Intel Confirms 8th-Generation CPUs Will Stick to 14nm Node

[Megalith]

I guess this should have gone along with that Cannonlake post I made earlier, but It sounds like Intel’s three-stage “Process-Architecture-Optimization” is already out. Okay, let’s play marketing department—what should we call the fourth step? And does this basically mean that 10nm yields aren’t as good as Intel had hoped?

In a somewhat confusing move, this week Intel confirmed that its next 8th generation Core processors would remain on the 14nm node, despite previously saying that it would have 10nm ready to go later this year. Speaking at its annual investor day in California, Intel spoke about its plans for 8th gen, adding that it would be sticking with 14nm for its mainline CPUs for the fourth year in a row. Earlier this year we saw the launch of Kaby Lake, which was the ‘Optimization’ step in Intel’s new three-stage ‘Process-Architecture-Optimization’ release schedule, which replaced the old Tick-Tock cycle last year. By returning to 14nm for 8th generation, it seems that a fourth step is being added to Intel’s cycle.

 

[Rvenger]

Skylake-E, Kabylake-E all dead on arrival now?

They sat on their hands with the intentions of AMD being dead already I guess, they are getting bit in the ass now and are scuffling it seems.

 

[Shintai]

The bar keeps getting lowered to run a "techsite".

Mobile:
CNL-Y(2+2), CNL-U(2+3e) are 10nm.
KBL-R(4+2) is 14nm+, CFL-H(4+2 and 6+2) is 14nm++.
TGL is 10nm+

Desktop:
CFL-S is 14nm++.(4+2 and 6+2)
TGL-S is 10nm+.

HEDT/Server:
SKL-X/EP is 14nm++. KBL-X is 14nm+, CFL-X is 14nm++. CNL-X/EP is 10nm+. TGL-X/EP is 10nm++.

Then later Sappfire Rapid comes on 7nm.

 

[Gideon]

Node shrinks are getting far more difficult now. I am guessing yields are what are making them use 14nm still.

 

[Ultima99]

Node shrinks are getting far more difficult now. I am guessing yields are what are making them use 14nm still.

Yields are undoubtedly bad especially at higher clocks so they'll use all the first Cannonlake chips for mobile while they use Coffee Lake which will bring 6 core to the mainstream on the very mature high yield 14nm process.

Skylake-X and Kaby Lake-X will be out this summer.

 

[Shintai]

Its more about cost and design time than yields. Also its silly to go for a new node when you can archive the same with an older node. TSMC is doing 4 nodes on 16nm, despite calling one for 12nm now.

With Intel its node first, then node+ and lastly node++.

14nm to 14nm+ for example gave plus ~20% for mobile. While desktop for obvious reasons was much lower. But even in the top it went from 4.0/4.2 to 4.2/4.5. For 14nm++ we will most likely get a CFL 6 core at ~4Ghz give or take within the same 95W.

 

[Gideon]

Its more about cost and design time than yields. Also its silly to go for a new node when you can archive the same with an older node. TSMC is doing 4 nodes on 16nm, despite calling one for 12nm now.

With Intel its node first, then node+ and lastly node++.

14nm to 14nm+ for example gave plus ~20% for mobile. While desktop for obvious reasons was much lower. But even in the top it went from 4.0/4.2 to 4.2/4.5. For 14nm++ we will most likely get a CFL 6 core at ~4Ghz give or take within the same 95W.

Oh look the Intel spin rep. Going for a third time on the same node was not in their plans originally. Mobile gain of 20% was mostly clock speed, that is why for desktop trying to find 5% of a gain was almost impossible. I highly doubt they will be able to clock up on 14nm much more then they already have. At some point you have pulled every trick in the book to make it better.

 

[JosiahBradley]

So Intel claimed 15% from 6 to "7" and got ~5, so 6v2 might be another 6%? And that's just pushing the clock higher and burning more power at that curve. Looks like the CPU is as dead as ever. Really wish multithreaded software would start to be written now.

 

[Deleted member 93354]

Earlier this year we saw the launch of Kaby Lake, which was the ‘Optimization’ step in Intel’s new three-stage ‘Process-Architecture-Optimization’ release schedule

What optimization? You mean the fact they improved the 3D transistor that did nothing for clock speed or IPC?

 

[Deleted member 93354]

So Intel claimed 15% from 6 to "7" and got ~5, so 6v2 might be another 6%? And that's just pushing the clock higher and burning more power at that curve. Looks like the CPU is as dead as ever. Really wish multithreaded software would start to be written now.

Software improvement isn't just as easy as saying "Just multi thread it" There's VERY little that benefits from multi threading. Pixel processing/3D, audio transform functions, simulations, statistical analysis/file compression, tournament sorts, and movie processing functions are about it. I know that sounds like a lot, but it's not. It's actually a very small percentage of code. And when you are dealing with code which requires concurrent read/write access to the same resources like databases, thread locks are very time consuming, hampering the effectiveness of those task.

