EETimes: ST plans for Dresden FDSOI production

[mrmt]

**Based on what I've read lately, AMD really needs to become competitive in Big Cores again, the market for small core servers is becoming red hot, with many players with their own high speed switching networks (Meshes).

AMD is pretty much behind Intel in big core. They need a huge amount of R&D just to level the playfield in design, and GLF needs to do a lot of work for that and even with flawless execution they would have problems catching up. Catching up with fewer and spreader resources will be quite a challenge.

As for the small core server market, it was meant to get crowded long ago, the writing was always there. Samsung, Qualcomm, Calxeda, Nvidia and Intel were going for that market. How such a market wouldn't be crowded? What margins do you think AMD will get for a vanilla ARM core on this market?

And here the current executive team have only themselves to blame, the microserver/ARM focus was Rory's, Su's and Papermaster's baby.

 

[Ajay]

AMD is pretty much behind Intel in big core. They need a huge amount of R&D just to level the playfield in design, and GLF needs to do a lot of work for that and even with flawless execution they would have problems catching up. Catching up with fewer and spreader resources will be quite a challenge.

As for the small core server market, it was meant to get crowded long ago, the writing was always there. Samsung, Qualcomm, Calxeda, Nvidia and Intel were going for that market. How such a market wouldn't be crowded? What margins do you think AMD will get for a vanilla ARM core on this market?

And here the current executive team have only themselves to blame, the microserver/ARM focus was Rory's, Su's and Papermaster's baby.

Staying on big core for the x86 server market, where AMD had only one competitor (with a huge financial advantage) could still work out for them. SR Opterons on 28nm FD-SOI would easily give a 50% performance increase (ST). I now view AMD's move into ARM as dumb. Maybe they could get more money selling SeaMicro to Qualcomm. It is obvious to me now, that AMD's wheelhouse is x86. And they are producing some very attractive low power SoC right now - excellent execution. So AMD is capable of this.

They could deliver on big core Desktop APUs, big core Server parts and a couple of small core low power SoC for tablets and eventually even smartphones. Four sectors (plus ATI). More than enough to keep them busy and they don't need to match Intel's release schedule on Server CPUs (which is actually slowing down anyway), so long as each 2 year iterations delivers substantial performance/watt gains over the previous model. I'm thinking that AMD got it all wrong by extrapolating one blunder into culture of fatalism.

 

[Khato]

SR Opterons on 28nm FD-SOI would easily give a 50% performance increase (ST).

Where exactly do they claim that 28nm FD-SOI 'would easily give a 50% performance increase' over 32nm bulk? The only 50% number I recall seeing was the increase in energy efficiency compared to 28nm bulk when run at extremely low voltages.

Going by the Multi-core A9 dynamic power versus frequency graph comparing 28nm bulk to 28nm FD-SOI on page 27 of the more detailed presentation ST gave on FD-SOI results back in December - linked again for convenience - they saw something around a 17.5% increase in frequency (1.85 GHz to about 2.17 GHz) at the same power consumption. aka, it does realize some gains in frequency at same power consumption, and can be run quite a bit faster at the expense of more power... sure running at 2.5 GHz is a 35% increase in frequency, but it's at the cost of a 64% increase in dynamic power! No, the real gains for FD-SOI are found by running at same frequencies with lower voltage - at the same 1.85 GHz the dynamic power goes from about 1.4W to 0.9W, a ~36% reduction.

 

[Greenlepricon]

Going by the Multi-core A9 dynamic power versus frequency graph comparing 28nm bulk to 28nm FD-SOI on page 27 of the more detailed presentation ST gave on FD-SOI results back in December - linked again for convenience - they saw something around a 17.5% increase in frequency (1.85 GHz to about 2.17 GHz) at the same power consumption. aka, it does realize some gains in frequency at same power consumption, and can be run quite a bit faster at the expense of more power... sure running at 2.5 GHz is a 35% increase in frequency, but it's at the cost of a 64% increase in dynamic power! No, the real gains for FD-SOI are found by running at same frequencies with lower voltage - at the same 1.85 GHz the dynamic power goes from about 1.4W to 0.9W, a ~36% reduction.

