[WCCF] Intel Skylake 2015 Platform Details Revealed

[jdubs03]

I will personally be surprised to see Intel 10nm products available and in stock on Newegg in 2016, let alone 2015.

Personally I think 10nm retail availability (the only kind that matters to me) won't happen until sometime 2017.

IMO 10nm will be delayed for an economic reason, not a technology reason, same as what has happened with the 14nm ramp.

agreed. i think there is a shot at H2 2016. but yea i think H1 2017 for product availability.

 

[StrangerGuy]

I mean exactly what i say.


Desktop has no need for more cores - no averege joe WILL buy Haswell-E because of TRUE 8 CORE blablaba.

Only us performance geeks will, average joe won't care.
(I'm one of them - YAY FOR 16 INTEL THREADS EPENIS!).

That market will also in time reach a "fast enough" stage - but it's not there yet and everyone buys to get faster and newer.
...without actually considering phone bloatware and a million other reasons as to why they're device is slow.
Even my old parents know that reinstalling windows and cleaning stuff up helps on speed "feeling".


All of the cycle things we experienced are to come - but it's not there yet for mobile.
So it works for an OEM to go moar cores and sell more.

It doesn't for us - without knowing quite clearly what we gain from it.
And in bulldozers' case - it was... very little if anything.

Ooo, finally somebody gets it correct. Even for mobile where general consumer productivity apps sucks has everything to do with the limitations of touchscreens and UI and nothing with CPU power.

 

[witeken]

I will personally be surprised to see Intel 10nm products available and in stock on Newegg in 2016, let alone 2015.

Personally I think 10nm retail availability (the only kind that matters to me) won't happen until sometime 2017.

IMO 10nm will be delayed for an economic reason, not a technology reason, same as what has happened with the 14nm ramp.

Intel's recent future doesn't make me sure about volume production in 2015, but if Intel promised it, I don't yet see a reason to suspect a delta of more than 24 months between Broadwell and Cannonlake, possibly even less if 10nm isn't affected by 14nm's delay.

 

[Fjodor2001]

Intel's recent future doesn't make me sure about volume production in 2015, but if Intel promised it, I don't yet see a reason to suspect a delta of more than 24 months between Broadwell and Cannonlake, possibly even less if 10nm isn't affected by 14nm's delay.

The average period between CPU releases from SB -> B-K is more than 16 months, and increasing. I don't see any reasons or signs indicating that this trend should not continue.

 

[IntelUser2000]

The average period between CPU releases from SB -> B-K is more than 16 months, and increasing. I don't see any reasons or signs indicating that this trend should not continue.

Only true for Ticks(process change). Tocks can vary. Haswell was 13-14 months.

That varies depending on the market segment as well. Mobile U series launch was almost same, at June/July for Sandy Bridge(while one model was available by March or so, the rest weren't until Summer. Ivy Bridge, and Haswell were both Summer too.

With Broadwell the Y chips are early Q4, and maybe we'll see U chips at mid-Q4. While Haswell had Y chips launched at Q3, there were almost nothing until September, and actual shipment was Q4. I'd expect Broadwell Y devices in the hands of consumers by Q4.

I'd expect some limited SKUs of Skylake to launch somewhere in Q4 of 2015. If we assume like how 22nm and 14nm turned out, we may see Cannonlake in Mar/April of 2017.

 

[Fjodor2001]

Only true for Ticks(process change). Tocks can vary. Haswell was 13-14 months.

Not really sure what you're trying to say? Both Ticks and Tocks take about the same time lately, and both can be delayed. E.g. both SB and H were delayed, in fact every Intel CPU generation since (and including) SB has been delayed. H could have been delayed even further if Intel had decided to fix the USB chipset bug before release.

That varies depending on the market segment as well.

I was referring to desktop CPUs when I calculated the time between CPU generations.

With Broadwell the Y chips are early Q4, and maybe we'll see U chips at mid-Q4.

Has Intel indicated when in Q4 any of these chips will be released? If not, it might as well be late Q4.

I'd expect some limited SKUs of Skylake to launch somewhere in Q4 of 2015. If we assume like how 22nm and 14nm turned out, we may see Cannonlake in Mar/April of 2017.

What do you base those assumptions on? If we go by the latest track record from SB->B-K, desktop S will be released ~16 months after B-K, so that will be around 2016Q2/Q3.

 

[Idontcare]

Only true for Ticks(process change). Tocks can vary. Haswell was 13-14 months.

That varies depending on the market segment as well. Mobile U series launch was almost same, at June/July for Sandy Bridge(while one model was available by March or so, the rest weren't until Summer. Ivy Bridge, and Haswell were both Summer too.

