Coffeelake thread, benchmarks, reviews, input, everything.

[Dufus]

Myth? Really?
https://www.gamersnexus.net/guides/...iquid-metal-vs-thermal-paste-benchmarks-7900x

That doesn't make sense. The effectiveness of the block and water for removing the heat should remain the same as should the temperature at the block under the same power and ambient conditions. The core temperature is referenced to the block temperature so if a higher temperature differential exists at the internal IHS TIM then core temperature will be higher, pretty straight forward. His argument seems back to front, having lower temps with better internal TIM means the enthusiast is able to use more power and using more power would generate hotter water with the same water cooling conditions.

Why not do what they have done for mobile CPU's for many years, get rid of the IHS for good. Something to do with using LGA maybe, idk.

Why is IPC so misunderstood? It's pretty much self-defined.

Because majority rules

 

[WhoBeDaPlaya]

Why is IPC so misunderstood? It's pretty much self-defined.

I guess it's like that manager at work this week who claimed that Voltus said a chip's average current was larger than the peak current.

 

[VirtualLarry]

https://videocardz.com/72817/intel-coffelake-s-8th-gen-desktop-and-z370-controversy

Intel decided that starting from Coffeelake-S launch, they will not be releasing detailed information on turbo clock speeds per each core. As you know different frequency will be applied once more cores are in use. Basically, the more cores are in use, the lower the frequency.

Such an obvious thing to know, especially when you are a reviewer, but here’s what Intel said when they were asked for a table of clock speeds for Coffelake-S:

We won’t be sharing these details going forward at this point anymore.

If you think that’s strange, then let me add that Intel even refused to share die sizes of Coffeelake-S.

--

So, Intel isn't even documenting the different core count turbos anymore? Is this normal? Seems... pretty crappy, but par for the part for tech companies these days. (HDD companies don't even tell you what RPM their drives are, half of the time.)

 

[Arachnotronic]

https://videocardz.com/72817/intel-coffelake-s-8th-gen-desktop-and-z370-controversy

Intel decided that starting from Coffeelake-S launch, they will not be releasing detailed information on turbo clock speeds per each core. As you know different frequency will be applied once more cores are in use. Basically, the more cores are in use, the lower the frequency.

Such an obvious thing to know, especially when you are a reviewer, but here’s what Intel said when they were asked for a table of clock speeds for Coffelake-S:

We won’t be sharing these details going forward at this point anymore.

If you think that’s strange, then let me add that Intel even refused to share die sizes of Coffeelake-S.

--

So, Intel isn't even documenting the different core count turbos anymore? Is this normal? Seems... pretty crappy, but par for the part for tech companies these days. (HDD companies don't even tell you what RPM their drives are, half of the time.)

They never used to publish the various turbo frequencies. Sometimes a site like AnandTech would ask and Intel would provide it to them, but it seems that this practice is done too.

So strange of Intel considering that anybody can just buy the darn things and test.

 

[coercitiv]

They never used to publish the various turbo frequencies.

Turbo speed tables on Intel's website, last updated on Sep. 2017:

• Intel® Turbo Boost Technology Frequency Table for Intel® Core™ i7 Desktop Processors
• Intel® Turbo Boost Technology Frequency Tables for Intel® Core™ i5 Desktop Processors

PS: these pages have been around for quite while, I remember them since before Kaby Lake was added.

 

[Arachnotronic]

Turbo speed tables on Intel's website, last updated on Sep. 2017:

• Intel® Turbo Boost Technology Frequency Table for Intel® Core™ i7 Desktop Processors
• Intel® Turbo Boost Technology Frequency Tables for Intel® Core™ i5 Desktop Processors

PS: these pages have been around for quite while, I remember them since before Kaby Lake was added.

Well then, I stand corrected. Nice find.

 

[epsilon84]

It has been shown again and again that the IPC difference is like 8-10 percent, not 10-20 percent. Gaming performance isn't a measure of IPC. Performance scaling with frequency isn't linear for all but canned benchmarks like Cinebench. So 30 percent improvement is for situations with ideal ST scaling.

Shown where exactly?

https://www.techpowerup.com/reviews/AMD/Ryzen_3_1300X/20.html

The i5 7500 @ 3.4GHz has a weighted average of 114.5% across all CPU benchmarks compared to the 1300X @ 3.4GHz

So ~15% falls right in my 10 - 20% IPC figure. CPU bound gaming is closer to 20% on average, but yes, I'll agree with you, that's not IPC. That's simply gaming performance.