Applying this to a 8 core (or 4 modue if that's how you look at it) cpu would probably yield the amazing results AMD needs to be a little more competitive. While I usually don't view power as that significant for a desktop, the 200-250 watts needed for bulldozer and vishera is pushing it a little. I would like to see it where most of the cores can stay at very low power until they're needed, which wouldn't be almost at all except for extreme multitasking or something very multithreaded. I expect the idle power usage would lower enough to impress some people. Honestly I don't think it will be THAT much more powerful than what they have now besides what you mentioned with the dynamic frequency. To be able to nearly double the voltage without touching the settings when you need the extra power would be really cool.

 

[mrmt]

Staying on big core for the x86 server market, where AMD had only one competitor (with a huge financial advantage) could still work out for them.

No, they can't. When AMD devoted most of their engineering resources, which were more than they were today, they were getting hopelessly behind, even if we discount Bulldozer. Now with significant diminished resources and spread thin between three core developments, I don't see AMD come back in conventional servers as more than a pipe dream.

In servers AMD situation is even worse than in desktops, because in servers there are benefits from packing more cores AND benefits for power consumption, so Intel can bring to bear a lot of high power SKUs with a lot of big cores AND a lot of SKUs with less cores but better power consumption, all this in more mature nodes, filling up all relevant price and performance price points and effectively shutting everything but niches for AMD.

There will be a time when AMD low sales figures won't justify all the investment in testing, validation and SG&A for the big core server market. And given AMD current server numbers, that time isn't too far down the road.

SR Opterons on 28nm FD-SOI would easily give a 50% performance increase (ST).

I'm really tired to see every company + dog from the SOI club touting benefits of the thing and nobody whose money depends on selling chips adopting the thing, except for AMD, and they have a reason for that.

AMD only went with SOI because IBM was willing to sell them a process, and AMD didn't have the money to develop process in-house. (for those who tend to see Intel and AMD as two giants of the same category fighting for the market, they aren't, and the differences start here).

And if the WSA is correct by ditching SOI AMD is going to save the paltry sum of 80 million dollars per year. When you have a magic dust that can give you 40% improvements in voltage for 0 additional manufacturing costs, this magic dust *is* worth 80 million dollars per year, and yet AMD refrained to pay those 80 millions.

While we all agree that AMD management isn't the most brilliant management team of the world, don't you think this decision is too stupid to pass on? If not for their intelligence, they would for their want of make money, and SOI would raise their annual bonus big time.

As for me, until I see someone whose life depends on manufacturing chips going with SOI, I'd refrain from believing on STM/GLF claims.

I now view AMD's move into ARM as dumb.

Agreed here. I don't think AMD has a better prospect with ARM than with x86, and in ARM they will face *a lot* of companies with *a lot* more financial muscles than them.

More than enough to keep them busy and they don't need to match Intel's release schedule on Server CPUs (which is actually slowing down anyway), so long as each 2 year iterations delivers substantial performance/watt gains over the previous model.

Not really. In order to succeed they must diminish the gap with Intel. There won't be any point in take 2 years to deliver substantial improvements in perf/watt and raw performance if Intel deliver more improvements and the gap grows. AMD will be lower and lower in the value ladder, and soon it won't be worth to build an AMD server.

Wanna have an idea on what I'm talking about? Look at AMD marketing. They went from talking big, things like data centers, cloud computing, HPC servers to this make huge events to announce affordable, single socket chips for... web hosting.

 

[Ajay]

Where exactly do they claim that 28nm FD-SOI 'would easily give a 50% performance increase' over 32nm bulk? The only 50% number I recall seeing was the increase in energy efficiency compared to 28nm bulk when run at extremely low voltages.

The gains would be from:

1. Architectural improvements in SR (some say up to +30% in IPC**)
2. The Shrink to 28nm
3. The addition of FD-SOI

So the 50% number is likely a bit conservative, for single thread performance. AMD could choose to hit the 50% bump in IPC, but drop the power usage as well - offering a substantial improvement in performance/watt. The would class SR Opterons as a huge boost for current AMD customers and would grab back a few % of the server market as well (I'm sure Dell, IBM, HP, etc would like an alternate source of CPUs for at least certain parts of their markets, if only to have some leverage to use when negotiating with Intel).


** probably on code profiles that s*ck in BD/PD. Less in other areas.

 

[SocketF]

AMD only went with SOI because IBM was willing to sell them a process, and AMD didn't have the money to develop process in-house. (for those who tend to see Intel and AMD as two giants of the same category fighting for the market, they aren't, and the differences start here).