With Broadwell the Y chips are early Q4, and maybe we'll see U chips at mid-Q4. While Haswell had Y chips launched at Q3, there were almost nothing until September, and actual shipment was Q4. I'd expect Broadwell Y devices in the hands of consumers by Q4.

I'd expect some limited SKUs of Skylake to launch somewhere in Q4 of 2015. If we assume like how 22nm and 14nm turned out, we may see Cannonlake in Mar/April of 2017.

The ambiguity in consistency of release schedule when factored by market segment is all due to OEM driven feedback and has little to do with Intel's capability.

We are witnessing the results of what happens when a near monopoly enters into an optimized "tidal lock" with its customers (the OEMs, not us end-users) where the tail wags the dog.

Intel listens, and rightfully so, to what its customers are telling it to do when it comes to inventory, production volumes, and release timelines.

What this creates, from our limited visibility, is a seemingly disjointed chaotic release schedule.

Which, when combined with our limited insight into the financial mechanization's that exist between Intel and its customers, is erroneously explained away as being the result of a technology shortfall (14nm isn't ready, etc etc) rather than justified as an unavoidable consequence of an economically optimized business relationship.

 

[jpiniero]

We are witnessing the results of what happens when a near monopoly enters into an optimized "tidal lock" with its customers (the OEMs, not us end-users) where the tail wags the dog.

Except the vast majority of end users don't want or even need faster computers. Intel could delay or even cancel Skylake and it wouldn't change sales much.

 

[Hulk]

Except the vast majority of end users don't want or even need faster computers. Intel could delay or even cancel Skylake and it wouldn't change sales much.

Unless they are wanting to sell tablets and have ultra portable laptops with tablet like run times and desktop like compute.

ARM is creeping up from phones to tablets, and now laptops. I'd think Intel would want the creep going the other way and Broadwell, Skylake, etc... is how they get there.

 

[jpiniero]

I'd think Intel would want the creep going the other way and Broadwell, Skylake, etc... is how they get there.

Intel probably thinks that way, but they will fail if they try to force Core through. Something radically new is going to be needed. And I do think it's coming - note that there haven't been any rumors about what's after Cannonlake. A rebuilt Atom Conroe-style could be a candidate for this, but I was thinking something more radical.

 

[Arachnotronic]

Intel probably thinks that way, but they will fail if they try to force Core through. Something radically new is going to be needed. And I do think it's coming - note that there haven't been any rumors about what's after Cannonlake. A rebuilt Atom Conroe-style could be a candidate for this, but I was thinking something more radical.

Goldmont should be a nice big "tock" for Atom, and Skylake should be a pretty big change from the current Haswell core.

Sandy Bridge was all about turbo, Haswell was all about power management and ridiculously low idle power in mobile, and Skylake will probably return to a focus on bring out some nice performance gains.

 

[IntelUser2000]

Goldmont should be a nice big "tock" for Atom, and Skylake should be a pretty big change from the current Haswell core.

Sandy Bridge was all about turbo, Haswell was all about power management and ridiculously low idle power in mobile, and Skylake will probably return to a focus on bring out some nice performance gains.

I think its important to manage realistic expectations.

The best they brought was 30% from Pentium III to Pentium M. It might have been 20%, but they have been sitting on not changing Pentium III for so long.

I think it'll be a surprise if they manage to do 20% on Skylake, even though IMO they need to do more. If they are going to do more though, I'd expect clock rates to go down, so it remains about a 20% gain.

With Sandy Bridge, it was 15% per clock increase plus Turbo. You may or may not remember it made a HUGE difference on Laptops. The gain was enormous. Since then, the gains are decreasing with expectations that it'll be harder and harder to get gains.

 

[jdubs03]

I think its important to manage realistic expectations.

The best they brought was 30% from Pentium III to Pentium M. It might have been 20%, but they have been sitting on not changing Pentium III for so long.

I think it'll be a surprise if they manage to do 20% on Skylake, even though IMO they need to do more. If they are going to do more though, I'd expect clock rates to go down, so it remains about a 20% gain.

With Sandy Bridge, it was 15% per clock increase plus Turbo. You may or may not remember it made a HUGE difference on Laptops. The gain was enormous. Since then, the gains are decreasing with expectations that it'll be harder and harder to get gains.

Agreed, we can't be too optimistic.

To keep the 2/1 ratio for Core/Atom, Intel has to scale performance equally on a percentage basis. IMO, Intel needs to boost the CPU for CT by >30%, meaning Broadwell needs that same increase to maintain the ratio, I don't know if Intel can do that, so Atom is more likely to encroach on the ratio then fall behind.

