Intel Skylake / Kaby Lake

Page 513 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Edrick

Golden Member
Feb 18, 2010
1,939
230
106
July 11. 9+ months after the first shipments (Google, etc.).

Considering you seem to know the most about these, can you divulge any information on how Intel will segment these (feature wise) across the spectrum (bronze-platinum)? For example, will Intel gimp on the AVX-512 pipes on any of these the same way they did with some Skylake-X cpus?
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Considering you seem to know the most about these, can you divulge any information on how Intel will segment these (feature wise) across the spectrum (bronze-platinum)? For example, will Intel gimp on the AVX-512 pipes on any of these the same way they did with some Skylake-X cpus?

Someone said earlier the reason some of the Xeon *Color* chips have low TDP and very high clocks and others high TDP and lower clocks is because of one having AVX-512 and one not having one.

That makes perfect sense. When you are running a workload that really takes advantage of AVX, majority of the core power consumption is taken up by the vector unit. If you double that with AVX-512 the proportion increases further.

It doesn't take much sense in the consumer sector though. I'd like to see Platform 2015 vision(talked about in IDF about a decade ago) come true.

You can put 10 or so of the Atom cores used in Xeon Phi in roughly the same space taken up by the GT2 iGPU in Skylake and Kabylake. Running at 1GHz, they'd offer 320 DP GFlops or 640 SP GFlops. That's better than the 384 SP GFlops offered by the GT2 iGPU and it'll be easier to use it because its a full Atom core! Maybe then they can add graphics-related units and we'll see Larrabee-iGPU finally come to a reality. Then you can just stick with AVX2(or AVX-512 but with half the flops just for compatibility sake) on the CPU core.

The CPU and the GPU guys talk about convergence but it seems to get forever delayed. Perhaps this is telling us such ideal visions don't work.
 
Last edited:

beginner99

Diamond Member
Jun 2, 2009
5,208
1,580
136
You can put 10 or so of the Atom cores used in Xeon Phi in roughly the same space taken up by the GT2 iGPU in Skylake and Kabylake. Running at 1GHz, they'd offer 320 DP GFlops or 640 SP GFlops

I always wonder why not create a quad-core + 10 of these Atom cores in 1 chip? Too difficult for thread-scheduling?
 

vinhom

Member
Oct 14, 2016
25
5
81
I'm a bit confused regarding the Intel roadmap for mobile, are Kaby Lake U 4+2 and 4+3e 15W still coming?
 

Sweepr

Diamond Member
May 12, 2006
5,148
1,142
131
I mean, that is what gets me. Aren't these the same dies? 18W 4+2 gets cancelled but 15W 4+2 is still coming?

Launching in Q3 (BTS) is Kaby Lake Refresh, Intel's first mobile 14nm++ product, a 4C/8T CPU at 15W TDP. The 4+3e SKU is Coffee Lake-U (15W and 28W versions), but this one comes in 2018.
 
  • Like
Reactions: vinhom

tech960

Member
Sep 17, 2006
76
2
71
I don't think so, these chips sell for so much that saving a few bucks per unit in exchange for tons of bad press (which will probably turn some enthusiasts off from the parts) would be...well, it'd be penny wise, pound foolish, IMHO.

There's got to be something else behind it. I wish we knew what.

But it's not pennies if you're selling 100M units. That cost savings is 25M if you factor the solder to be $.25. It adds up. Especially if you make a 3, 5 or even 10 year metric.
 

Topweasel

Diamond Member
Oct 19, 2000
5,436
1,654
136
But it's not pennies if you're selling 100M units. That cost savings is 25M if you factor the solder to be $.25. It adds up. Especially if you make a 3, 5 or even 10 year metric.

On the other hand these are $600-$2k CPU's that are really $600-$4K CPU's. These are made from $80-$160 worth of sand and tech. That might be $25 million in savings. But these are the guys that make Intel the most profit margins and the 25 million they are saving is a drop in the bucket to the multibillion that they are jeopardizing with bad press. The Time to do this was with Haswell or Broadwell so no one would think twice about alternatives and it would be old news before Ryzen came out. They picked the wrong time to make a polarizing change like this.

That said I am not sure the worry about this is all it's cracked up to be. Specially the higher the core count.
 

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
But it's not pennies if you're selling 100M units. That cost savings is 25M if you factor the solder to be $.25. It adds up. Especially if you make a 3, 5 or even 10 year metric.
100M units at $600 is $60 Billion.

Saying that saving $25M on $60B in sales is significant when you seriously compromise the product is very, very shortsighted. That is around 0.04%. Are Intel's margins so low that they need this?
 

lolfail9001

Golden Member
Sep 9, 2016
1,056
353
96
100M units at $600 is $60 Billion.
Now consider for a second how many people will seriously not purchase a given CPU because it did not have solder. Seriously. I would be very surprised if it added up to even 100000 total.

In the end, even with bad press and all, it's clearly the profitable move for Intel.
 
Aug 11, 2008
10,451
642
126
100M units at $600 is $60 Billion.

Saying that saving $25M on $60B in sales is significant when you seriously compromise the product is very, very shortsighted. That is around 0.04%. Are Intel's margins so low that they need this?
How many HEDT units do they sell though? It certainly isnt 60 billion in sales, unless you count servers as well. That said, it certainly seems unwise not to solder, considering it both alienates loyal intel supporters and is like throwing gasoline on a fire to intel haters. My feeling is that the savings is more than just the cost of solder vs tim, and comes from simplified production.
 