And if the WSA is correct by ditching SOI AMD is going to save the paltry sum of 80 million dollars per year. When you have a magic dust that can give you 40% improvements in voltage for 0 additional manufacturing costs, this magic dust *is* worth 80 million dollars per year, and yet AMD refrained to pay those 80 millions.

But what if you can get the magic dust for free?
What AMD ditched was their special SHP process that was PD-SOI, not FD-SOI. PD-SOI is running out of steam, technically not usable any longer, that is why STM and IBM make/made research about FD-SOI/ET-SOI.

Now if you have to pay for 28nm PD-SOI but can get 28nm FD-SOI for free what would you do?

The disadvantage is loosing a few ~100 MHz clock headroom because the FD-SOI is based on the LP-process, but in return you get much lower leakage, especially at lower clocks. And if you want invest the additional wattage for higher clocks e.g. for some FX CPUs, you can do so by body biasing.

Obviously an easy decision...

 

[Ajay]

No, they can't. When AMD devoted most of their engineering resources, which were more than they were today, they were getting hopelessly behind, even if we discount Bulldozer. Now with significant diminished resources and spread thin between three core developments, I don't see AMD come back in conventional servers as more than a pipe dream.

In servers AMD situation is even worse than in desktops, because in servers there are benefits from packing more cores AND benefits for power consumption, so Intel can bring to bear a lot of high power SKUs with a lot of big cores AND a lot of SKUs with less cores but better power consumption, all this in more mature nodes, filling up all relevant price and performance price points and effectively shutting everything but niches for AMD.

There will be a time when AMD low sales figures won't justify all the investment in testing, validation and SG&A for the big core server market. And given AMD current server numbers, that time isn't too far down the road. [AJL, Agreed, but with 28nm FD-SOI, AMD could offer a substantial improvement in performance and core count, so the time to strike is now. I think if AMD could make this kind of jump every two years, which is possible with an responsible adult overseeing the CPU architectures. Jim Keller will likely re-introduce sanity back into AMD CPU development]

I'm really tired to see every company + dog from the SOI club touting benefits of the thing and nobody whose money depends on selling chips adopting the thing, except for AMD, and they have a reason for that.

AMD only went with SOI because IBM was willing to sell them a process, and AMD didn't have the money to develop process in-house. (for those who tend to see Intel and AMD as two giants of the same category fighting for the market, they aren't, and the differences start here).

And if the WSA is correct by ditching SOI AMD is going to save the paltry sum of 80 million dollars per year. When you have a magic dust that can give you 40% improvements in voltage for 0 additional manufacturing costs, this magic dust *is* worth 80 million dollars per year, and yet AMD refrained to pay those 80 millions.

While we all agree that AMD management isn't the most brilliant management team of the world, don't you think this decision is too stupid to pass on? If not for their intelligence, they would for their want of make money, and SOI would raise their annual bonus big time. [AJL, Yes and Yes!]

As for me, until I see someone whose life depends on manufacturing chips going with SOI, I'd refrain from believing on STM/GLF claims. [AJL, A fair point, technology demos do not impress as much as HVM does.]

Agreed here. I don't think AMD has a better prospect with ARM than with x86, and in ARM they will face *a lot* of companies with *a lot* more financial muscles than them. [AJL, Yes, this is such a 'DUH' that I'm really worried that Papermaster sold AMD execs an idea as useful as screen doors for a submariine ]


Not really. In order to succeed they must diminish the gap with Intel. There won't be any point in take 2 years to deliver substantial improvements in perf/watt and raw performance if Intel deliver more improvements and the gap grows. AMD will be lower and lower in the value ladder, and soon it won't be worth to build an AMD server.

Wanna have an idea on what I'm talking about? Look at AMD marketing. They went from talking big, things like data centers, cloud computing, HPC servers to this make huge events to announce affordable, single socket chips for... web hosting.

Their confidence collapsed, in large part because Bulldozer was such an expensive and miserable failure. Dirk deserves to be out because of this. But then in comes Read with better accountability, but he also appears to be using the failure Bulldozer as a whipping boy on the value of big cores. The failure of Bulldozer was a failure of responsible adult supervision; it should have been seen for what it was and killed off in time to design a new 32nm octo-core Thuban as a fall back till a more successful follow-on architecture could be developed (which still could have been CMT, but CMT done right, as is likely the case with steamroller). Look at how long it took for Intel to get HT right, it started out life as a disable feature, so that Intel could enable it on ES processors in order to characterize it's strength and weaknesses which could be re-mediated in compiler improvements b/4 HT processors were released en-masse.