Intel is facing the same problem it had years ago. Processors are fast enough for things like word and excel (unless running VBA). Things like database processing are where intel likes to rape companies with Xeons that offer only a small improvement. So what's intel's answer to stalled clock speeds and inability to increase IPC and a lack of a reason to upgrade? "Improve a totally crappy iGPU and just make it mostly crappy for the 1% of people who could use it and don't want to buy a dedicated GPU which is much more powerful."

 

[N4CR]

IPC doesn't really give Intel shit these days for their main markets, this is why they had that rather unusual quad 5ghz xeon for the freeloader financial transaction market types being discussed later last year. The sort of parasitic industry that build private fibre lines because it gives them a 2ms advantage over their rivals...

So Intel realized that doesn't sell them new chips and instead decided to build in more specialised areas in their chips tailored for various vendors, e.g. Amazon has far higher database hits per second than e.g. Coke Cola webservers, so they get tailored Xeons to take advantage of this. Each company of large enough size gets this special treatment, this has been mentioned in articles before and it's rather secret sauce... that's how and why we get all those xeons dumped on the market for peanuts every few year cycles.

We just focus on IPC gains because it's all most of us know and really care about. Meanwhile Intel is playing a different game with stuff we don't even get to test or know about.

So sure, on paper they're probably failing but I'd say in these speciality but high volume markets, they're not too worried about AMD.

 

[IdiotInCharge]

Intel has been focused on the mobile experience for a while- and here, they've succeeded, even when staying on a node far longer than tick-tock originally prescribed.

6000 -> 7000 is a good example of this; it did f-all for desktops, but for mobile to be able to sustain higher clockspeeds? That's gold for the user experience.

I'm good with my 6700k and will be so for years, barring a disruptive application and accompanying hardware, but for mobile? I'm waiting for ASUS to get the 7000-series on the shelves.

 

[thesmokingman]

IPC doesn't really give Intel shit these days for their main markets, this is why they had that rather unusual quad 5ghz xeon for the freeloader financial transaction market types being discussed later last year. The sort of parasitic industry that build private fibre lines because it gives them a 2ms advantage over their rivals...

So Intel realized that doesn't sell them new chips and instead decided to build in more specialised areas in their chips tailored for various vendors, e.g. Amazon has far higher database hits per second than e.g. Coke Cola webservers, so they get tailored Xeons to take advantage of this. Each company of large enough size gets this special treatment, this has been mentioned in articles before and it's rather secret sauce... that's how and why we get all those xeons dumped on the market for peanuts every few year cycles.

We just focus on IPC gains because it's all most of us know and really care about. Meanwhile Intel is playing a different game with stuff we don't even get to test or know about.

So sure, on paper they're probably failing but I'd say in these speciality but high volume markets, they're not too worried about AMD.

It's not just Intel producing custom chips, AMD have been doing that since way back just like Intel.

 

[Shintai]

*waah waaah*. Going for a third time on the same node was not in their plans originally. Mobile gain of 20% was mostly clock speed, that is why for desktop trying to find 5% of a gain was almost impossible. I highly doubt they will be able to clock up on 14nm much more then they already have. At some point you have pulled every trick in the book to make it better.

Its been their plan for quite some time now. And whatever makes the improvement is irrelevant as long as its there.

You are not getting something big for desktop even if you shrinked it to 5nm. That's a whole other issue that you obviously haven't understood yet.

 

[Gigus Fire]

Its been their plan for quite some time now. And whatever makes the improvement is irrelevant as long as its there.

You are not getting something big for desktop even if you shrinked it to 5nm. That's a whole other issue that you obviously haven't understood yet.

Why not?
If you used the same architecture and stuffed in 4 more cores at 5nm, wouldn't that be an improvement?

 

[Shintai]

Why not?
If you used the same architecture and stuffed in 4 more cores at 5nm, wouldn't that be an improvement?

I think the reference is core speed. You can already get 22 cores on a die.

 

[sboucher]

—what should we call the fourth step?

Tick, Tock, Twang, Sproing!

 

[Gigus Fire]

I think the reference is core speed. You can already get 22 cores on a die.

The 22 core xeon processor sits on a 454 mm² die size. I can't find the exact die size for a 7700k, but a 4790 is 177mm2.
The larger the die size, the higher the costs. If you can fit 8 cores in the same space as 4 cores, the die size stays the same, thus reducing costs.

 

[Shintai]

The 22 core xeon processor sits on a 454 mm² die size. I can't find the exact die size for a 7700k, but a 4790 is 177mm2.
The larger the die size, the higher the costs. If you can fit 8 cores in the same space as 4 cores, the die size stays the same, thus reducing costs.

Its not that simple anymore. A 7700K is 122mm2 and may cost more to make than a 177mm2 4790.

From a foundry perspective its much easier. A 14/16nm 5B transistor die from Glofo/Samsung/TSMC pretty much cost the same of a 28nm 5B transistor die one from the same foundries.

 

[Gigus Fire]

Its not that simple anymore. A 7700K is 122mm2 and may cost more to make than a 177mm2 4790.

From a foundry perspective its much easier. A 14/16nm 5B transistor die from Glofo/Samsung/TSMC pretty much cost the same of a 28nm 5B transistor die one from the same foundries.

I would just assume that costs over time would decrease.

