Intel Skylake / Kaby Lake

Page 530 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

TheF34RChannel

Senior member
May 18, 2017
786
309
136
Well I hope you are right, because I am losing patience sometimes....



Well I have a good experience going from more cores to less with much higher frequency. This is it. If SKL-X supported ECC I would probably posting some results here already. Then Xeon can not be OCed that way. A shame.

The solver vs computer idle time- that is always the same. At the start of the project you are iterating to model the right "reality" conditions. At the end, no computer is fast enough :)

Hovewer I must say the computing power of last years is heavily based on individual and custom optimization which can do more than any HW.

Why not opt for Skylake-SP if you must upgrade? Your system does look the part right now.
 

Bouowmx

Golden Member
Nov 13, 2016
1,138
550
146
Last edited:

TahoeDust

Senior member
Nov 29, 2011
557
404
136
Has anyone done a real thorough review of a Gigabyte Gaming 7 board? I would like to see if their findings/thoughts are similar to mine.
 

VirtualLarry

No Lifer
Aug 25, 2001
56,227
9,990
126
I think I know why Coffee Lake has been (?) delayed. Too much Kaby Lake stock still left over. Nobody's buying anything but the G4560 and i7-7700K, for the most part. (Except a few OEMs offering i5-7400 rigs.)

(Edit: Above post is my opinion only.)
 
  • Like
Reactions: Drazick

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Skylake-SP Core: Designed for Data Center Workloads
Skylake-SP cache hierarchy architected specifically for Data Center use case

Skylake-SP cache latency:
https://www.servethehome.com/wp-con...-SP-Microarchitecture-L2-L3-Cache-Latency.jpg

Intel's testing shows the L3 cache latency increase is less than 10% when compared to Broadwell-EP.

These sub-tests from SpecCPU 2006 shows that except for very rare cases the new cache hierarchy affords significantly reduced miss rates for L2 cache without impacting it for the L3.

https://www.servethehome.com/wp-con...3-Cache-Inclusive-v-Non-Inclusive-SPECint.jpg
https://www.servethehome.com/wp-con...L3-Cache-Inclusive-v-Non-Inclusive-SPECfp.jpg

Clearly, the reason SKL-X isn't an ideal enthusiast CPU is because its a server-focused architecture. Marketing of course makes it unclear, but that's not news to us.
 
  • Like
Reactions: formulav8

jj109

Senior member
Dec 17, 2013
391
59
91
I'm fairly tickled by the fact that the extra L2 and 512b unit was just stapled right onto the base Skylake core.

servethehome has a 8180/8176 review up as well:

https://www.servethehome.com/quad-intel-xeon-platinum-8180-initial-benchmarks/

Quad socket inter-node latencies are lower than 140ns, how does that compare to previous generations?

Much better dual socket c-ray and NAMD results relative to dual E5-2699 v4 compared to Anandtech's results, but why though?
 
  • Like
Reactions: Sweepr

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
I'm fairly tickled by the fact that the extra L2 and 512b unit was just stapled right onto the base Skylake core.

It's not just "stapled". Because if you look at the cores, they are rectangular. Adding units would require the CPU floorplan to change to keep the rectangular shape. I'd like to at least acknowledge the engineering work required for that. The changes are actually significant.

Much better dual socket c-ray and NAMD results relative to dual E5-2699 v4 compared to Anandtech's results, but why though?

Tomshardware results are better too.

I see two reasons. Johan, personally favors AMD. Now, proper journalism dictates you be 100% fair. But as human beings, we aren't 100% neutral. We do lean based on our inclinations and past experiences. So he does say in his review he uses applications and OS settings that would seem "typical" to an enterprise user. You need to make sure you aren't biased in any way. You can't favor AMD because Intel is the monopoly and AMD was shown in a worse light by others. It needs to be fair for both.

My analysis is that he made a wrong conclusion. You cannot, as a review consisting of often less than a dozen benchmark out of thousands, or tens of thousands usage cases perfectly represent the server market. Whether the server adminstrator decides to do due diligence to have it performing it as much or leave it as it came is entirely up to them. As a reviewer you should use the best case scenario for each platform because of this. The administrators would try their best in reaching maximum performance allowed within their budget and time.

I've seen another person pointing out that by not using the latest distro in Johan's review it underrates EPYC by not taking advantage of the IF differences. The only way to quash most of those claims is to show EPYC AND Xeon Scalable in the best light.