EDIT:
Using Anandtechs own data, as another point of reference:
https://www.anandtech.com/show/11658/the-amd-ryzen-3-1300x-ryzen-3-1200-cpu-review/17

The i5 7400 (3GHz Base / 3.5GHz Max Turbo) is ~13% faster overall compared to the R3 1200 (3.1GHz Base / 3.4GHz Max Turbo)

I'll leave it at this because if I say any more I'll be accused of derailing another thread with IPC talk

 

[Zucker2k]

Shown where exactly?

https://www.techpowerup.com/reviews/AMD/Ryzen_3_1300X/20.html

The i5 7500 @ 3.4GHz has a weighted average of 114.5% across all CPU benchmarks compared to the 1300X @ 3.4GHz

So ~15% falls right in my 10 - 20% IPC figure. CPU bound gaming is closer to 20% on average, but yes, I'll agree with you, that's not IPC. That's simply gaming performance.

Great post. I don't even put much stock in tests done at a certain clock speed in the name of highlighting ipc because as far as I'm concerned, it's a minefield; for many reasons.

 

[mikk]

Shown where exactly?

https://www.techpowerup.com/reviews/AMD/Ryzen_3_1300X/20.html

The i5 7500 @ 3.4GHz has a weighted average of 114.5% across all CPU benchmarks compared to the 1300X @ 3.4GHz

So ~15% falls right in my 10 - 20% IPC figure. CPU bound gaming is closer to 20% on average, but yes, I'll agree with you, that's not IPC. That's simply gaming performance.

The problem is that most gaming tests are bottlenecked by its GPU which hurts the faster CPU more than the slower one. The real IPC difference is way bigger, in many games more than 20%.

 

[tamz_msc]

The i5 7400 (3GHz Base / 3.5GHz Max Turbo) is ~13% faster overall compared to the R3 1200 (3.1GHz Base / 3.4GHz Max Turbo)

It has been shown, by averaging over multiple reviews, that the deficit in IPC(using the correct way of measurement according to definition), is around 8-10%. Citing just one result out of multiple data points doesn't prove you right, because what you try to prove with TPU graphs is invalidated by the Anandtech graph-

Ratio of data: 40% ST, 60% Mixed/ST. Meaning they're running at turbo frequency for most of the tests. i5 7400 ACT 3.3GHz, Ryzen 3 1200 ACT 3.1GHz. So i5 7400 ~13% faster = 1.13(3.1/3.3) = 6% adjusting for clock speed.

If we now use the TPU graph with the 1300X and 7500, assuming they're clocking identically, then (6% + 14%)/2 = 10%.
So we're back where we started.

"Gaming IPC" is a stupid metric because gaming workloads aren't as well-defined as application workloads, and GPU/CPU interaction plays a big part in gaming performance.

 

[epsilon84]

It has been shown, by averaging over multiple reviews, that the deficit in IPC(using the correct way of measurement according to definition), is around 8-10%. Citing just one result out of multiple data points doesn't prove you right, because what you try to prove with TPU graphs is invalidated by the Anandtech graph-

Ratio of data: 40% ST, 60% Mixed/ST. Meaning they're running at turbo frequency for most of the tests. i5 7400 ACT 3.3GHz, Ryzen 3 1200 ACT 3.1GHz. So i5 7400 ~13% faster = 1.13(3.1/3.3) = 6% adjusting for clock speed

If we now use the TPU graph with the 1300X and 7500, assuming they're clocking identically, then (6% + 14%)/2 = 10%.
So we're back where we started.

"Gaming IPC" is a stupid metric because gaming workloads aren't as well-defined as application workloads, and GPU/CPU interaction plays a big part in gaming performance.

The 7400 has a lower base clock (3.0GHz) than the Ryzen 3 1200 (3.1GHz). Its maximum turbo is 3.5GHz, 100MHz faster than the R3 1200 (3.4GHz).

The only time where the chips should be running at their max turbo would be the ST tests, where the 7400 would be 100MHz faster. Conversely, if all cores are pegged at 100% and the chips are close to their TDP limits, they will run at the base clock, in which case the R3 1200 would be running 100MHz faster.

Please get your facts right instead of pulling numbers out of nowhere to suit your argument.
https://ark.intel.com/products/97147/Intel-Core-i5-7400-Processor-6M-Cache-up-to-3_50-GHz
https://www.amd.com/en/products/cpu/amd-ryzen-3-1200

 

[tamz_msc]

The 7400 has a lower base clock (3.0GHz) than the Ryzen 3 1200 (3.1GHz). Its maximum turbo is 3.5GHz, 100MHz faster than the R3 1200 (3.4GHz).