I do see why AMD has faced defections from the ranks of SoC developers. AMD has perfectly good SoCs based around the Jaguar core, but they are pointless pursuing ARM SoCs against a mass of giants as you point out (and midgets with lots of VC and DARPA money backing them, like Tilera).

I wonder if the delay in getting FD-SOI into Dresden is that GF won't go with it unless they have a big customer on board like AMD, but perhaps AMD is gun shy since they will be back on the hook for $80M (which they'd rather pour into their efforts around ARM) because GF won't even cut AMD a deal even if that 'goodwill' would attract more customers than their current inflexible stance does. What a cluster fcuk! As an enthusiast, it's just disgusting to see AMD, which IMHO, could still recover, being burned at every corner by a magnitude of 'small mindedness' that isn't often seen outside of politics D:

 

[Ajay]

But what if you can get the magic dust for free?
What AMD ditched was their special SHP process that was PD-SOI, not FD-SOI. PD-SOI is running out of steam, technically not usable any longer, that is why STM and IBM make/made research about FD-SOI/ET-SOI.

Now if you have to pay for 28nm PD-SOI but can get 28nm FD-SOI for free what would you do?

The disadvantage is loosing a few ~100 MHz clock headroom because the FD-SOI is based on the LP-process, but in return you get much lower leakage, especially at lower clocks. And if you want invest the additional wattage for higher clocks e.g. for some FX CPUs, you can do so by body biasing.

Obviously an easy decision...

We can only hope, it seems like such an easy decision, unless there is some odd problem in scaling FD-SOI that we are unaware of. Other than that, all I can imagine is something like what I wrote above in my last paragraph being true

 

[ShintaiDK]

The gains would be from:

1. Architectural improvements in SR (some say up to +30% in IPC**)
2. The Shrink to 28nm
3. The addition of FD-SOI

So the 50% number is likely a bit conservative, for single thread performance. AMD could choose to hit the 50% bump in IPC, but drop the power usage as well - offering a substantial improvement in performance/watt. The would class SR Opterons as a huge boost for current AMD customers and would grab back a few % of the server market as well (I'm sure Dell, IBM, HP, etc would like an alternate source of CPUs for at least certain parts of their markets, if only to have some leverage to use when negotiating with Intel).


** probably on code profiles that s*ck in BD/PD. Less in other areas.

You are in hype territory.

 

[MisterMac]

The gains would be from:

1. Architectural improvements in SR (some say up to +30% in IPC**)
2. The Shrink to 28nm
3. The addition of FD-SOI

So the 50% number is likely a bit conservative, for single thread performance. AMD could choose to hit the 50% bump in IPC, but drop the power usage as well - offering a substantial improvement in performance/watt. The would class SR Opterons as a huge boost for current AMD customers and would grab back a few % of the server market as well (I'm sure Dell, IBM, HP, etc would like an alternate source of CPUs for at least certain parts of their markets, if only to have some leverage to use when negotiating with Intel).


** probably on code profiles that s*ck in BD/PD. Less in other areas.

No new architecture has EVER been released with 50% IPC.

Nevermind 30% IPC.


So your stating your internet avatar reputation on - AMD somehow improving IPC\ST more than ANY other CPU uARCH in history?

After PD, after BD, after AMD in this sad state - they're supposed to somehow do better than anything ever done before?


....really?


Listen - i have some investment opportunities id like to talk to you about.

 

[MisterMac]

Thanks for my first, signature Ajay

 

[mrmt]

But what if you can get the magic dust for free?

(...)

Now if you have to pay for 28nm PD-SOI but can get 28nm FD-SOI for free what would you do?

You don't get the magic dust for free, as you have to pay royalties to STM on top of the more expensive wafer. And so far nobody wants this magic dust, except a zombie JV where one of the partners is the developer of the node.

You can be right and SOI can be the game changer STM is teasing, but I prefer to wait and see. This isn't the first time the SOI consortium touts advantages and delivers nothing. Credibility is easy to lose, and the SOI consortium lost its credibility among MPU companies long time ago.

