Skylake/Broadwell Roadmap Update @Vr-zone

[imported_ats]

For me the desktop BW-K always sounded like an unexciting product. Ever since I have been buying Intel CPUs, I have never seen an impressive Tick. The Ticks matter more for laptops and 6-8 cores where the shrink helps with lower power usage/battery life. Otherwise, I find upgrading on the Tock most impressive:

1) Larger gains in IPC
2) Usually more dramatic changes to motherboard chipsets and features.

At this point I think 2-year-old Haswell and especially 4-year-old Sandy owners could care less about a Broadwell-K on 1150 since Skylake brings way higher IPC and more advanced features. For new PC builders, BW-K also makes no sense since they would be buying obsolete DDR3. At least with DDR4, you could reuse it in 2-3 years with another Intel Tock.

It just sucks that Intel continues to launch the latest architecture on the mainstream socket first, putting their flagship X99 socket 1-1.5 years behind. If BW-E is scheduled for Q1 2016, I presume we won't even see Skylake-E until Q4 2016 or later. That is what I hate the most as I think the premium platform should always get the latest architecture like the old i7 920. Right now it seems Intel is either not confident enough to go 6-8 cores on the latest architecture until they see yields on Skylake or its a permanent strategy to treat their workstation/premium customers as 2nd class citizens.

I bet if a 6-core Skylake-E launched Q3-4 2015, a lot more PC enthusiasts would have considered ditching the mainstream platform.

-E parts are and have always been cutdown server parts. There has never been enough of a market for -E parts to make them a priority. -EP and -EX drive the design. Realistically, -E parts only make sense if you need the additional memory capacity and/or the addiction PCI-E bandwidth. In general the -E/EP/EX design isn't aimed at desktop workloads and is optimized around high core counts (currently 15-18c).

 

[arandomguy]

Also, I think I read somewhere that one team designed Haswell and Skylake, while the other designed Broadwell. This might have some bearing on whether FIVR is present or not.

I believe Skylake is primarily being developed at Intel Israel while Haswell was Intel California?

If we want to parallel Zen and Keller than Intel Israel was also the primary for Sandybridge :awe:

Because the Intel management wants to save a few cents per unit according to the article. Can that be true?

The cost reasoning behind it never made sense to me.

Intel has shown they are willing to have separate packaging lines with Devil's Canyon. They've also shown they are willing to charge extra for unlocked CPUs and that customers will pay for them. So why not just pass that cost on along with a healthy profit margin? Even at something like 200%+ gross margins on soldered parts that can't possible add more than $20 to the price which I'm sure people would pay.

I believe the other explanation that has been given before is the small die sizes have issues if soldered to the IHS?

 

[SAAA]

So what could be the difference between K and "unlocked", does this mean that Broadwell will feature unlimited multiplier like high end X models? Or maybe the IGP is locked on Skylake (and GT2 model only) and you can overclock it on Broadwell?

Too much speculation here... I'm starting to think we won't see GT4 graphic anytime soon so that would probably be a refresh model, like Devil's Canyon, sometime in 2016.
That could cover up delays in 10nm products and satisfy people if it brings some base clock speed increase too.

 

[witeken]

If the roadmap is accurate, then it seems like maybe they finally got their 14nm yield problems mostly fixed.

It seems they will take a bit of a margin hit. Problems will be fixed by Q2 (wafers from Q2 will sell in Q3):

 

[witeken]

Something tells me we won't be seeing 10nm anytime soon :hmm:

What?

 

[Idontcare]

The cost reasoning behind it never made sense to me.

Intel has shown they are willing to have separate packaging lines with Devil's Canyon. They've also shown they are willing to charge extra for unlocked CPUs and that customers will pay for them. So why not just pass that cost on along with a healthy profit margin? Even at something like 200%+ gross margins on soldered parts that can't possible add more than $20 to the price which I'm sure people would pay.

I believe the other explanation that has been given before is the small die sizes have issues if soldered to the IHS?

