Intel Skylake / Kaby Lake

Page 617 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

phillyman36

Golden Member
Jun 28, 2004
1,762
160
106
Question if someone wanted to run
2 Gpus( I know they would run at 8x8)
2 Nvme ssds
1 Sata ssd
1 optical drive
Would Kaby/Coffee Lake paired with a z270/z370 be able to handle that or would a Skylake X be better suited for the task?
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
You would think Intel would have given up on Skull Canyon due to it not selling well, especially with Raven Ridge coming.

Why would Raven Ridge prevent Intel from releasing their own NUC? If Raven Ridge is good then there's more incentive to release a high performance NUC. I really don't get your logic sometimes.

The reason Skull Canyon was a crappy seller is because it was barely faster than Iris Pro 6200. Part of the problem is that it only had 45W to play with versus 65W for Iris Pro 6200. Part of it was that it just sucked.

Some games, Iris Pro 580 in Skull Canyon really sucked: http://www.anandtech.com/show/10343/the-intel-skull-canyon-nuc6i7kyk-minipc-review/4

How do you go from 22nm to 14nm, and bump the EU up by 50%, with an architecture that's 2 generations ahead and end up being 30% faster in some cases? Something seriously went wrong with their GT4e chip because GT2/GT3e Gen 9 is pretty good.
 
Last edited:
  • Like
Reactions: Arachnotronic
Mar 10, 2006
11,715
2,012
126
Why would Raven Ridge prevent Intel from releasing their own NUC? If Raven Ridge is good then there's more incentive to release a high performance NUC. I really don't get your logic sometimes.

The reason Skull Canyon was a crappy seller is because it was barely faster than Iris Pro 6200. Part of the problem is that it only had 45W to play with versus 65W for Iris Pro 6200. Part of it was that it just sucked.

Some games, Iris Pro 580 in Skull Canyon really sucked: http://www.anandtech.com/show/10343/the-intel-skull-canyon-nuc6i7kyk-minipc-review/4

How do you go from 22nm to 14nm, and bump the EU up by 50%, with an architecture that's 2 generations ahead and end up being 30% faster in some cases? Something seriously went wrong with their GT4e chip because GT2/GT3e Gen 9 is pretty good.

I'd expect a high performance NUC to have a Coffee Lake 4+3e chip. 14nm++ and whatever physical design improvements could help such a product be better than the ill-fated SKL 4+4e part.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
I'd expect a high performance NUC to have a Coffee Lake 4+3e chip. 14nm++ and whatever physical design improvements could help such a product be better than the ill-fated SKL 4+4e part.

If you look at the Kabylake and Kabylake Refresh benchmarks, the part where the gains are nonexistent to miniscule are on the GPU side.

The 28W Iris 550 which is a Skylake GT3e part is 10-20% slower than the Iris Pro 6200. The CFL 4+3e chip would have to be 20-40% better just to be equal to the Skull Canyon part. Considering the 86% increased thermals improve performance over the 15W Iris 540 by only 5-10%, I don't count a 65W part closing that gap.

I can't see how a Coffeelake GT3e part would do any better. It's basically same architecture. If its going to be Q2 2018 I would also think it would be absolutely crazy to release a Coffeelake part. I mean, I would like them to stay in execution, not be worse than they are at now. Cannonlake GT3e at the minimum please.
 
Mar 10, 2006
11,715
2,012
126
If you look at the Kabylake and Kabylake Refresh benchmarks, the part where the gains are nonexistent to miniscule are on the GPU side.

The 28W Iris 550 which is a Skylake GT3e part is 10-20% slower than the Iris Pro 6200. The CFL 4+3e chip would have to be 20-40% better just to be equal to the Skull Canyon part. Considering the 86% increased thermals improve performance over the 15W Iris 540 by only 5-10%, I don't count a 65W part closing that gap.

I can't see how a Coffeelake GT3e part would do any better. It's basically same architecture. If its going to be Q2 2018 I would also think it would be absolutely crazy to release a Coffeelake part. I mean, I would like them to stay in execution, not be worse than they are at now. Cannonlake GT3e at the minimum please.

There is no Cannon Lake GT3e planned based on everything that's leaked. It's just the 2+2 U/Y part.

Intel-Gemini-Lake-1.jpg


I think if you want a GT3e on 10nm, you're going to need to wait for Ice Lake.
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
In older Roadmaps there is a Coffeelake based 4+3e for 15W/28W. Such a NUC with a GPU Gen9 tech from 2015 would be a shame in mid 2018, I can't see how this can be a success. It's a wasted effort. Intel should put more effort into Icelake and 10nm+. Also for Cannonlake ULT 2+2 in mid 2018, they should cancel it. There is no realistic chance against a better clocking 14nm+ with four cores. It's way too late.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
There is no Cannon Lake GT3e planned based on everything that's leaked. It's just the 2+2 U/Y part.