 

[Ajay]

Thanks for my first, signature Ajay

That's pretty immature. Making a sig out of a one-liner taken out of context.
I hope I at least get an invited to you 14th Bday

 

[Ajay]

You don't get the magic dust for free, as you have to pay royalties to STM on top of the more expensive wafer. And so far nobody wants this magic dust, except a zombie JV where one of the partners is the developer of the node.

You can be right and SOI can be the game changer STM is teasing, but I prefer to wait and see. This isn't the first time the SOI consortium touts advantages and delivers nothing. Credibility is easy to lose, and the SOI consortium lost its credibility among MPU companies long time ago.

Seems like IBM usually get what it wants out of these consortia.

 

[AtenRa]

What we dont know is the differences from 32nm HKMG SOI to 28nm HKMG bulk/SOI. Density should be close to 10-15% higher for the 28nm, the real question is how much lower does the power goes.
I dont believe 28nm HKMG bulk will have any advantages over 32nm HKMG SOI except perhaps a 10-15% higher density. And i will challenge that because 28nm HKMG bulk will need bigger transistors to get the same performance as with 32nm HKMG SOI. To conclude, i dont see AMD using the 28nm HKMG bulk for the big cores.

How about 28nm HKMG FD-SOI ?? Too late for AMD, i believe they have been following the 20nm train and they dont look back.

 

[mrmt]

Seems like IBM usually get what it wants out of these consortia.

http://www.commonplatform.com/about/manufacturing_partners.asp

IBM provides expertise in advanced research, process-aware chip design, technology development, and system design.

GLOBALFOUNDRIES enhances the Common Platform collaboration with its pure-play foundry model, manufacturing economies of scale and support across a full lifecycle of products.

Samsung complements the team with its deep sub-micron expertise, strong consumer- and low-power focus, along with manufacturing economies of scale.

Of course IBM gets. They are doing (and probably paying) the research, making the design rules, roadmaps and other critical definitions. It's IBM's turf, GLF just pays to have access to the tech and implement the process and Samsung is doing legal industrial espionage.

 

[Ajay]

You are in hype territory.

Maybe. But the hype, as far as I've read was 50% increase in IPC for SR, before FD-SOI! Someone (on the internet - I know, silly me) estimated that if all the changes to SR works as advertised that it would increase IPC by 30%. Now that I think of it, that was over BD. Taking that estimate, and the estimates of the performance benefits of 28nm FD-SOI and backing them off a bit, 50% sounds more reasonable (over two generations). Now if the clocks drop a bit on 28nm FD-SOI, the actual raw performance increase wouldn't be as high.

PD was what 10-15% on the same exact node with reduced power as well. So if the team working on SR manages to get rid of the major performance roadblocks found in BD & PD and it is produced on a process that allows for much faster xtor switching/Hz and AMD has a few more xtors to work with if need because of the shrink, then maybe, just maybe, it isn't really hype.

Unfortunately, we won't know anything unless GloFo signs on to 28nm FD-SOI and is able to show appreciable performance gains. Then toward the end of next year we will see what SR (-L3$) can do on 28nm bulk. So yeah, all of this is speculation at this point.

 

[Ajay]

What we dont know is the differences from 32nm HKMG SOI to 28nm HKMG bulk/SOI. Density should be close to 10-15% higher for the 28nm, the real question is how much lower does the power goes.
I dont believe 28nm HKMG bulk will have any advantages over 32nm HKMG SOI except perhaps a 10-15% higher density. And i will challenge that because 28nm HKMG bulk will need bigger transistors to get the same performance as with 32nm HKMG SOI. To conclude, i dont see AMD using the 28nm HKMG bulk for the big cores.

How about 28nm HKMG FD-SOI ?? Too late for AMD, i believe they have been following the 20nm train and they dont look back.

Well, if that's true - to bad. When is 20nm bulk going to come out. GF announced it had produced 20nm test shuttles (very small ARM cortex cores) in 4Q12, when will the process be mature enough for producing CPUs? Would ~two years be a reasonable guess given GF's performance on 28nm production?