Mar 10, 2006
11,715
2,012
126
How many HEDT units do they sell though? It certainly isnt 60 billion in sales, unless you count servers as well. That said, it certainly seems unwise not to solder, considering it both alienates loyal intel supporters and is like throwing gasoline on a fire to intel haters. My feeling is that the savings is more than just the cost of solder vs tim, and comes from simplified production.

Intel uses TIM for lifetime reliability concerns, AFIAK, not for cost reasons.

The bad PR and the potential lost sales due to this would not be made up for in saving a couple of bucks in IC packaging, especially since these are very high value parts not Atom/Pentium/Celeron chips.

Also, $60B is not even close to the right value. Most of Intel's CPU volumes are into mobile/embedded where IHS isn't even used. In mainstream desktop, which is a large part of the business, lifetime reliability/warranty concerns are much more important than overclocking capability.

In workstations/data centers, where component failures are bad, bad, bad, you want to do everything you can to maximize such reliability.
 
Last edited:
  • Like
Reactions: Conroe

scannall

Golden Member
Jan 1, 2012
1,944
1,638
136
Intel uses TIM for lifetime reliability concerns, AFIAK, not for cost reasons.

I really don't buy that though. Doing maintenance, repairs and building computers for 40 years now. And I have yet to see a single failure due to solder. As *speculation*, soldering may damage more dies in assembly/packaging? A buck worth of solder wouldn't make a difference. But losing *more speculation* %10 of your very valuable dies in assembly would be a whole lot more money than a squirt of solder. Perhaps someone here knows more about the assembly/packaging part than I do.
 

arandomguy

Senior member
Sep 3, 2013
556
183
116
If the future now seems to be multi die packages, specifically in Intel's case with EMIB, what are the possible challenges and reliability issues with solder interfaces?

What are challenges with modern chips which can have large varying surface temperature spots? An example would be the half with the IGP at idle and the CPU half at maximum load. Does that type of thermal stress affect the viability of solder?

Could there actually be a longer term move to move away from using an IHS on packaging all together?

Are there back channel or other logistics issues the public isn't aware of? Especially as Intel's volumes are quite a bit different. What other companies are using solder versus TIM? Does Intel use solder for Xeon Phi? AMD has only switched back to solder for Ryzen currently (any confirmation for Threadripper or Epyc?). AMD GPUs have been direct die. Nvidia did not use solder on their Fermi IHS, since then all GPUs have switched to direct die. IBM's Power9?
 

LTC8K6

Lifer
Mar 10, 2004
28,520
1,575
126
The number of people who even know what "Intel doesn't solder the IHS." means is tiny. Intel does not really need to solder any IHS ever.
The number of people who care is insignificant.

Intel only needs to manufacture chips that perform as advertised and compare favorably with any competition.

When they do that, the chips sell well. When they don't, they don't.

If that requires soldering, then Intel will solder. If it doesn't, then they won't.

That's it.

Either buy the chips or don't.

Quit telling Intel (or AMD) how to make their chips.
 

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
Now consider for a second how many people will seriously not purchase a given CPU because it did not have solder. Seriously. I would be very surprised if it added up to even 100000 total.

In the end, even with bad press and all, it's clearly the profitable move for Intel.

It's rather obvious I was replying to a post where certain assumptions were made. Namely 100M units sold and a total of $25M in savings, and using those values to justify the TIM decision. My reply was simply to put that $ value saved into perspective.

Now I see you arbitrarily assuming sale quantities that make the case for Intel's choice of TIM seem reasonable. Who are you really?

Also, a few others seem to be stuck in the thought train that it MUST be some other reason, even though no other one is apparent, because to admit Intel doing this for economic reasons, is too much of a shock.

Intel has great engineers, but I've seen many dominant successful organizations hijacked by more political short sighted ambitious players as they rode that success wave. Any mediocre leader can appear to be great when you have such a lead. Sports, the military, businesses, etc, all have examples of this.

This, in my view, is Intel's true battle.
 
Last edited:
  • Like
Reactions: coercitiv

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
Especially since the "bad press" is limited to some obscure tech forums and not general media. Most people buying these servers won't even know about this.
You are making the assumption that the TIM is the issue.

The primary issue is the high temps. Talking about the TIM is to try and understand why this is happening.

Trust me, the wider public will see the higher temps and have reservations.
 

lolfail9001

Golden Member
Sep 9, 2016
1,056
353
96
The primary issue is the high temps.
Evidently it is not an issue at all since Intel used TIM on their highest TDP chips for last 3 or something years. I guess public really does not care about temperatures... not too surprising either.
It's rather obvious I was replying to a post where certain assumptions were made.
Yes, it is, but while you're correct that $25M is not very significant in the face of billions of profits, it is still 25 millions dollars (in that particular numbers case).
Now I see you arbitrarily assuming sale quantities that make the case for Intel's choice of TIM seem reasonable. Who are you really?
Do i? I just wonder how many people actually care.
Also, a few others seem to be stuck in the thought train that it MUST be some other reason
Any other reason ends up tying into economics, they may be just as right.
 

beginner99

Diamond Member
Jun 2, 2009
5,208
1,580
136
You are making the assumption that the TIM is the issue.

I'm making the assumption the process variation in applying the TIM and the IHS is the issue. The TIM itself is actually pretty good but if you use too much and/or have gap to the IHS because you applied too much glue, then temps will suck. Just removing the too much glue and fixing contact with IHS helped a lot back then with ivy bridge. It's beyond me how they can manage to create billions of tiny transistors with tiny fins and then fuck up the application of glue.