

However, x86 is so ingrained inside Intel, that many of their businesses went nowhere because of it eg. StrongARM/Xscale.

I don't think it's just a matter of x86 only.

You don't change your foundry partner each node, you usually go for partnerships that stick for years. And here we have a problem because Intel isn't with structurally low fab utilization rates, so in order to get the big fabless guys they would need to develop more capacity, which would need a lot more capex, which is exactly your main worry about Intel, and that for have return rates well below their current rates.

Once you see x86 ROI falling below foundry rates, Intel will open even its bleeding edge node to fabless companies, until then, forget it. They will be better putting their money on x86.

 

[Ajay]

And ASP is 121$ in average.

Not for the SoCs, IIRC. That figure looks like BOM costs.

 

[ShintaiDK]

Not for the SoCs, IIRC. That figure looks like BOM costs.

And your math on that? Even discount AMD got a 48$ average ASP.

Clovertrail is also half the size of competing SoCs.

 

[Ajay]

And your math on that? Even discount AMD got a 48$ average ASP.

Clovertrail is also half the size of competing SoCs.

I've generally heard under $40 for the SoC (we are talking tablets and smart phones). Here is an example. I seem to recall that the QCOM Snapdragon cost around the same as the Tegra3.

 

[pablo87]

I've generally heard under $40 for the SoC (we are talking tablets and smart phones). Here is an example. I seem to recall that the QCOM Snapdragon cost around the same as the Tegra3.

Don't take what clever CEO's say at face value. if anything, it costs less. That thing is made on old process, NVDA is most likely #1 customer status at TSMC with special relationship, NV's gross margins are sky high, what's wrong with the picture? JHH doesn't want his customers to know his cost, they'll use that info. to grind him down, and he doesn't want to boast either, so he says it's more...

Now if it was RR, he'd say their cost is very low, see the difference?

 

[PlasmaBomb]

Not for the SoCs, IIRC. That figure looks like BOM costs.

$121 is Intels average selling price for CPUs.

 

[Idontcare]

Don't take what clever CEO's say at face value. if anything, it costs less. That thing is made on old process, NVDA is most likely #1 customer status at TSMC with special relationship, NV's gross margins are sky high, what's wrong with the picture? JHH doesn't want his customers to know his cost, they'll use that info. to grind him down, and he doesn't want to boast either, so he says it's more...

Now if it was RR, he'd say their cost is very low, see the difference?

Yep, CEO's have to speak out of both sides of their mouths.

In front of customers and employees they have to plead poverty lest the customers decide the CEO can afford to cut the price some or employees decide they deserve a larger cut of the profits in terms of bonuses and salary bumps.

In front of the BoD and shareholders they have to tout their financial accomplishments like a pimp who is paying his ho's pennies on the dollar for their back-breaking labor while charging the Johns absurd prices for something that ordinarily would cost no more than a drink or two at a bar in college

And if you study the financials of a CEO's compensation package, the pimp usually wins in this situation

 

[pablo87]

Yes, the alliances formed would be interesting. (Competitors cooperating with each other? Well, it is happening with Samsung and Apple....but maybe that is bad example. )

So who could be the possible clients for a second tier Intel fab? (Assuming Intel's second tier is better than a first tier fab from TSMC?) Anyone want to venture a guess?

Don't think too many 1st tier clients would like to be second class citizens...

I think if Intel acquired Micron and managed to integrate memory manufacturing into their fab processes, that would be a big win because of the sheer volume involved. Also, when you look at smartphones, tablets, Chromebook's and value PC's, memory of all kinds is a bigger slice of the BOM than the processor/SOC. There also seems to be some new memory innovations on the horizon that could benefit wag from leading edge processes.

Now Intel had a bad experience in memory but times are a changing - it's basically down to Micron vs Korea Inc. at least in DRAM. Korea Inc. also leads in smartphones...

I still think x86 is the most logical candidate for older fabs. The lead is so large, despite wasting IMO a lot of die real estate on IGP.

 

[tommo123]

^ good point batman! they just need a way to keep the foundrys as busy as possible.

side note - would love it if AMD got intel to do their GPUs for them

 

[Ajay]

$121 is Intels average selling price for CPUs.

So we weren't talking about ARM SoCs. O.K.

 

[IntelUser2000]

Don't think too many 1st tier clients would like to be second class citizens...

And in a theoretical world where Intel's lead is so large their last generation node is better than the cutting edge of competitors that matters how....? It'll be that or using something worse. The most baseless statement I've heard in a while.

But I doubt they want to do that. Maybe if they can do that using last-last process tech. Like imagine if everyone else is stuck at 28nm where Intel is at 10nm. Then Intel can give them their 22nm. The reason is because their own products use previous gen processes too, like server chips.

I still think x86 is the most logical candidate for older fabs. The lead is so large, despite wasting IMO a lot of die real estate on IGP.

The topic of "Waste" is different depending on the individual. You can say the same about AMD's chips and even Apple's SoC's.

Once you see x86 ROI falling below foundry rates, Intel will open even its bleeding edge node to fabless companies, until then, forget it.

I think its almost certain if Intel starts fabbing for others majority of their advantage will go away. Their biggest strength and the reason they have such a lead is because they are such a well executed IDM. That's why I believe if something happens that breaks that trend, you'll wonder where their advantage has went. Something like the company splitting up or even fabbing for others.

There's always a human aspect to business that can't be ignored. Two different companies mean two different goals, and conflict of interest can occur(think of Global Foundries and AMD). That's a hindrance being an IDM does not have, because everything is controlled under one roof.

 

[cbn]

I think if Intel acquired Micron and managed to integrate memory manufacturing into their fab processes, that would be a big win because of the sheer volume involved.

Pablo,

According to the following Anandtech article here Intel uses a process for CPUs that is different than DRAM.

Intel and Micron came up with an idea. Take a DRAM stack and mate it with a logic process (think CPU process, not DRAM fabs) layer for buffering and routing and you can deliver a very high bandwidth, low power DRAM.