Intel Skylake / Kaby Lake

[blue11]

I'm making no claims about energy efficiency, just that there are people who do overclock and the TIM might be a problem. Why is it so hard to understand.

For some people that tune their cars, the clean-burning nature of modern engines is also a problem. I remain unconcerned.

Pretty sure that Google doesn't pay ark prices for their chips. And Uncore does nothing for the most part except waste power on Broadwell-E.

"The uncore does nothing except waste power." Maybe if you only run single-threaded workloads like most "enthusiasts" that buy HEDT CPUs and park 8 out of 10 cores. Reminds me a lot of car enthusiasts in fact. Why would you expect to double core frequency (6 GHz!) at the same L3 frequency and see linear scaling?

That clown you refer to helps Asus design their boards.

Designing them for useless overclocking stunts? Not exactly confidence inspiring here.

 

[tamz_msc]

For some people that tune their cars, the clean-burning nature of modern engines is also a problem. I remain unconcerned.


"The uncore does nothing except waste power." Maybe if you only run single-threaded workloads like most "enthusiasts" that buy HEDT CPUs and park 8 out of 10 cores. Reminds me a lot of car enthusiasts in fact. Why would you expect to double core frequency (6 GHz!) at the same L3 frequency and see linear scaling?


Designing them for useless overclocking stunts? Not exactly confidence inspiring here.

Then just buy a Xeon. Why are you so salty? You're laughing at the face of all those who buy HEDT for OCing a 8,10 core CPU to 4.3-4.5GHz.

I don't understand your problem - you are clearly not interested in Intel Core X, so I say it again, calm down and get your Xeon Gold with fully functional AVX512.

To be honest, if I weren't interested at all in gaming, I'd rather have Xeons as well. I presently work with an i7 3770 and a e5 2640v3. I don't feel that I'm lacking in low-threaded tasks with the Xeon.

 

[blue11]

Then just buy a Xeon. Why are you so salty? You're laughing at the face of all those who buy HEDT for OCing a 8,10 core CPU to 4.3-4.5GHz.

I don't understand your problem - you are clearly not interested in Intel Core X, so I say it again, calm down and get your Xeon Gold with fully functional AVX512.

To be honest, if I weren't interested at all in gaming, I'd rather have Xeons as well. I presently work with an i7 3770 and a e5 2640v3. I don't feel that I'm lacking in low-threaded tasks with the Xeon.

I never claimed to be uninterested in Core X. In fact, if the AVX-512 is fully functional, the 12C-18C SKUs will be great, since they can reach 4.5 GHz in single-thread and be more responsive in GUIs and browsers. What is silly is wanting to put every single core at 5 GHz just so you can claim to have done it.

Also, you failed to address the truth of the L3/uncore in your fear-mongering about IPC.

 

[tamz_msc]

I never claimed to be uninterested in Core X. In fact, if the AVX-512 is fully functional, the 12C-18C SKUs will be great, since they can reach 4.5 GHz in single-thread and be more responsive in GUIs and browsers. What is silly is wanting to put every single core at 5 GHz just so you can claim to have done it.

It's funny that this is exactly the same reason people are looking forward to CFL-S 6C12T (not the same as HEDT, but power efficiency argument applies here as well), as a few persons on these forums would have you believe.

 

[Bassman2003]

I think a lot of HEDT users would find a good use for all cores at 5GHz Paying up is usually for need, not for bragging.

 

[lolfail9001]

Better than you might think http://edgeup.asus.com/2016/06/17/broadwell-e-overclocking-guide/

So, 1 hour Handbrake, as expected. Also as expected, it does not pass P95 even for half an hour (people who tried reported thermal throttling under 280mm AIOs).

Now-now, SKL-X is certainly not making latter any better. But with delid i can totally see a 10 core get to 4.8Ghz under top tier water and pass that 1 hour handbrake. What then?

 

[blue11]

It's funny that this is exactly the same reason people are looking forward to CFL-S 6C12T (not the same as HEDT, but power efficiency argument applies here as well), as a few persons on these forums would have you believe.

Are you going to retract your fear-mongering about IPC? I remember when Skylake first came out and this silly narrative of "no IPC or worse IPC" than Haswell/Broadwell was spreading around all the review sites and forums.

