Intel Skylake / Kaby Lake

[Dasa2]

Skylake on the other hand seems to show much bigger differences.

I do agree though the tests should be re-done with modern games and a 980ti.

sandybridge for me was seeing up to 15% performance improvement in moving to faster ram thats not all that different to what skylake is seeing from different ram speeds and i think haswell was much the same from the few tests i have seen that were not gpu bottlenecked

980ti can still bottlneck a 6700k in a lot of new games even a pentium in a few of the game tests anandtech uses
its not just the games they use but the sections of the games they use

a lot of reviews use sections of a game thats on rails where everything happens exactly the same every time which is great for consistency but it can significantly reduce the load on the cpu vs other sections of a game

 

[steve wilson]

The mounting holes are the same as 1150. Yes its fits. Im going to use a NH-U14s with my Asus z170 Deluxe.(just waiting to find a cpu)

http://www.noctua.at/main.php?show=compatibility_gen&products_id=64&lng=en#LGA1151_Asus

Awesome, thanks for the info and the link.

 

[Pwndenburg]

Still a bit confused on one thing. Let's say you left it at stock and bought 3000 memory. Would the memory actually operate at that frequency or just not because the core is not overclocked. I mainly ask b/c I am running a locked Haswell i7 4770 with 1866 mhz ram. The board takes it and all programs correctly identify my speed settings, but is it actually running at that speed?

 

[BFG10K]

Not sure if this was posted: http://pc.watch.impress.co.jp/docs/column/sebuncha/20150806_715335.html

The delidding shows a really small die, and thermal paste that yet again isn't optimal.

 

[witeken]

There's no presentation about Skylake at Hotchips, although there are a few interesting ones on day 2 (from Intel Cherry Trail and Knights Corner).

http://www.hotchips.org/

 

[ShintaiDK]

What I don't get is...if Skylake scales so well with memory speeds, why did Intel validate it for such (relatively) slow memory?

For the non-k skus, aren't people going to be SOL, or can non-k skus still oc the men controller?

Its already answered previously in the thread.

You cant validate something that runs out of spec or doesnt exist when validation of the chip begins. (Think 12months+ ago)

And I dont think DDR4 reached more than 2400Mhz today within the specs.

 

[steve wilson]

Not sure if this was posted: http://pc.watch.impress.co.jp/docs/column/sebuncha/20150806_715335.html

The delidding shows a really small die, and thermal paste that yet again isn't optimal.

How difficult is Delidding? Are there special tools you need and is the Cool laboratory liquid pro expensive?

 

[steve wilson]

Still a bit confused on one thing. Let's say you left it at stock and bought 3000 memory. Would the memory actually operate at that frequency or just not because the core is not overclocked. I mainly ask b/c I am running a locked Haswell i7 4770 with 1866 mhz ram. The board takes it and all programs correctly identify my speed settings, but is it actually running at that speed?

I should think it is running at the desired speed. You could use something like CPUZ to make sure in windows.

 

[BFG10K]

How difficult is Delidding? Are there special tools you need and is the Cool laboratory liquid pro expensive?

You basically need a vise and something to lift the IHS, like a razor. Some people use wooden mallets to apply force from momentum. See the article for how they did it.

But generally speaking if you have to ask, you shouldn't be doing it. I'm not trying to patronize anyone with that statement because I myself stay well away from things like that. But I'm always interested in the results.

 

[steve wilson]

You basically need a vise and something to lift the IHS, like a razor. Some people use wooden mallets to apply force from momentum. See the article for how they did it.

But generally speaking if you have to ask, you shouldn't be doing it. I'm not trying to patronize anyone with that statement because I myself stay well away from things like that. But I'm always interested in the results.

It does look a bit risky, but I believe I could learn to do it. I've watched a few videos on it, but would be nice to hear from someone with experience.

 

[podspi]

Its already answered previously in the thread.

You cant validate something that runs out of spec or doesnt exist when validation of the chip begins. (Think 12months+ ago)

And I dont think DDR4 reached more than 2400Mhz today within the specs.

I think that is a pretty poor excuse - Intel has partnerships with large memory manufacturers, I'm skeptical that they couldn't get it done.

Intel has always been pretty conservative when it comes to memory speed validation (if memory serves me correctly) but I don't remember it having as large of an impact in recent years.