Another reason compounds the first problem. Actually the two feeds into each other. He says he didn't have enough time. Yes, I do feel they raised the price of Skylake-SP too much. Yes I don't think its entirely right that you need Platinum to get max 2P performance. Yes, I don't like the branding and segmentation. The previous branding was logical. But do I feel the way Johan made his review could have been done better? Yes, definitely.
 
  • Like
Reactions: JoeRambo and Sweepr

JoeRambo

Golden Member
Jun 13, 2013
1,814
2,105
136
Clearly, the reason SKL-X isn't an ideal enthusiast CPU is because its a server-focused architecture. Marketing of course makes it unclear, but that's not news to us.

Actually not for every server either. It caters to 2 markets:

1) A real cloud market, where Amazon/MS etc requested large L2 sizes to keep things local to core, since obviously once you partition a server, you don't want a single customer workload to chew L3 and MC.
2) Imaginary HPC market that forced AVX512 onto core. While it looks nice on marketing department TFlops chart, and easy to compare with competitors it does inflict insane power penalty ( and it can be argued performance penalty, by virtue of eating area/power budged that could be used for cores and/or clock).

I will concede that Mesh had to go in, since ring could not scale, but Intel ruined what was a proper low/mid tier server CPUs and released cloud oriented CPU. Thankfully for customers, at this critical junction in time, Ryzen has arrived to take that sweet 1-2S market from Intel. And Intel's marketing department is working hard to help them with insane market segmentation and pricing.

P.S. i am of course biased, since we only used to buy E5-26N0 vX for my company workloads.
 
  • Like
Reactions: Ajay

wahdangun

Golden Member
Feb 3, 2011
1,007
148
106
It's not just "stapled". Because if you look at the cores, they are rectangular. Adding units would require the CPU floorplan to change to keep the rectangular shape. I'd like to at least acknowledge the engineering work required for that. The changes are actually significant.



Tomshardware results are better too.

I see two reasons. Johan, personally favors AMD. Now, proper journalism dictates you be 100% fair. But as human beings, we aren't 100% neutral. We do lean based on our inclinations and past experiences. So he does say in his review he uses applications and OS settings that would seem "typical" to an enterprise user. You need to make sure you aren't biased in any way. You can't favor AMD because Intel is the monopoly and AMD was shown in a worse light by others. It needs to be fair for both.

My analysis is that he made a wrong conclusion. You cannot, as a review consisting of often less than a dozen benchmark out of thousands, or tens of thousands usage cases perfectly represent the server market. Whether the server adminstrator decides to do due diligence to have it performing it as much or leave it as it came is entirely up to them. As a reviewer you should use the best case scenario for each platform because of this. The administrators would try their best in reaching maximum performance allowed within their budget and time.

I've seen another person pointing out that by not using the latest distro in Johan's review it underrates EPYC by not taking advantage of the IF differences. The only way to quash most of those claims is to show EPYC AND Xeon Scalable in the best light.

Another reason compounds the first problem. Actually the two feeds into each other. He says he didn't have enough time. Yes, I do feel they raised the price of Skylake-SP too much. Yes I don't think its entirely right that you need Platinum to get max 2P performance. Yes, I don't like the branding and segmentation. The previous branding was logical. But do I feel the way Johan made his review could have been done better? Yes, definitely.

thats why it need more time to review server cpu because the difficulties to get the maximum performance, especially since the server market use fully customized software to extract every inch of performance they can get.

so sometimes using standard build will be not representative of actual capability the software have.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
1) A real cloud market, where Amazon/MS etc requested large L2 sizes to keep things local to core, since obviously once you partition a server, you don't want a single customer workload to chew L3 and MC.

The cloud is fast becoming the majority of the server market if it isn't already. So we shouldn't be surprised that Intel is designing processors for them. The non-cloud market is in decline. Doesn't mean Intel hasn't made a mess of things with this insane segmentation.

2) Imaginary HPC market that forced AVX512 onto core. While it looks nice on marketing department TFlops chart, and easy to compare with competitors it does inflict insane power penalty ( and it can be argued performance penalty, by virtue of eating area/power budged that could be used for cores and/or clock).

Yeah it's a problem. Intel did manage to get three systems in the Top500 that use Xeon Gold/Plat so it's not going to waste but getting good use out of it outside of that market is going to be tough.
 

itsmydamnation

Platinum Member
Feb 6, 2011
2,744
3,077
136
I'm fairly tickled by the fact that the extra L2 and 512b unit was just stapled right onto the base Skylake core.

servethehome has a 8180/8176 review up as well:

https://www.servethehome.com/quad-intel-xeon-platinum-8180-initial-benchmarks/

Quad socket inter-node latencies are lower than 140ns, how does that compare to previous generations?