 

[tamz_msc]

So, 1 hour Handbrake, as expected. Also as expected, it does not pass P95 even for half an hour (people who tried reported thermal throttling under 280mm AIOs).

Now-now, SKL-X is certainly not making latter any better. But with delid i can totally see a 10 core get to 4.8Ghz under top tier water and pass that 1 hour handbrake. What then?

Same as with 6700K@4.7 vs 7700K@5.0. With people complaining on Intel forums and them responding with "overclocking not recommened" 3-4 months down the line.

 

[tamz_msc]

Are you going to retract your fear-mongering about IPC? I remember when Skylake first came out and this silly narrative of "no IPC or worse IPC" than Haswell/Broadwell was spreading around all the review sites and forums.

I get 10% more CB15 score with 8% more clock speed in hwbot submissions, plus numbers from Intel regarding SPEC2006_Int which are useless given the error bar.

So, no significant IPC improvements, given what we have, and considering what we got from Haswell to Skylake on mainstream platforms.

 

[blue11]

I get 10% more CB15 score with 8% more clock speed in hwbot submissions, plus numbers from Intel regarding SPEC2006_Int which are useless given the error bar.

So, no significant IPC improvements, given what we have, and considering what we got from Haswell to Skylake on mainstream platforms.

I don't understand what you are claiming. In your previous post, you said that the IPC increase from Broadwell-E to Skylake-X will be less than the IPC increase from Haswell/Broadwell-S to Skylake-S ("disappointing"). Now you are saying that the IPC gain matches client SKUs?

 

[tamz_msc]

I don't understand what you are claiming. In your previous post, you said that the IPC increase from Broadwell-E to Skylake-X will be less than the IPC increase from Haswell/Broadwell-S to Skylake-S ("disappointing"). Now you are saying that the IPC gain matches client SKUs?

The question you asked makes no sense.

What's so difficult to understand? Look at any generational chart - Skylake-S is 5-10% ahead in IPC compared to Haswell-S.

This won't be the case with Skylake-X, which is what I'm saying.

 

[blue11]

The question you asked makes no sense.

What's so difficult to understand? Look at any generational chart - Skylake-S is 5-10% ahead in IPC compared to Haswell-S.

This won't be the case with Skylake-X, which is what I'm saying.

OK, so we are on the same page. You think because some guy OC'd a chip to 6 GHz without touching the uncore, Skylake-X won't have any IPC increase over Haswell/Broadwell-E. I bet we won't hear a correction from you when benchmarks are released, and the world turns out to not in fact have ended.

 

[Ajay]

I get 10% more CB15 score with 8% more clock speed in hwbot submissions, plus numbers from Intel regarding SPEC2006_Int which are useless given the error bar.

So, no significant IPC improvements, given what we have, and considering what we got from Haswell to Skylake on mainstream platforms.

The error indicates uncertainty - not certainty. We don't know what the change in IPC, and we can't know base on this one result. Geez, you are on a rampage here arguing points that you don't even understand.

 

[tamz_msc]

The error indicates uncertainty - not certainty. We don't know what the change in IPC, and we can't know base on this one result. Geez, you are on a rampage here arguing points that you don't even understand.

Error by definition means uncertainty. That much everybody, including me, understands.

 

[lolfail9001]

Same as with 6700K@4.7 vs 7700K@5.0. With people complaining on Intel forums and them responding with "overclocking not recommened" 3-4 months down the line.

People complaining... do you reference the temp spike story? It had nothing to do with overclocking, no matter what Intel PR has you believe, but w/e.

 

[tamz_msc]

OK, so we are on the same page. You think because some guy OC'd a chip to 6 GHz without touching the uncore, Skylake-X won't have any IPC increase over Haswell/Broadwell-E. I bet we won't hear a correction from you when benchmarks are released, and the world turns out to not in fact have ended.

Intel could have sold you 18C Broadwell-E dies at these exact same prices and the world would not have ended. Most review sites report IPC gains by comparing CPUs at 3 or 3.5GHz - and I'm looking forward to that. Then we'll have known for certain about IPC. The new cache structure has both advantages and disadvantages compared to the previous one, would like to see an in-depth analysis on that as well.

Uncore doesn't matter. Check it out here

 

[tamz_msc]

People complaining... do you reference the temp spike story? It had nothing to do with overclocking, no matter what Intel PR has you believe, but w/e.