With Ivy Bridge and Haswell, the focus was getting down to lower TDP's, and Broadwell will continue that for some SKU's (obviously depending on TDP), but I don't see a reason why Intel could just keep the U-Series at 15W and add a solid 25-30% in performance improvement. If I remember correctly there was talk of Broadwell-Y being 30% more frugal with energy usage yet still increasing performance by about 10%, I would consider that as a win. We don't know as much about Skylake, but like I said on the other thread, at least to me, performance would be more important there as their power levels are already down to the point where they can have fanless tablets with BD-Y. Cherry Trail will serve the lower TDP segments just fine underneath BD-Y (maybe higher multi-thread performance too).

 

[shady28]

I went back and looked at the original article considering the excitement about GT4e -

Looks like to get GT4e, you'll be buying an i7. The lower two tiers are GT2 (HD 4600). i7 is really not in the same price segment as A10-7850K. It's priced to compete with something along the lines of a higher end i3.

4600 is a big improvement over 4000 for the i3 / i5, but it's non competitive with the current crop of APUs as far as GPU goes.

Now if they were doing GT3e in all i3 / i5 CPUs and GT4e in the i7s, then yes AMD and Nvidia would probably have some issues. But that's not what appears to be in the plan.

 

[Ajay]

Sandy Bridge was all about turbo, Haswell was all about power management and ridiculously low idle power in mobile, and Skylake will probably return to a focus on bring out some nice performance gains.

There are some limitations to FinFet that will hamper faster switching in 14nm - Homeles brought it up. So unless Intel performs some process Kung-Fu, we won't be seeing higher clocks.

Haswell was a pretty big change in uArch, and yet netted relatively small integer performance gains - though we are likely to see a big jump in FP (via AVX 3.x) over Haswell.

Intel's biggest problem will be that Skylake will be a much smaller die than Haswell, and for a Desktop CPUs, heat will still be a problem. The biggest gains will likely be in the low power versions of Skylake which will offer even higher performance/watt over BW.

I do wish we would see a +15% gain in integer performance over Broadwell - I just don't see anything that's changed to allow this. Maybe BW will surprise us, and then we can be more optimistic about Skylake.

 

[Homeles]

I think its important to manage realistic expectations.

The best they brought was 30% from Pentium III to Pentium M. It might have been 20%, but they have been sitting on not changing Pentium III for so long.

I think it'll be a surprise if they manage to do 20% on Skylake, even though IMO they need to do more. If they are going to do more though, I'd expect clock rates to go down, so it remains about a 20% gain.

With Sandy Bridge, it was 15% per clock increase plus Turbo. You may or may not remember it made a HUGE difference on Laptops. The gain was enormous. Since then, the gains are decreasing with expectations that it'll be harder and harder to get gains.

I think there's a couple of possibilities, and it's largely hinging on 3D. If Skylake has stacked memory and leverages it, it could be a decent bump. If not, we're looking at typical small gains from reinventing the 2D wheel. Maybe a bit bigger than Haswell in terms of perf/clock, but nothing too interesting.

However, we've already seen the performance figures for the 14nm process (up to 40%), which is actually better than 45 -> 32nm, where Intel claimed a "20-35% improvement over 45 nm high-k/MG logic transistors." So there is hope there. And I'm becoming increasingly confident that 10nm is going to be a huge node not only for Intel, but for everyone else, as it's the likely introduction of "post-silicon" channel materials.

There are some limitations to FinFet that will hamper faster switching in 14nm - Homeles brought it up. So unless Intel performs some process Kung-Fu, we won't be seeing higher clocks.

Actually, my argument was that moving from planar to FinFET is a one time penalty for overclocking. 22nm -> 14nm should allow for traditional performance gains.

Here's a beautiful MS paint example:

 

[IntelUser2000]

To keep the 2/1 ratio for Core/Atom, Intel has to scale performance equally on a percentage basis.

They are at LEAST 2x now, so Atom has a room to improve without violating what Intel wants.

If I remember correctly there was talk of Broadwell-Y being 30% more frugal with energy usage yet still increasing performance by about 10%,

It's 30% lower power at SAME performance. The rumors say the reason that Intel is able to put Broadwell-Y at such low TDP envelopes(4.5W) is mostly because the iGPU perf/watt increases dramatically. It's not enough with just process improvements.

If Skylake has stacked memory and leverages it, it could be a decent bump. If not, we're looking at typical small gains from reinventing the 2D wheel. Maybe a bit bigger than Haswell in terms of perf/clock, but nothing too interesting.

I don't doubt that Skylake can be a bigger gain than Haswell, oh no. The Oregon team that worked on Nehalem and Haswell focuses on "not-core" aspects, like chipset, platform, ecosystem, etc. The Haifa team that worked on Merom and Sandy Bridge focuses on the CPU core. Because of that, Desktop enthusiasts generally favor the Haifa designs while Server and Mobile benefit greatly by Oregon team designs.