Years ago when Intel released the G0 stepping to their B3 Q6600 they actually presented a slide at IDF in which they publicly went on record stating that the sole reason they increased the TJmax rating for the new G0 stepping chips over that of the B3 stepping was because they estimated could save ~$1-$2 per retail chip because they could use a cheaper stock HSF (higher TJmax means a crappier HSF is OK, but temps run higher of course).

Fast forward to present day, and yeah, Intel would absolutely make those kinds of choices just to save a nickle or a penny here and there.

And their shareholders expect nothing less. So as an end-user expect nothing more.

 

[Idontcare]

It seems they will take a bit of a margin hit. Problems will be fixed by Q2 (wafers from Q2 will sell in Q3):

Ouch. That is brutal. Haswell reduced costs over Ivy Bridge, without a change in node, more than Broadwell will reduce costs over Haswell with a change in node

Didn't Intel make a HUGE to-do about how their 14nm was not only in keeping with Moore's Law but that it actually accelerated it? WTF happened with 14nm broadwell then?

 

[arandomguy]

Years ago when Intel released the G0 stepping to their B3 Q6600 they actually presented a slide at IDF in which they publicly went on record stating that the sole reason they increased the TJmax rating for the new G0 stepping chips over that of the B3 stepping was because they estimated could save ~$1-$2 per retail chip because they could use a cheaper stock HSF (higher TJmax means a crappier HSF is OK, but temps run higher of course).

Fast forward to present day, and yeah, Intel would absolutely make those kinds of choices just to save a nickle or a penny here and there.

And their shareholders expect nothing less. So as an end-user expect nothing more.

Except in this case it makes them more money. Intel could be selling something that would cost <$5 extra to produce at a charge of $20+ extra to the end customer? That would be a gross margin of 300% for this add on feature versus Intel's current roughly 60% average on its products.

Unlike in the past they are selling separate overclocking enthusiast oriented SKUs at a premium, a premium in which customers are willing to pay and whom would surely also be willing to pay another premium just for a better thermal interface.

 

[AtenRa]

It seems they will take a bit of a margin hit. Problems will be fixed by Q2 (wafers from Q2 will sell in Q3):

After this does people still believe 14nm is cheaper than 22nm ?? In Q4 2015 Broadwell (smaller die) will reach the same cost as Haswell in Q2 2014.
Transistor price may fallen but actual product price is not, even with a smaller die size.

 

[myocardia]

Ouch. That is brutal. Haswell reduced costs over Ivy Bridge, without a change in node, more than Broadwell will reduce costs over Haswell with a change in node

Didn't Intel make a HUGE to-do about how their 14nm was not only in keeping with Moore's Law but that it actually accelerated it? WTF happened with 14nm broadwell then?

Hah, sounds as if Hector Ruinz found himself a new job, doesn't it? He's the best in the business at that!

Since I just built a new system, I'll be watching this one from the sidelines. Do we have any reference at all, as to either IPC gains, or to new instructions on either Broadwell or Skylake?

edit: Higher than normal IPC or clockspeed gains, not the usual 5-10%.

 

[elemein]

Hah, sounds as if Hector Ruinz found himself a new job, doesn't it? He's the best in the business at that!

Since I just built a new system, I'll be watching this one from the sidelines. Do we have any reference at all, as to either IPC gains, or to new instructions on either Broadwell or Skylake?

edit: Higher than normal IPC or clockspeed gains, not the usual 5-10%.

http://www.anandtech.com/show/8355/intel-broadwell-architecture-preview/2

Yes, for Broadwell, we get the usual 5% increase with some small internal changes to already existing components in the arch, as well as some new (niche) ISA extensions.

It's a nice tick.

 

[Essence_of_War]

How can one draw any conclusions from this graph at all with an unlabeled y-axis?

 

[Sweepr]

Braswell in Q2 might be acceptable (earlier than previous H2/2015 launch) but Skylake-U as soon as Q3? Not impossible but sounds fishy.