The only time where the chips should be running at their max turbo would be the ST tests, where the 7400 would be 100MHz faster. Conversely, if all cores are pegged at 100% and the chips are close to their TDP limits, they will run at the base clock, in which case the R3 1200 would be running 100MHz faster.

Please get your facts right instead of pulling numbers out of nowhere to suit your argument.
https://ark.intel.com/products/97147/Intel-Core-i5-7400-Processor-6M-Cache-up-to-3_50-GHz
https://www.amd.com/en/products/cpu/amd-ryzen-3-1200

Apparently you cannot comprehend how turbo modes work, even when somebody posted them on this same page!

Turbo speed tables on Intel's website, last updated on Sep. 2017:

• Intel® Turbo Boost Technology Frequency Table for Intel® Core™ i7 Desktop Processors
• Intel® Turbo Boost Technology Frequency Tables for Intel® Core™ i5 Desktop Processors

Now compare with Ryzen 3 specs.

 

[eddman]

Performance scaling with frequency isn't linear for all but canned benchmarks like Cinebench. So 30 percent improvement is for situations with ideal ST scaling.

How is CB canned? It's based on a real application and measures real 3D rendering performance. It shows that Lake CPUs have a 20-25% better raw ST performance, (edit) compared to ryzen 7s. Just because other programs do not display a similar advantage, it does not lessen their ST potential.

The mere measure that 6 cores are able to almost match and even occasionally surpass 8 cores says a lot too.

P.S. IPC ≠ ST performance.

 

[Zucker2k]

How is CB canned? It's based on a real application and measures real 3D rendering performance. It shows that Lake CPUs have a 20-25% better raw ST performance. Just because other programs do not display a similar advantage, it does not lessen their ST potential.

The mere measure that 6 cores are able to almost match and even occasionally surpass 8 cores says a lot too.

P.S. IPC ≠ ST performance.

Please see below

I don't like the data equates to [canned] data. OK.

 

[tamz_msc]

How is CB canned? It's based on a real application and measures real 3D rendering performance. It shows that Lake CPUs have a 20-25% better raw ST performance. Just because other programs do not display a similar advantage, it does not lessen their ST potential.

The mere measure that 6 cores are able to almost match and even occasionally surpass 8 cores says a lot.

P.S. IPC ≠ ST performance.

I should have worded it better, but what I meant was that Cinebench, as long as it scales, scales almost perfectly with (freq*core count) on a 1:1 basis. Majority of multithreaded workloads aren't like this. This means that in workloads like Cinebench a higher clocked/higher ST CPU with fewer cores can match a lower clocked/lower ST CPU with more cores.

Does that mean that the 2nd CPU is worse for multithreaded performance? Absolutely not. If you believe otherwise then explain why the Platinum 8160 exists when it should in theory just barely beat the Gold 6154 in a Cinebench-like workload with 33% more cores(freq*core count = 67.2 for 8160, 66.6 for 6154)

 

[tamz_msc]

Please see below

Says the guy who invents his own definition of IPC. Get a break.

 

[MarkPost]

How is CB canned? It's based on a real application and measures real 3D rendering performance. It shows that Lake CPUs have a 20-25% better raw ST performance. Just because other programs do not display a similar advantage, it does not lessen their ST potential.

The mere measure that 6 cores are able to almost match and even occasionally surpass 8 cores says a lot too.

P.S. IPC ≠ ST performance.

Just to clarify, compared to what?

 

[MarkPost]

@eddman If you are referring to Ryzen, thats far from true. My 1700X (fixed at 3.5ghz) = 147 CB (ST); My 7800X (fixed at 3.5ghz) = 149 CB (ST)

AMD (Ryzen) main disadvantage is clock frequency. Intel (Lake) advantage is clock frequency. Thats the reason AMD fight agains it with core count, due to the fact that the majority of tasks that take time (video encoding, rendering etc) are normally heavy MT. These days ST apps are, generally speaking and saving some exception, irrelevant. It wasnt the same ten years ago for example. Present is MT, not to mention the future.

 

[eddman]

I should have worded it better, but what I meant was that Cinebench, as long as it scales, scales almost perfectly with (freq*core count) on a 1:1 basis. Majority of multithreaded workloads aren't like this. This means that in workloads like Cinebench a higher clocked/higher ST CPU with fewer cores can match a lower clocked/lower ST CPU with more cores.

POV-ray also seems to scale well with clock speeds, and is a real application. It seems most (if not even all) rendering tasks benefit from lake's high clocks. As I mentioned, just because not all workloads benefit from high clocks, does not diminish intel's raw ST potential.

Does that mean that the 2nd CPU is worse for multithreaded performance? Absolutely not.

Where did I say that exactly? This is about ST performance only. Ryzen CPUs are still great for full MT tasks; better than CFL if those are your main workloads.