 

[ShintaiDK]

Maybe. But the hype, as far as I've read was 50% increase in IPC for SR, before FD-SOI! Someone (on the internet - I know, silly me) estimated that if all the changes to SR works as advertised that it would increase IPC by 30%. Now that I think of it, that was over BD. Taking that estimate, and the estimates of the performance benefits of 28nm FD-SOI and backing them off a bit, 50% sounds more reasonable (over two generations). Now if the clocks drop a bit on 28nm FD-SOI, the actual raw performance increase wouldn't be as high.

PD was what 10-15% on the same exact node with reduced power as well. So if the team working on SR manages to get rid of the major performance roadblocks found in BD & PD and it is produced on a process that allows for much faster xtor switching/Hz and AMD has a few more xtors to work with if need because of the shrink, then maybe, just maybe, it isn't really hype.

Unfortunately, we won't know anything unless GloFo signs on to 28nm FD-SOI and is able to show appreciable performance gains. Then toward the end of next year we will see what SR (-L3$) can do on 28nm bulk. So yeah, all of this is speculation at this point.

Didnt Phenom and Bulldozer learn you the hard way about these hypes?

It doesnt become more real the more you times you write it.

 

[AtenRa]

Well, if that's true - to bad. When is 20nm bulk going to come out. GF announced it had produced 20nm test shuttles (very small ARM cortex cores) in 4Q12, when will the process be mature enough for producing CPUs? Would ~two years be a reasonable guess given GF's performance on 28nm production?

20nm will be ready for 2014

 

[Ajay]

Didnt Phenom and Bulldozer learn you the hard way about these hypes?

It doesnt become more real the more you times you write it.

Yeah, you have a point. Intel delivered Core2, Nehalem/Bloomfield and Sandy Bridge close to expectations (well, Core2 was probably above expectations). There is no point in being an optimist wrt AMD. I formally unsubscribe from this thread.

 

[MightyMalus]

SR won't fix the L3 cache latency problem, I doubt it will increase performance that much over PD. Excavator tho, will fix it(rumor) and will use that automated design they mentioned awhile ago. Hopefully that will be a winning core.

If AMD manages the 15% IPC gain with SR over PD, and 30% with EXV over SR, that will put them much nearer to Intel, if Intel keeps focusing on lower watts over IPC gains like they did with IVB and now possibly with HSW.

 

[SocketF]

You don't get the magic dust for free, as you have to pay royalties to STM on top of the more expensive wafer.

HMm good question if AMD had to pay or not. I guess STM and GF are discussing that right now.

You can be right and SOI can be the game changer STM is teasing, but I prefer to wait and see. This isn't the first time the SOI consortium touts advantages and delivers nothing. Credibility is easy to lose, and the SOI consortium lost its credibility among MPU companies long time ago.

Yes I am also a bit cautious, everything just sounds soooo good ... suspicious ^^
By the way what where the old promises that they failed to deliver? I wasnt interested in that topic back then I guess, dont know.

Didnt Phenom and Bulldozer learn you the hard way about these hypes?

First generation is always bad (Ph1), second is better(Ph2), 3rd is great (Ph2 X6 ;-)) At least we can hope that it will be the same with BDv1,2,3 ;-)

20nm will be ready for 2014

Yes but not many benefits and expensive due to double pattering. In my opinion a node to skip, especially if 14nm is ready just a year after.

SR won't fix the L3 cache latency problem, I doubt it will increase performance that much over PD.

Dedicated decoders, bigger caches, more dispatch bandwidth will surely due their job ;-)

Excavator tho, will fix it(rumor) and will use that automated design they mentioned awhile ago. Hopefully that will be a winning core.

Maybe we'll already see that with the first Kaveris. Would be an explanation for the postponement of the server parts to 2014 (end of 2013 for Kaveri) , too. There were also rumors that we'll see the Steamroller B core. As Steamroller is the core-architecture, not a chip it cannot be the chip-revision, thus I assume they changed the core a bit. Something in beetween Steamroller and Excavator maybe, like Vishera's Piledriver cores, which are not 100% identical with Trinity's PD-cores.

 

[AtenRa]

Yes but not many benefits and expensive due to double pattering. In my opinion a node to skip, especially if 14nm is ready just a year after.

32nm to 20nm will have a substantial increase in performance and lower power, it is after all a full node shrink unlike the half node that 28nm is over the 32nm. Also, price is not a number one problem because they will manufacture the big cores with it, that means server and high end desktop chips.

14XM is only available for LPM, not usable for high performance chips. They may use it for SoCs after 28nm.