I think if you want a GT3e on 10nm, you're going to need to wait for Ice Lake.

Yea, yea I know the roadmaps. For Q2 release Coffeelake is extremely underwhelming. If it were to release in Jan, maybe we can cut them slack. Barely.

Before the release, the GT4e parts promised up to 50% gains over predecessor's GT3e. We know in GT2 vs GT2 comparisons though the Gen 9 is 15-30% better than Gen 8. Basically, what they are saying is if its not thermally bound, Gen to Gen comparison, GT4e is only 20% better than GT3e.

Also for Cannonlake ULT 2+2 in mid 2018, they should cancel it.

Cannonlake isn't mid 2018. We should see it by the holiday season, at the most maybe February.

Some rumors have it possibly delayed to next year. The way it could pan out is announcement at CES, with availability in the coming months. I figure unlike Desktops where you can buy the chip at release and have the system ready at most in 2 days, it takes time for Laptop guys to put it in new designs.
 
Last edited:

Bouowmx

Golden Member
Nov 13, 2016
1,138
550
146
Question if someone wanted to run
2 Gpus( I know they would run at 8x8)
2 Nvme ssds
1 Sata ssd
1 optical drive
Would Kaby/Coffee Lake paired with a z270/z370 be able to handle that or would a Skylake X be better suited for the task?

Depends on your speed requirements for the NVME drives. On Z[23]70, they are connected to chipset, which itself has 24 PCIE lanes, but is connected to CPU using DMI: PCIE 3.0 ×4. On Skylake-X, because CPU has more PCIE lanes, you have the option to connect NVME drives directly to CPU, to avoid DMI.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Depends on your speed requirements for the NVME drives. On Z[23]70, they are connected to chipset, which itself has 24 PCIE lanes, but is connected to CPU using DMI: PCIE 3.0 ×4. On Skylake-X, because CPU has more PCIE lanes, you have the option to connect NVME drives directly to CPU, to avoid DMI.

Considering the complexity of application and performance, I wonder how much of a difference there is in actual usage?

Transfer rates moving large files?
Responsiveness?
Boot time?
Loading time?

Due to reviewers focusing on mobile tech, we'll never see articles in detail like that again. Maybe we don't though. The difference might end up being 5%.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
Why would Raven Ridge prevent Intel from releasing their own NUC?

Because Intel can't compete because the IGP drivers are so bad. Especially now that Apple is dumping Iris for dGPUs there really isn't much point in continuing the Iris line altogether since AMD will be so far ahead. And you know AMD hates money so Raven Ridge will be cheaper too.
 
Mar 10, 2006
11,715
2,012
126
In older Roadmaps there is a Coffeelake based 4+3e for 15W/28W. Such a NUC with a GPU Gen9 tech from 2015 would be a shame in mid 2018, I can't see how this can be a success. It's a wasted effort. Intel should put more effort into Icelake and 10nm+. Also for Cannonlake ULT 2+2 in mid 2018, they should cancel it. There is no realistic chance against a better clocking 14nm+ with four cores. It's way too late.

Cannon Lake 2+2 is being released, IMO, as a pipe cleaner for the 10nm process. Intel will run a bunch of those wafers through the fabs mainly for yield learning ahead of the "real" 10nm product family launch -- Ice Lake.

Unlike with Cannon Lake, which is basically going to be a low-volume goof while the 14nm+/14nm++ products handle the bulk of the volume, Ice Lake needs to span the entire range of products from Y mobile all the way through desktop. Any learning Intel can do ahead of that Ice Lake ramp should be helpful.
 
Mar 10, 2006
11,715
2,012
126
Cannonlake isn't mid 2018. We should see it by the holiday season, at the most maybe February.

Some rumors have it possibly delayed to next year. The way it could pan out is announcement at CES, with availability in the coming months. I figure unlike Desktops where you can buy the chip at release and have the system ready at most in 2 days, it takes time for Laptop guys to put it in new designs.

Intel is being super vague about when we should see CNL. They have indicated that they'll announce CNL by year's end, but have said 1H 2018 volume ramp. We could very well not see systems until mid-2018 on this schedule.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Because Intel can't compete because the IGP drivers are so bad. Especially now that Apple is dumping Iris for dGPUs there really isn't much point in continuing the Iris line altogether since AMD will be so far ahead. And you know AMD hates money so Raven Ridge will be cheaper too.