 

[Zucker2k]

Says the guy who invents his own definition of IPC. Get a break.

Oh come on now; you and I know what IPC means. I've probably defined it more times than I care to cite. It is exactly because of people like you, who try to bend the rules to suit their narrative that 'Clock for Clock' IPC tests are not necessarily a good idea. Case in point, look at your posts above? You're doing everything in your power to downplay the scaling efficiency of CB. You should be applauding that, but no, you'd rather see less desirable scaling in order to suppress the fact, I suppose?

 

[eddman]

@eddman If you are referring to Ryzen, thats far from true. My 1700X (fixed at 3.5ghz) = 147 CB (ST); My 7800X (fixed at 3.5ghz) = 149 CB (ST)

What? This is about ST performance, NOT IPC. I clearly stated that. Stop confusing the two people.

It very roughly looks like this: IPC * clock = ST performance.

These days ST apps are, generally speaking and saving some exception, irrelevant. It wasnt the same ten years ago for example. Present is MT, not to mention the future.

ST apps? No one mentioned ST apps. It's about ST performance. There are still many situations where a task does not scale well beyond 4-6 threads, or even less. For those, you still need good ST performance.

P.S. Again, I'm not saying ryzens are weak, no, quite the contrary; just that you cannot so easily dismiss high ST potential.

 

[MarkPost]

What? This is about ST performance, NOT IPC. I clearly stated that. Stop confusing the two people.

It very roughly looks like this: IPC * clock = ST performance.

you were talking about Cinebech ST:

"How is CB canned? It's based on a real application and measures real 3D rendering performance. It shows that Lake CPUs have a 20-25% better raw ST performance. Just because other programs do not display a similar advantage, it does not lessen their ST potential."

 

[eddman]

you were talking about Cinebech ST:

Yes, as a measure of ST performance, that's all. I never meant that there are any ST only applications out there. I doubt any proper developer would make such an application these days.

 

[tamz_msc]

POV-ray also seems to scale well with clock speeds, and is a real application. It seems most (if not even all) rendering tasks benefit from lake's high clocks. As I mentioned, just because not all workloads benefit from high clocks, does not diminish intel's raw ST potential.

To realize the ST potential that Intel has, more often than not, one needs to have applications behave more like Cinebench and POV-ray, where BOTH frequency AND core scaling is nearly-linear. My point is that many workloads aren't like this. An example of an application which doesn't care much about core count if frequency is high enough is Photoshop/Lightroom. An example where the opposite is true would be encoding. A high ST perf CPU would excel in the former, while a high core count CPU, should scaling work, excel in the latter. However this detail is apparently lost by many because they only tend to focus on Cinebench/POV-ray type benchmarks.

Where did I say that exactly? This is about ST performance only. Ryzen CPUs are still great for full MT tasks; better than CFL if that's your main workloads.

Maybe you didn't specifically state this, but many seem to be of the opinion that just because the 8700K should score around 1450cb, it means that the competition's 8-core CPU which also scores similarly is obsolete in MT workloads. Hence this meme CFL 6C/12T = competition's 8C/16T.

Oh come on now; you and I know what IPC means. I've probably defined it more times than I care to cite. It is exactly because of people like you, who try to bend the rules to suit their narrative that 'Clock for Clock' IPC tests are not necessarily a good idea. Case in point, look at your posts above? You're doing everything in your power to downplay the scaling efficiency of CB. You should be applauding that, but no, you'd rather see less desirable scaling in order to suppress the fact, I suppose?

What? You did exactly the opposite when you claimed this:

I don't even put much stock in tests done at a certain clock speed in the name of highlighting ipc because as far as I'm concerned, it's a minefield; for many reasons.

So what is it? Clock vs clock - is it right or wrong? Make up your mind. I remember it was you who claimed that it's irrelevant for something like Bulldozer because that has a high-clocking architecture

 

[MarkPost]

ST apps? No one mentioned ST apps. It's about ST performance. There are still many situations where a task does not scale well beyond 4-6 threads, or even less. For those, you still need good ST performance.

well to me I know ST perf is irrelevant. And I think I'm an average user. I mean, using office apps, creating some PDF, using a web browser, mail... even audio encoding... heck, encoding to mp3 a whole album just take less than a minute, or just some seconds if its to lossless. I use a little Photoshop (a lot of filters arent MT) but takes miliseconds or some seconds in my use anyways. Perhaps you use some app that takes a lot of time doing whatever job and its ST yet, I dont know if its your case. Then yeah ST is relevant for that. But tasks/apps I do taking time is specially video encoding, and its heavy MT.