 

[Abwx]

Ouch. That is brutal. Haswell reduced costs over Ivy Bridge, without a change in node, more than Broadwell will reduce costs over Haswell with a change in node

Didn't Intel make a HUGE to-do about how their 14nm was not only in keeping with Moore's Law but that it actually accelerated it? WTF happened with 14nm broadwell then?

IDC, you seem to forget that the x axys is shifted for each CPU, to get an accurate picture time reference should be normalized and the curves shifted accordingly.

Haswell didnt reduce the cost, the curve start at Q3 13, look at the price of Ivybridge at Q1 13, this is the price at wich Haswell started two quarter later and the curve say that the cost of 22nm wasnt reduced much since these days but certainly that 14nm is very expensive, something like 10k$/waffer as a very conservative figure.

I guess that you re not a marketing man, isnt it, beware of thoses peoples ability to make apples look like oranges...

 

[witeken]

Ouch. That is brutal. Haswell reduced costs over Ivy Bridge, without a change in node, more than Broadwell will reduce costs over Haswell with a change in node

Didn't Intel make a HUGE to-do about how their 14nm was not only in keeping with Moore's Law but that it actually accelerated it? WTF happened with 14nm broadwell then?

14nm heavily reduces normalized cost per transistor, so with iso yields. You should make the comparison with Haswell in about a year.

Also, I though were reversed to using public slides? This one doesn't even have quantitative number, let alone a scale.

 

[witeken]

After this does people still believe 14nm is cheaper than 22nm ??

Yes it is. Density increased 2.0-2.2x, while wafer cost rose 1.3x, so transistor cost has fallen by at least 1.5x. But of course, transistor count has also increased.

Also note that it isn't a very specific slide: there are no numbers or scale. It doesn't even say what kind of cost it is. Is it cost per transistor? Platform cost?

Fortunately, this slide does not live in a vacuum. From the presentation (http://intelstudios.edgesuite.net/im/2014/live_im.html):

Intel expect 2015 platform cost to increase (the slide before that one) because of 14nm ramp. Stacy says that that is a normal phenomenon for Intel (you can also see it with Ivy Bridge, right). But in 2015 it will be a higher increase in platform cost than traditional node changes because of the low yields in H1 (by the way: even when 14nm catches up to 22nm in terms of yield, it still is 2 years behind, because the yield comparisons are offset by 2 years and in that time 22nm obviously is also improved).

This will come down like Ivy Bridge did, and it will asymptote lower than IB. Skylake will likely be lower than BDW, just like HSW.

 

[Eug]

Arrgghhh... I came in here to find out the roadmap for Skylake U and Skylake Y, but the Y is missing.

 

[witeken]

Arrgghhh... I came in here to find out the roadmap for Skylake U and Skylake Y, but the Y is missing.

You should expect a holiday release for SLK-Y. A 1 year delta makes sense to me.

 

[Fjodor2001]

You should expect a holiday release for SLK-Y. A 1 year delta makes sense to me.

What's the logic behind that? Broadwell-U -> Skylake-U only has a ~6 months delta for example.

 

[erunion]

The 4770R and the 65 W Broadwell shown on the latest roadmap are totally different. BGA vs LGA, locked vs unlocked. So I don't see how this 65 W Broadwell can be a replacement for the 4770R.

Also, actually a Broadwell-K is what has been expected for desktop LGA. What's strange is that it now appears to be 65 W TDP, and released at the same time as the Skylake 95 W LGA SKU. The whole point of Broadwell-K would be that it is unlocked and thus if overclocked potentially faster than the locked Skylake, otherwise it serves no purpose. But at 65 W TDP it is questionable if that can be achieved.

Earlier roadmap:

Believe what you like, but what I told you is official intel.

Intel talked about Broadwell iris pro when they announced Devil's Canyon. Key words: Desktop, Iris Pro.

http://www.anandtech.com/show/7875/new-unlocked-iris-pro-cpu-broadwell

So what could be the difference between K and "unlocked", does this mean that Broadwell will feature unlimited multiplier like high end X models? Or maybe the IGP is locked on Skylake (and GT2 model only) and you can overclock it on Broadwell?