Much better dual socket c-ray and NAMD results relative to dual E5-2699 v4 compared to Anandtech's results, but why though?

It's not just "stapled". Because if you look at the cores, they are rectangular. Adding units would require the CPU floorplan to change to keep the rectangular shape. I'd like to at least acknowledge the engineering work required for that. The changes are actually significant.



Tomshardware results are better too.

I see two reasons. Johan, personally favors AMD. Now, proper journalism dictates you be 100% fair. But as human beings, we aren't 100% neutral. We do lean based on our inclinations and past experiences. So he does say in his review he uses applications and OS settings that would seem "typical" to an enterprise user. You need to make sure you aren't biased in any way. You can't favor AMD because Intel is the monopoly and AMD was shown in a worse light by others. It needs to be fair for both.

My analysis is that he made a wrong conclusion. You cannot, as a review consisting of often less than a dozen benchmark out of thousands, or tens of thousands usage cases perfectly represent the server market. Whether the server adminstrator decides to do due diligence to have it performing it as much or leave it as it came is entirely up to them. As a reviewer you should use the best case scenario for each platform because of this. The administrators would try their best in reaching maximum performance allowed within their budget and time.

I've seen another person pointing out that by not using the latest distro in Johan's review it underrates EPYC by not taking advantage of the IF differences. The only way to quash most of those claims is to show EPYC AND Xeon Scalable in the best light.

Another reason compounds the first problem. Actually the two feeds into each other. He says he didn't have enough time. Yes, I do feel they raised the price of Skylake-SP too much. Yes I don't think its entirely right that you need Platinum to get max 2P performance. Yes, I don't like the branding and segmentation. The previous branding was logical. But do I feel the way Johan made his review could have been done better? Yes, definitely.

I did a quick and dirty and i get this:

server%20comparision.png


How is c-ray different? outside of NAMD which you dont know if there setup is v4 performing well/poor or v5 performing well/poor or a combination of both.

Also they kind of showed Xeon in one of its best light ( DB ) if you wanted to show EPYC in its best light, then it would be VM host with guests in the 2 to 4 vcpu range. The memory bandwidth + more core + more SMT perf per core would really show EPYC strong suite.
 

StefanR5R

Elite Member
Dec 10, 2016
5,459
7,718
136
The administrators would try their best in reaching maximum performance allowed within their budget and time.

thats why it need more time to review server cpu because the difficulties to get the maximum performance, especially since the server market use fully customized software to extract every inch of performance they can get.
That's true for a part of the market. Other customers do not have the means or the rights to tune the software they use for best utilization of the hardware they have. Anyway, both types of hardware purchasers obviously evaluate (to various extent) which kind of hardware fits "their" software and their workloads best.

Actually not for every server either. It caters to 2 markets:

1) A real cloud market, where Amazon/MS etc requested large L2 sizes to keep things local to core, since obviously once you partition a server, you don't want a single customer workload to chew L3 and MC.
2) Imaginary HPC market that forced AVX512 onto core.
Such HPC use cases may be broader and more divers than one would initially think. E.g. financial services have been mentioned among them.
While it looks nice on marketing department TFlops chart, and easy to compare with competitors it does inflict insane power penalty ( and it can be argued performance penalty, by virtue of eating area/power budged that could be used for cores and/or clock).
There is an area penalty, but less so a power penalty. If the units aren't used, they should mostly function as black silicon. If they are used (by properly tuned software in workloads which can actually benefit), not just throughput but also performance per Watt should increase compared to AVX-2 or classic FP code.

On Haswell-EP, running an AVX workload on at least one core downclocked all cores. But since Broadwell-EP, only the cores which actually run AVX workload clock down; the other cores continue to run at non-AVX base clock or non-AVX turbo. I do not know though to which extent the workload on some cores influences the ability of other cores to clock up from base to turbo. (The Broadwell-EP boxes which I have, with Supermicro workstation-class frequency-optimized DP mainboard, always run all-core non-AVX turbo and/or all-core AVX turbo when many cores are loaded. I have never seen these particular boxes run at base clocks; it's either few-core turbo, all-core-turbo, or idle clock. IOW in my case at least, AVX on some cores does not prevent turbo on other cores. But yes, I can almost determine by looking at a Watt meter whether or not these boxes are runing AVX code at a given moment. Meaning, classic FP code is unable to fully utilize the package power budget of these processors.)
 

TheF34RChannel

Senior member
May 18, 2017
786
309
136
Last edited:

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
I did a quick and dirty and i get this:

How is c-ray different? outside of NAMD which you dont know if there setup is v4 performing well/poor or v5 performing well/poor or a combination of both..