I find it hard to fault AT for using the fastest validated memory - Non-OCed tests are non-OCed tests. Fact is, running the processor in spec - all of it in spec- results in disappointing (imho) performance. It'll be interesting to see whether we will be able to use faster memory in non-k skus. Something tells me we won't - and that'll make the issue much more important for most of the processors that'll be out there.

 

[Sweepr]

We can look at the Bioshock 720p framerates and notice some nice scaling, but we can also look at the Tomb Raider results and see no scaling whatsoever. We have a healthy jump in memory performance (2133->3200) and the FPS meter doesn't even budge. Could it be this situation is similar to the one in the Anandtech's Haswell DDR3 scaling review?

Haswell with 780Ti does 133FPS in Tombraider @720p, meanwhile in the memory scaling review it does 47FPS on single dGPU and 90FPS in the Crossfire test. Now imagine if we ran the test again on Nvidia's or AMD's latest.

And there's also tests where it shows impressive scaling behaviour.

I'm not sure what happened in AnandTech's review but something went very wrong there. Multiple websites made it clear that Skylake can deliver >10% better performance per clock in CPU limited games while AnandTech is selling the lie that Haswell is a better gaming chip.

There’s no easy way to write this.

Discrete graphics card performance decreases on Skylake over Haswell.

We need some proper DDR3/DDR4 testing using the latest dGPUs from the same website comparing both chips, for now I think Skylake scales better with 2000MHz+ speeds.

 

[BFG10K]

I find it hard to fault AT for using the fastest validated memory - Non-OCed tests are non-OCed tests. Fact is, running the processor in spec - all of it in spec- results in disappointing (imho) performance.

I agree completely. The fact is, in order for Skylake to actually beat a 4790K at stock speeds, the former needs to be running memory speeds far outside spec, memory that's even more expensive than the already expensive stock versions.

On top of that you need a whole new motherboard which adds nothing of significance to older chipsets, and causes issues with installing Windows 7 off USB sticks. Considering Windows 7 is the massively dominant platform, that's a problem.

People are desperately trying to see performance that simply isn't there, and an upgrade that simply isn't there.

If I was buying a brand new system right now, I'd still get a 4790K. DDR3 memory is dirt cheap compared to DDR4, and even the slowest kits cause virtually no performance loss.

 

[Sweepr]

Not true. We have reviews using DDR4-2133 showing better gains than AnandTech too. And most reviews I quote where Skylake is beating Haswell by >10% per clock are using DDR4-2666, which is not a lot more expensive than (crap) DDR4-2133.

Example 1 - PC Perspective (using DDR4-2133)

Example 2 - Hardware Canucks (DDR4-2666)

Example 3 - Eurogamer (DDR4-2666)

Core i5 6600K vs Core i5 4690K (same 3.5-3.9GHz base/turbo)
- 17% faster @ The Witcher 3
- 11% faster @ GTA V
- 10% faster @ Battlefield 4

Example 4 - PCLab (DDR4-2666)

Core i5 6600K vs Core i5 4690K at fixed 4.5GHz, Skylake-S is:
16.2% faster @ Battlefield 4 MP
10.9% faster @ Counter Strife Global Offensive
6% faster @ Crysis 3
22.7% faster @ Far Cry 4
13% faster @ GTA V
12.5% faster @ The Witcher 3
9.2% faster @ Watch Dogs
14.7% faster @ Project Cars
14.8% faster @ Starcraft 2
24.5% faster @ Total War Attila

I agree that Core i7 4790K combos might be the better deal right now, but Skylake's price will come down too.

 

[coercitiv]

I find it hard to fault AT for using the fastest validated memory - Non-OCed tests are non-OCed tests.

Original Haswell review. (Ananad Lal Shimpi)

I used DDR3-2400 for most of my testing

Haswell refresh review. (Ian Cutress)

Memory Settings 1600 9-11-9-27 1T tRFC 240

Skylake review.(Ian Cutress)

G.Skill DDR3-1866 4x4
*Memory Timings used were the supported frequencies of each architecture, except DDR3L vs DDR4 testing, which used DDR3-1866 C9.

Earlier on this thread there were some comments regarding Haswell having and advantage by using DDR3 1866 vs DDR4 2133. Turns out his is not the case, the note in the review implies Haswell was used with DDR3 1600. At CL 9 it would still have a slight latency advantage, while at CL11 it would be on par with DDR3 latency wise.