 

[xorbe]

How long did they both spend on 22? What did anyone expect to happen on 14? How long will they be on 10 and 7? Just sayin' ... things ain't hoppin' like they used to.

 

[Shintai]

I would just assume that costs over time would decrease.

Yes, but so does the old nodes. Its never going to be the same as before the 28nm cut. Its the past.

 

[nutzo]

The 22 core xeon processor sits on a 454 mm² die size. I can't find the exact die size for a 7700k, but a 4790 is 177mm2.
The larger the die size, the higher the costs. If you can fit 8 cores in the same space as 4 cores, the die size stays the same, thus reducing costs.

There are usually diminishing returns as you add more cores. For most desktop users, it doesn't make sense to go with more than 4. Most office workers are better of with a fast i5 with dual core/hyper threading than a slower i5 quad core.

On servers there is also the issue of software costs. With something like SQL server that is priced per core, a lower number of faster cores can end up being cheaper than a larger number of slower cores when you include the software licensing.

 

[Gigus Fire]

There are usually diminishing returns as you add more cores. For most desktop users, it doesn't make sense to go with more than 4. Most office workers are better of with a fast i5 with dual core/hyper threading than a slower i5 quad core.

On servers there is also the issue of software costs. With something like SQL server that is priced per core, a lower number of faster cores can end up being cheaper than a larger number of slower cores when you include the software licensing.

Where are you basing your figures from?
From a cost perspective, sure. Going with a cheaper chip will always been more cost effective. But for efficiency? If i had 6 apps running, wouldn't an 8 core be a faster solution where every core could be dedicated to a single process than a quad core which would have to swap programs on the stack back and forth?
I get it there's a diminishing returns, but it certainly doesn't start at 4 or 8 even.

On another related note, i never understood what exactly intel is marketing their high end desktop chip for. 7700k is their top of the line quad core model which is marketed for gamers and people who want performance, yet there's an igpu which takes up 30-40% of the space. They could have easily added 2 more cores with that space.

 

[Gideon]

Its been their plan for quite some time now. And whatever makes the improvement is irrelevant as long as its there.

You are not getting something big for desktop even if you shrinked it to 5nm. That's a whole other issue that you obviously haven't understood yet.

You also said Ryzen will be just like the Bulldozer launch, looking pretty much like your wrong (not surprised). I dont think you get what I am saying, Intel has to come out with a new design not a shrink and tweak to their current design. They have milked this design for long enough and it's time for a fresh new architecture. Problem for Intel is they figured AMD could not close the gap, by their reaction lately, it tells me AMD has in fact closed the gap. Also look at that newegg has a sale on Intel chips.. go figure. Keep grasping for straws why Intel has not failed us enthusiasts.

 

[westrock2000]

The bar keeps getting lowered to run a "techsite".

Mobile:
CNL-Y(2+2), CNL-U(2+3e) are 10nm.
KBL-R(4+2) is 14nm+, CFL-H(4+2 and 6+2) is 14nm++.
TGL is 10nm+

Desktop:
CFL-S is 14nm++.(4+2 and 6+2)
TGL-S is 10nm+.

HEDT/Server:
SKL-X/EP is 14nm++. KBL-X is 14nm+, CFL-X is 14nm++. CNL-X/EP is 10nm+. TGL-X/EP is 10nm++.

Then later Sappfire Rapid comes on 7nm.

Next time....use more insider short hand.

 

[CombatChrisNC]

Why can't Intel stick to a socket like they stick to a node?

 

[Snowdog]

Where are you basing your figures from?
From a cost perspective, sure. Going with a cheaper chip will always been more cost effective. But for efficiency? If i had 6 apps running, wouldn't an 8 core be a faster solution where every core could be dedicated to a single process than a quad core which would have to swap programs on the stack back and forth?
I get it there's a diminishing returns, but it certainly doesn't start at 4 or 8 even.

If you think you need cores equal to the the number of simultaneous running apps, then you don't get this at all. Diminishing returns start at 2 cores for office work.
Modern programs on modern OS/CPU don't hog resources, or take meaningful time to switch. You could run 10 apps on one core and not have an issue.

You likely have 100+ threads and processes running on your computer right now, but switching between them is nothing on a modern CPU.

 

[Shintai]

Where are you basing your figures from?
From a cost perspective, sure. Going with a cheaper chip will always been more cost effective. But for efficiency? If i had 6 apps running, wouldn't an 8 core be a faster solution where every core could be dedicated to a single process than a quad core which would have to swap programs on the stack back and forth?
I get it there's a diminishing returns, but it certainly doesn't start at 4 or 8 even.

On another related note, i never understood what exactly intel is marketing their high end desktop chip for. 7700k is their top of the line quad core model which is marketed for gamers and people who want performance, yet there's an igpu which takes up 30-40% of the space. They could have easily added 2 more cores with that space.

Faster cores is better then more cores in most cases outside something with high concurrency like servers. With Coffee Lake you get the extra cores, but only because 14nm++ will allow this without too much penalty to per core performance.

For scaling, here is a real world example:
https://www.pugetsystems.com/labs/articles/Adobe-After-Effects-CC-2015-3-Multi-Core-Performance-843/