About stacked memory: That's NOT coming in Skylake. They are still using MCP for their PCH. The next logical step would be to integrate the MCH, not stack the memory. Intel has also stated that they'll progressively increase integration of memory, not go straight away to stacking.

However, we've already seen the performance figures for the 14nm process (up to 40%), which is actually better than 45 -> 32nm, where Intel claimed a "20-35% improvement over 45 nm high-k/MG logic transistors.

They are pretty ambiguous on what voltage levels they are talking about. I BET its at lower voltages again. Remember the hype around 22nm? The 37% frequency gain is explicitely stated to be on Bay Trail...

At some point, they might need a new architecture for Core where it sacrifices frequency and significantly ups perf/clock.

And I'm becoming increasingly confident that 10nm is going to be a huge node not only for Intel,

And I am more and more inclined to believe the "node after revolution" that really brings exciting products for Intel. Remember 90nm and 45nm? They were kinda disappointing. The 65nm and 32nm were pretty awesome though.

 

[Shivansps]

I went back and looked at the original article considering the excitement about GT4e -

Looks like to get GT4e, you'll be buying an i7. The lower two tiers are GT2 (HD 4600). i7 is really not in the same price segment as A10-7850K. It's priced to compete with something along the lines of a higher end i3.

4600 is a big improvement over 4000 for the i3 / i5, but it's non competitive with the current crop of APUs as far as GPU goes.

Now if they were doing GT3e in all i3 / i5 CPUs and GT4e in the i7s, then yes AMD and Nvidia would probably have some issues. But that's not what appears to be in the plan.

Actually we need to wait and see, the changes to gen8 seems to be big, it could be one big jump, like the one from Intel Gen 4 to Gen 5, and them from Gen 5 to Gen 6, those where 2 BIG jumps right there. It could happen again, and HD4600 can already get near 7850K with OC.

 

[Idontcare]

And I am more and more inclined to believe the "node after revolution" that really brings exciting products for Intel. Remember 90nm and 45nm? They were kinda disappointing. The 65nm and 32nm were pretty awesome though.

Very true, and 14nm vs 22nm is looking likely to keep that "every other node" trend alive.

 

[witeken]

I went back and looked at the original article considering the excitement about GT4e -

Looks like to get GT4e, you'll be buying an i7. The lower two tiers are GT2 (HD 4600). i7 is really not in the same price segment as A10-7850K. It's priced to compete with something along the lines of a higher end i3.

4600 is a big improvement over 4000 for the i3 / i5, but it's non competitive with the current crop of APUs as far as GPU goes.

Now if they were doing GT3e in all i3 / i5 CPUs and GT4e in the i7s, then yes AMD and Nvidia would probably have some issues. But that's not what appears to be in the plan.

You can't compare a Gen7 HD4600 to a Gen9 GPU, certainly if you don't even know the amount of EUs.

 

[witeken]

They are at LEAST 2x now, so Atom has a room to improve without violating what Intel wants.

What is this ratio thing I've never heard of before?

It's 30% lower power at SAME performance.

Broadwell was actually slightly faster.

 

[Homeles]

What is this ratio thing I've never heard of before?

Intel's design teams used to, and possibly still holds to the mantra that an design change that results in a 1% increase in power requires at least a 2% increase in performance to be implemented.

 

[witeken]

Oh, that rule. I misread it and thought it was referring to something else.

 

[Ajay]

I think there's a couple of possibilities, and it's largely hinging on 3D. If Skylake has stacked memory and leverages it, it could be a decent bump. If not, we're looking at typical small gains from reinventing the 2D wheel. Maybe a bit bigger than Haswell in terms of perf/clock, but nothing too interesting.

However, we've already seen the performance figures for the 14nm process (up to 40%), which is actually better than 45 -> 32nm, where Intel claimed a "20-35% improvement over 45 nm high-k/MG logic transistors." So there is hope there. And I'm becoming increasingly confident that 10nm is going to be a huge node not only for Intel, but for everyone else, as it's the likely introduction of "post-silicon" channel materials.

Actually, my argument was that moving from planar to FinFET is a one time penalty for overclocking. 22nm -> 14nm should allow for traditional performance gains.

Here's a beautiful MS paint example:

I thought you also found a presentation by Kelin Kuhn showing an electrostatic limitation on FinFet that extended to 14nm visa vi frequency scaling. I think it may have been a problem with gate capacitance, but I'm not sure about that. I'm searching for it now, but maybe you can find it faster.

 

[Arachnotronic]

I thought you also found a presentation by Kelin Kuhn showing an electrostatic limitation on FinFet that extended to 14nm visa vi frequency scaling. I think it may have been a problem with gate capacitance, but I'm not sure about that. I'm searching for it now, but maybe you can find it faster.