Thanks for your graph and comments.

From comparing the two sites, Anandtech's results use the "Medium" rather than the "Hard" setting which you are comparing to. STH just shows time elapsed, while Anandtech shows jobs per hour. You'll see what I am talking about.

Medium
STH 8176- 7.5
STH 2699- 12.5

Anandtech
8176- 340
2699- 263

With pixel counting I got 9 seconds for 8176 and 12.96 seconds for 2699. If you convert that to jobs per hour you get

8176- 400
2699- 278

The 2699 numbers at Medium settings roughly correspond to Anandtech numbers. The 8176 is 44% faster according to STH numbers and 29% faster according to Anandtech's.

On NAMD the only other result is Tomshardware which shows 45% advantage for Platinum 8176 over E5 2697 v4. In comparison, Anandtech's Xeon E5 v4 results show the slightly slower E5 2695 v4 gets 3.4. The previous generation 2697 v3 gets 3.6. The 10% advantage of going from 18 v3 to 22 v4 cores suggest that its quite memory bandwidth bound in that benchmark. Interestingly, the memory bandwidth advantage for the v4 over v3 is 15%.

With AVX enabled, Skylake-SP has 35% advantage over Broadwell-EP bandwidth. Based on that it should show greater gain than 5%. The advantages with the AVX-enabled one is even less at 2%. It might as well be equal.
 
Last edited:
Aug 11, 2008
10,451
642
126
Hmm, I think Intel may set the clocks on 6C/12T below the i7 7800X to preserve it's place is the market :neutral:
It it overclocks to 4.8 or so, then stock clocks dont really matter much to the enthusiast. If it doesnt overclock well, they have shot themselves in the foot again, after rushing out Skylake X with a less efficient cache, TIM instead of solder, and immature motherboards. Just goes to show that they should have increased cores on the mainstream a long time ago, instead of forcing enthusiasts onto the HEDT platform. Even worse, now that the HEDT platform is close to mainstream in process and architecture, instead of 2 generations behind like it was, they come out with a server oriented cache that is not optimized for consumers. I am still hoping CF hits it out of the park, but am becoming less optimistic after seeing Skylake X.

Edit: like Ajay, I hope they dont gimp CF so that it wont compete well against the lower end of Skylake X.
 
  • Like
Reactions: TheF34RChannel

beginner99

Diamond Member
Jun 2, 2009
5,208
1,580
136
Hmm, I think Intel may set the clocks on 6C/12T below the i7 7800X to preserve it's place is the market :neutral:

As long as they do the same with price, I'm good. ;) If it can't OC, then I won't buy and if it can, base clock hardly matters.

My bigger fear is they hike up the price to 7800x levels or above. It will certainly cost more than a 7700k.
 

Qwertilot

Golden Member
Nov 28, 2013
1,604
257
126
They'll 99% do a few of them :) One clocked as high (er!) than can be sanely stood, one clocked to fit into 65w & one for 35-45w.

Those last two obviously maybe higher if 6 cores == a bigger TDP window. I'm guessing that that SKU would expect to fit into ~65w?
 

VirtualLarry

No Lifer
Aug 25, 2001
56,227
9,990
126
It will certainly cost more than a 7700k.
That's what really gives me pause about Coffee Lake 6C/12T. I'm guessing clocks will be above 4.5Ghz when OCed, so a 6C CFL should compete well in MT with an 8C Ryzen CPU (1700), but if they price it higher than $320 or so, I'm out. Ryzen just costs less and offers so much value. (Ryzen 7 1700 was recently selling for $270 shipped. Now THAT's unbeatable value.)
 

Lodix

Senior member
Jun 24, 2016
340
116
116
Considering they are increasing transistors increasing the cores and the new 14nm++ is less dense ( bigger dies for the same chip so more wafers ) it all points outs to be more expensive considering Intel likes high profits. And for sure it will clock higher/better than the 7800k.
 

LTC8K6

Lifer
Mar 10, 2004
28,520
1,575
126
Hmm, I think Intel may set the clocks on 6C/12T below the i7 7800X to preserve it's place is the market :neutral:
They did not set the 7740X clock above the 7700K clock...unless you want to count the measly 100mhz of turbo.

And the 7700K holds it's own pretty well against the 7800X in lots of areas and of course defeats it in gaming

Intel seems to be into market segmentation, so they will probably not worry about desktop vs HEDT, imo.

If the new 6C CFL chip has a base below 3.5ghz, it will be stillborn. I won't even give it a second look.