 

[Sweepr]

True latency (ns) = clock cycle time (ns) x number of clock cycles (CL)

Lower is better:
Haswell DDR3 1600 CL9 = 1.25 ns x 9 CL = 11.25 ns true latency
Skylake DDR4 2133 CL15 = 0.94 ns x 15 CL = 14.06 ns true latency

There is no plausible reason not to use faster RAM (for both chips), given that they have used 'out of spec' memory in their past reviews. These are K chips, you don't see many people using DDR3-1600 with Devil's Canyon, same will happen with unlocked Skylake and DDR4-2133.

 

[AtenRa]

Same happen with DDR-2 vs DDR-3, DDR-2 had lower latency but DDR-3 had higher bandwidth.

When you review at default you should use the highest each CPU OFFICIALLY supports and that is DDR-3 1600MHz for Haswell and DDR-4 2133MHz for the Skylake.
Now if you want to measure at OC or out of specs with higher Memory, you should use higher memory for both systems, DDR-3 2400MHz or 2666MHz and DDR-4 2666MHz or 3000MHz+.

 

[Sweepr]

Some tips about Surface Pro 4's CPU:

- It's not Core-M
- Not Broadwell
- October launch

Most likely Skylake-U. I hope they offer different flavours, give me Skylake-U GT3e too M$.

www.diggita.it/v.php?id=1481582

 

[witeken]

True latency (ns) = clock cycle time (ns) x number of clock cycles (CL)

Why isn't the number of clock cycles the only thing that matters for performance?

 

[therealnickdanger]

Same happen with DDR-2 vs DDR-3, DDR-2 had lower latency but DDR-3 had higher bandwidth.

When you review at default you should use the highest each CPU OFFICIALLY supports and that is DDR-3 1600MHz for Haswell and DDR-4 2133MHz for the Skylake.
Now if you want to measure at OC or out of specs with higher Memory, you should use higher memory for both systems, DDR-3 2400MHz or 2666MHz and DDR-4 2666MHz or 3000MHz+.

I wouldn't mind three categories:

1. Default = stock clocks with best officially-supported RAM (represents the experience of probably 95% of users)
2. OC vs OC = run everything at 4.5GHz with the fastest RAM (closest thing to apple-to-apples)
3. Max OC vs Max OC = run everything as fast as it can go (not great due to binning)

 

[ShintaiDK]

Another way, like the E series. Would simply be to increase the cache on Skylake to 12-16MB. Or simply to use eDRAM as with Broadwell-C. It would reduce the impact instead of having to use afster memory to offset the worse latency of DDR4.

 

[LTC8K6]

I wouldn't mind three categories:

1. Default = stock clocks with best officially-supported RAM (represents the experience of probably 95% of users)
2. OC vs OC = run everything at 4.5GHz with the fastest RAM (closest thing to apple-to-apples)
3. Max OC vs Max OC = run everything as fast as it can go (not great due to binning)

I'd be happy if reviewers did that.

A little hard to believe that 95% of "K" cpu users are running at stock CPU and ram speeds, though.

 

[coercitiv]

Why isn't the number of clock cycles the only thing that matters for performance?

Warning, analogy following. Skip to end for tl;dr
Two things matter when I try to write "Blade Runner" in a thread post:
- my typing speed (equals bandwidth)
- time to remember the name the famous movie (equals latency)

If you ever had to learn a rather large text by heart, you probably know that once you get started writing it on a piece of paper you probably feel your hand is too slow compared to how fast you could reproduce it in your head (or through speech). You were bandwidth limited.

However, there must have been times when you had to pause writing a document simply because it took a bit of time to remember a name or equation. You were access time limited.

When CPU waits for lots of data, bandwidth matters most, but when CPU stalls operations for just a few bits, latency is paramount.

 

[phillyman36]

I talked to a local computer shop and asked them when/if they had and Skylake cpu. They told me they expect stock by OCTOBER 1. I hope because they are a small shop that isn't going to be the case for Amazon and Newegg.

 

[TeknoBug]

Surface Pro 4 sounds interesting, probably outrageously priced though- since Xeon mobile was mentioned by Intel, I wonder if that's what will be used in Surface Pro 4's.