Don't operate with parameters outside of what we specify - same thing, but w/e.

 

[blue11]

Uncore doesn't matter. Check it out here

Of course it matters. Uncore frequency directly determines the L3 cache bandwidth. You link to some result that shows quad-core CPUs overclocked by less than 25% (4.4 to 5 GHz) with already high client uncore frequencies (3.6 GHz or more) do not show L3 bottlenecks. Then you try to say the same has to be true for server/HEDT CPUs with way more cores overclocked by over 100% (3.x GHz to 6 GHz) and already low 2.4 GHz uncore frequencies. L3 bandwidth is critical for multi-threaded scalability (and "IPC") and one of the bright spots for IPC in Skylake compared to Haswell/Broadwell.

 

[RichUK]

Too much animosity in this thread.

Trolling is not allowed. This is a technical forum
Markfw
Anandtech moderator

 

[tamz_msc]

Of course it matters. Uncore frequency directly determines the L3 cache bandwidth. You link to some result that shows quad-core CPUs overclocked by less than 25% (4.4 to 5 GHz) with already high client uncore frequencies (3.6 GHz or more) do not show L3 bottlenecks. Then you try to say the same has to be true for server/HEDT CPUs with way more cores overclocked by over 100% (3.x GHz to 6 GHz) and already low 2.4 GHz uncore frequencies. L3 bandwidth is critical for multi-threaded scalability (and "IPC") and one of the bright spots for IPC in Skylake compared to Haswell/Broadwell.

1. Uncore doesn't matter for overclocking 6700K or 7700K.
2. It may matter for large core count chips with all the extra L3
3. I'm not taking your word for number 2 just because you say so, without independent testing.
4. We'll see how much IPC is improved when we get reviews.
5. I'm not under any obligation to redact any statements I make here, same being applicable to anyone, because nothing hinges on it.

 

[blue11]

1. Uncore doesn't matter for overclocking 6700K or 7700K.

Because some dudes on an OC forum said so? Even on quad-cores, L3 bandwidth limits scalability in numeric software if not carefully written. Skylake is better than Haswell in bandwidth, but running out of L3 bandwidth will clearly block your software from scaling. It doesn't take any genius insight to realize this.

2. It may matter for large core count chips with all the extra L3
3. I'm not taking your word for number 2 just because you say so, without independent testing.

You can write the code and see for yourself. Just because you don't believe it doesn't make it untrue.

4. We'll see how much IPC is improved when we get reviews.
5. I'm not under any obligation to redact any statements I make here, same being applicable to anyone, because nothing hinges on it.

Of course I expect to hear crickets from you when all your doomsday predictions fail to materialize.

 

[Bouowmx]

Just waiting patiently for Intel Coffee Lake stuff
Lots of vigorous back-and-forth

 

[tamz_msc]

Because some dudes on an OC forum said so? Even on quad-cores, L3 bandwidth limits scalability in numeric software if not carefully written. Skylake is better than Haswell in bandwidth, but running out of L3 bandwidth will clearly block your software from scaling. It doesn't take any genius insight to realize this.

You can write the code and see for yourself. Just because you don't believe it doesn't make it untrue

Of course I expect to hear crickets from you when all your doomsday predictions fail to materialize.

Why don't you provide your data with the "numeric software" to prove your claims? I can have one look at FFTW results for Ryzen and Haswell, for example, and say that they give the same IPC in FFTW. Too bad they don't have Skylake. What can you provide, except vapid remarks?

 

[blue11]

Why don't you provide your data with the "numeric software" to prove your claims? I can have one look at FFTW results for Ryzen and Haswell, for example, and say that they give the same IPC in FFTW. Too bad they don't have Skylake. What can you provide, except vapid remarks?

Why are you suddenly discussing Zen and Haswell? All I said was that scaling core throughput (by clock frequency) 2x without increasing bandwidth will never yield 2x the performance. Why are you so defensive about this?

 

[swilli89]

Its absolutely, stunningly mind blowing that Intel cheaped out on the TIM and didn't solder the IHS on these "high end" products. https://www.youtube.com/watch?v=I1Bv8Mxnnlc

Sure forget the fact that it took AMD making a move to force Intel above 12 cores.. but even now that they are forced to compete in a market segment AMD created.. they are doing everything they can to save a penny. Amazing.