Actually, their drivers are not that bad. Few years ago, yes. Looking at some driver overhead benchmarks, it looks like Intel might even have an advantage over the others in that department too(lower overhead means lower power for CPU).

AMD's performance will be bound in the same way as everyone else - memory bandwidth.
 
Last edited:
Mar 10, 2006
11,715
2,012
126
Because Intel can't compete because the IGP drivers are so bad. Especially now that Apple is dumping Iris for dGPUs there really isn't much point in continuing the Iris line altogether since AMD will be so far ahead. And you know AMD hates money so Raven Ridge will be cheaper too.

Apple isn't dumping Iris, they only dumped Iris Pro. Iris (2+3e) is used in the 13-inch MBP, which is a much higher volume design than the 15-inch MBP.
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
Cannonlake isn't mid 2018. We should see it by the holiday season, at the most maybe February.

I don't think so. My expectation is towards mid 2018.

Cannon Lake 2+2 is being released, IMO, as a pipe cleaner for the 10nm process. Intel will run a bunch of those wafers through the fabs mainly for yield learning ahead of the "real" 10nm product family launch -- Ice Lake.

Unlike with Cannon Lake, which is basically going to be a low-volume goof while the 14nm+/14nm++ products handle the bulk of the volume, Ice Lake needs to span the entire range of products from Y mobile all the way through desktop. Any learning Intel can do ahead of that Ice Lake ramp should be helpful.


That might be the only reason. Something like Broadwell, although Broadwell didn't face a Quadcore based Haswell ULV so it did make much more sense than Cannonlake next year against KBL-R.
 
Mar 10, 2006
11,715
2,012
126
I don't think so. My expectation is towards mid 2018.




That might be the only reason. Something like Broadwell, although Broadwell didn't face a Quadcore based Haswell ULV so it did make much more sense than Cannonlake next year against KBL-R.

That's true, Cannon Lake-U is basically going to have a hard time against KBL Refresh though for systems that can't support the peak power consumption of KBL Refresh, the CNL-U part could be a reasonable choice. Cannon Lake-Y is probably OK.

Intel really screwed up with 10nm, it's like the people who totally wrecked their mobile efforts are now running process tech development. I really don't understand why current TMG management is not out on their rear-ends.
 
Last edited:

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Intel is being super vague about when we should see CNL. They have indicated that they'll announce CNL by year's end, but have said 1H 2018 volume ramp. We could very well not see systems until mid-2018 on this schedule.

That sucks really, really bad. Let me explain.

If they release Cannonlake in mid-2018, there's zero chance we'll see Icelake before mid-2019.

Now that assumes they are still on the 12 months cycle. There's evidence we're past that.

June 2013: Haswell
Jul 2014: Haswell Refresh
August 2015: Skylake
Jan 2017: Kabylake
November 2017?: Coffeelake
Summer 2018: Cannonlake

Hmm. Yea we're not on the 12 month cycle. Based on that pattern,

Fall 2019: Icelake
Holiday 2021: Tigerlake
Spring 2022: 7nm

We are talking 3 year delay for their 10nm process. Basically, their claims of "3.5 year lead" doesn't exist even if you took their(likely marketing) claims at face value.

Also for Cannonlake ULT 2+2 in mid 2018, they should cancel it. There is no realistic chance against a better clocking 14nm+ with four cores. It's way too late.

Proper yield learning likely requires it to be field tested. The best way is to send it to HVM. If they cancel Cannonlake, then Icelake has to bear that burden. That's why when people claim they should "skip" a process, they have zero clue what they are talking about. If you are having a problem on a simpler process, you'll have worse times on a more advanced one.
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
If they release Cannonlake in mid-2018, there's zero chance we'll see Icelake before mid-2019.


I doubt that. If Icelake and 10nm+ is ready Intel is going to release Icelake as soon as possible. Same happened with Broadwell-->Skylake. Broadwell ULV was available in H1 2015 and Skylake ULV was available in H2 2015, less than a year. Kabylake-S and Coffeelake-S is another example, it's less than a year.
 
Mar 10, 2006
11,715
2,012
126
That sucks really, really bad. Let me explain.

If they release Cannonlake in mid-2018, there's zero chance we'll see Icelake before mid-2019.

Now that assumes they are still on the 12 months cycle. There's evidence we're past that.

June 2013: Haswell
Jul 2014: Haswell Refresh
August 2015: Skylake
Jan 2017: Kabylake
November 2017?: Coffeelake
Summer 2018: Cannonlake

Hmm. Yea we're not on the 12 month cycle. Based on that pattern,

Fall 2019: Icelake
Holiday 2021: Tigerlake
Spring 2022: 7nm

We are talking 3 year delay for their 10nm process. Basically, their claims of "3.5 year lead" doesn't exist even if you took their(likely marketing) claims at face value.