Too much speculation here... I'm starting to think we won't see GT4 graphic anytime soon so that would probably be a refresh model, like Devil's Canyon, sometime in 2016.
That could cover up delays in 10nm products and satisfy people if it brings some base clock speed increase too.

K is a brand, unlocked multiplier is a feature. The slide says that Desktop Broadwell will be unlocked, it doesn't commit them to using the K brand. So it may or may not carry the K suffix(but likely could).

Whats also interesting is that the Haswell Refresh shows "K and non-K" for 95W, while Skylake only says "K". Suggesting no more locked i7.

 

[erunion]

double post

 

[Fjodor2001]

Hmmm.... well, I guess it's hard to know what the 65 W TDP LGA model in the roadmap refers to then. I guess it could be either one. It would be simpler if the roadmap in the OP was more clear on that. :\

But if it's the one you refer to, does it mean that Broadwell-K has been scrapped? Because then there is no mentioning of it on the roadmap.

Also, is there any point in having it unlocked, if it's 4 cores with GT3 and Iris Pro, and only rated at 65 W TDP? How high CPU frequency can you reach within that power envelope anyway, with that quite beefy iGPU consuming power too?

 

[NTMBK]

Hmmm.... well, I guess it's hard to know what the 65 W TDP LGA model in the roadmap refers to then. I guess it could be either one. It would be simpler if the roadmap in the OP was more clear on that. :\

But if it's the one you refer to, does it mean that Broadwell-K has been scrapped? Because then there is no mentioning of it on the roadmap.

Also, it there any point in having it unlocked, if it's 4 cores with GT3 and Iris Pro, and only rated at 65 W TDP? How high CPU frequency can you reach within that power envelope anyway, with that quite beefy iGPU consuming power too?

They probably just dropped the "K" branding because marketing decided to brand it as a "5770R" to match the "4770R"- use R for the Iris Pro SKUs. It's probably the same part.

 

[Ajay]

They probably just dropped the "K" branding because marketing decided to brand it as a "5770R" to match the "4770R"- use R for the Iris Pro SKUs. It's probably the same part.

That makes allot sense - especially for higher end AIOs and NUC like devices, which, last I read, seem to be gaining on std desktops.

 

[erunion]

Fjodor2001;37128412]Hmmm.... well, I guess it's hard to know what the 65 W TDP LGA model in the roadmap refers to then. I guess it could be either one. It would be simpler if the roadmap in the OP was more clear on that. :\

But if it's the one you refer to, does it mean that Broadwell-K has been scrapped? Because then there is no mentioning of it on the roadmap.

Also, it there any point in having it unlocked, if it's 4 cores with GT3 and Iris Pro, and only rated at 65 W TDP? How high CPU frequency can you reach within that power envelope anyway, with that quite beefy iGPU consuming power too?

Now the conversation gets interesting!

Yes, broadwell desktop was cancelled. With the 14nm delay Intel pulled in haswell refresh and broadwell desktop disappeared from roadmaps( except one product).

With broadwell delayed, Intel had a couple options to adjust their roadmap. They could cancel broadwell and have haswell carry on until skylake. They could release broadwell and push everything else back to maintain a one year product life. Or they could release broadwell with a short life by not delaying skylake.

Intel has done a mixture of those. Broadwell desktop is cancelled and broadwell mobile will be quickly replaced(and we may still find out some skylake products were pushed back)

Remember all those rumors about BGA only broadwell? Then the later rumor that skylake would bring sockets back?

Plans change. Its possible that when those things leaked they were accurate, but the roadmap changed since. desktop broadwell Iris Pro remained on the roadmap, originally it should have released long ago. Intel showed a renewed interest in unlocked chips in mid 2014. Its possible the socketed, unlocked chip Intel announced along devils canyon is actually a broadwell refresh.

Broadwell's Iris Pro's BGA heritage explains the carry over 65w TDP. But now that it is LGA, will it be capable of pulling 100W+? Since its unlocked on z97, that suggests it could but we'll have to wait and see. Like I said, it'll be interesting.