Proper yield learning likely requires it to be field tested. The best way is to send it to HVM. If they cancel Cannonlake, then Icelake has to bear that burden. That's why when people claim they should "skip" a process, they have zero clue what they are talking about. If you are having a problem on a simpler process, you'll have worse times on a more advanced one.


Here's what BK said on the most recent CC:

And we'll be shipping our first 10-nanometer products near the end of the year beginning with a lower volume SKU and followed by multiple SKUs and a volume ramp in the first half of 2018.

And...

One of the things we're seeing is each one of these technologies, and actually products that we generate, we're generating quite a few more SKUs, Vivek. And so you'll see a variety of SKUs progressing through 2018 as we ramp the 10-nanometer products, starting with the, I'll call it, more simpler SKUs at the beginning, going all the way through the high performance, high complexity SKUs towards the middle and back half of the year. Which is it's a traditional ramp like you see us push out on a new product ramp.

It really looks to me like the small-die CNL-Y/U aren't the only 10nm products planned for 2018.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
It really looks to me like the small-die CNL-Y/U aren't the only 10nm products planned for 2018.

I've seen that CC.

Didn't we establish that what he's talking about may in fact apply to Icelake?

I doubt that. If Icelake and 10nm+ is ready Intel is going to release Icelake as soon as possible.

I think in extreme cases, they can pull forward the launch, especially when its going well, or the project itself isn't much of a change. But the 12 months minimum rule applies because the ecosystem puts immense pressure on Intel to resist that. In fact, it prefers something slightly longer than 12 months. Many articles I have seen about Intel not pulling it forward because the industry doesn't want it. Designing a system takes time, and there's an optimal period for how much revenue it'll bring in. Too early, not everyone that can buy has bought it. Too late, the financial effect is the same because customers lose interest.

That's if Intel is not in trouble. With the supreme difficulty they are having on 10nm, I doubt on a faster cycle happening.

I do concede on one thing though. I said earlier than mid-2019 for Icelake is not possible. That's not true. Because Desktops are skipping Cannonlake. I apologize. :)
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
I think we are at the point where you almost have to believe that Intel's 10 nm yield is so terrible that it's not really fixable. The only real solution is EMIB and make the tiles as small as possible and/or keep large portions of actual products on 14 nm.
 
  • Like
Reactions: beginner99

ZGR

Platinum Member
Oct 26, 2012
2,052
656
136
Why would Raven Ridge prevent Intel from releasing their own NUC? If Raven Ridge is good then there's more incentive to release a high performance NUC. I really don't get your logic sometimes.

The reason Skull Canyon was a crappy seller is because it was barely faster than Iris Pro 6200. Part of the problem is that it only had 45W to play with versus 65W for Iris Pro 6200. Part of it was that it just sucked.

Some games, Iris Pro 580 in Skull Canyon really sucked: http://www.anandtech.com/show/10343/the-intel-skull-canyon-nuc6i7kyk-minipc-review/4

How do you go from 22nm to 14nm, and bump the EU up by 50%, with an architecture that's 2 generations ahead and end up being 30% faster in some cases? Something seriously went wrong with their GT4e chip because GT2/GT3e Gen 9 is pretty good.

I believe it was TDP limited, was it not?
My i7-5775c is only 65W out of the box. It is allowed to use as much power as it wants in the BIOS. It is basically a 125W CPU at 4.2 GHz.... That is why Iris Pro 6200 is faster. If Skull Canyon could have an unlocked TDP it would be faster.

I mean you say that. 45w vs 65w. It really was a silly move putting the top end Iris part in a 45w NUC and calling it an upgrade.

But this is why I am kinda pissed at Intel. Broadwell C is DDR3 based while Skull Canyon supports fast DDR4. If it had an unlocked TDP on a proper desktop platform, it would annihilate Broadwell-C!
 
Last edited:

7uly1

Junior Member
Sep 3, 2017
1
3
1
The scaling on that 4447cb looks legit and the scaling is really impressive.

Comparing to my 7820x score.

8 cores x 5.0GHz = 40 2249cb/40 = 56.2cb per core per 1GHz
18 cores x 4.5GHz = 81 4447cb/81 = 54.9cb per core per 1GHz
Pretty sure it is under water. His other screenshots show it idling at 18c...I'm pretty sure LN2 does not do that.
yes I just using ek wb now. no delid :)