Intel to launch 6 core Coffee Lake-S CPUs & Z370 chipset 5 October 2017

[smilingcrow]

.. they will likely solve the problem completely by the time we're on Zen2 or Zen3 and support for 3600+ RAM. And in the longer term, DDR5 should put an end to it in ~2021, since the slowest standard speed will be 4266 MHz CAS 28. And a decent tuned speed would be 6400 MHz CAS 28.

Slow down dude and stay present as Zen3 and DDR5 are way into the future.
The question for me is whether Zeppelin's dual module design is significantly hurting performance for certain workloads and if that is the case is there ever going to be a solution for that which brings it into contention with whatever Intel have at the time.
I'm talking mainstream rather than HEDT which has other issues.

 

[AllBodies]

Slow down dude and stay present as Zen3 and DDR5 are way into the future.
The question for me is whether Zeppelin's dual module design is significantly hurting performance for certain workloads and if that is the case is there ever going to be a solution for that which brings it into contention with whatever Intel have at the time.
I'm talking mainstream rather than HEDT which has other issues.

100% yes, just will take optimisation by the devs. Ryzen is already very close to the 7700k in scenario's where it's been optimised. And kills it when it's optimised + uses 6+ cores.

And the reason I say 100% is because Intel's HEDT has the same issue, which automatically means Intel will push hard for optimisations for Skylake-X. And these same optimisations will help Ryzen.

Plus I did point out the problem will likely be solved with Zen2 and 3600 low-latency RAM. So it could go away for good within 18 months. Without dedicated software help.

EDIT: Oh also if you double-check the gamersnexus results, it shows barely any gain from the i3-7350k to the i7-7700k too. I hadn't spotted that at first. So the 7700k should also be doing better than it is. There's clearly a threading issue in general, but particularly with Ryzen since SMT isn't even working.

 

[gavinh87]

100% yes, just will take optimisation by the devs. Ryzen is already very close to the 7700k in scenario's where it's been optimised. And kills it when it's optimised + uses 6+ cores.

And the reason I say 100% is because Intel's HEDT has the same issue, which automatically means Intel will push hard for optimisations for Skylake-X. And these same optimisations will help Ryzen.

Plus I did point out the problem will likely be solved with Zen2 and 3600 low-latency RAM. So it could go away for good within 18 months. Without dedicated software help.

EDIT: Oh also if you double-check the gamersnexus results, it shows barely any gain from the i3-7350k to the i7-7700k too. I hadn't spotted that at first. So the 7700k should also be doing better than it is. There's clearly a threading issue in general, but particularly with Ryzen since SMT isn't even working.

Except ryzen is using 8 cores, just not the hyper threads which in most games should provide better performance.

 

[AllBodies]

Except ryzen is using 8 cores, just not the hyper threads which in most games should provide better performance.

Yes but it might be crossing CCX boundaries a lot, like how the windows scheduler treated Ryzen wrong at launch. And/or just not be optimised at all for Ryzen's specific pipeline.

By the nature of you saying that, you know there's something wrong (that the devs have/haven't done). If a program of any kind uses 8-cores correctly, an 8-core Ryzen will eat the 7700k for breakfast.

 

[triss]

Yes but it might be crossing CCX boundaries a lot, like how the windows scheduler treated Ryzen wrong at launch. And/or just not be optimised at all for Ryzen's specific pipeline.

By the nature of you saying that, you know there's something wrong (that the devs have/haven't done). If a program of any kind uses 8-cores correctly, an 8-core Ryzen will eat the 7700k for breakfast.

Correctly isnt really the right word as there are programs that can use 8 cores fine, just due to the different way cores are with ryzen it requires a different approach.
No longer a 1 fit for all deal. X299 shows this aswell

 

[gavinh87]

Yes but it might be crossing CCX boundaries a lot, like how the windows scheduler treated Ryzen wrong at launch. And/or just not be optimised at all for Ryzen's specific pipeline.

By the nature of you saying that, you know there's something wrong (that the devs have/haven't done). If a program of any kind uses 8-cores correctly, an 8-core Ryzen will eat the 7700k for breakfast.

Except the windows scheduler was proven to be working as intended, as said by AMD themselves. This whole Nvidia not being optimised for higher threaded CPUs is also an unproven rumour. The latency from ccx is hurting ryzen in games that can use more cores. It's low clock speed and haswell levels of IPC is also hurting it in games when compared to the competition. It's seems to be a lose lose situation.
Something is wrong yes, but are we really expecting AMD to jump in and help optimise every game that gets released?
I fear this is going to become the norm for ryzen. If the game is big enough it will get sorted, if not then your out of luck.

 

[voidshatter]

Yep, very annoying. I'd understand if things were being ironed out but I see no evidence of this, not in my case anyway.
This may very well be nothing but 3 of us were in game, 2 of us crashed and got booted out and 1 remained.
The 2 of us that got booted were using a 1700 and a 1600. The guy remaining had an i7.
This may very well be a coincidence but add this to the other quirks ive been experiencing and meh, not worth the hassle.

I remember that I mentioned AMD to have more bugs, and got attacked hard. Someone even believed me to have a personal vendetta against anything AMD... I was merely forwarding points by my friend T who happens to have worked for Baidu and Tencent.

Translation of conversation between me and my friend T below:

Me: Baidu has started using AMD (Ryzen)
T: Yes, my colleague has fallen for the sales pitch
T: AMD's instruction sets are not so compatible
Me: What instruction sets?
T: If they set high optimisations for GCC's compiling, then certain programs would crash
T: The market is like this. The ecology has been broken for a long time (due to AMD's hardware staying behind for too long).
T: It's of no use if there's no software support, no matter how cheap the hardware is.
Me: But the price performance ratio for Ryzen is already halved
T: Yes, that's why I said for home usage it's totally fine (to pick AMD)
T: I wouldn't risk using AMD for industry

 

[DragonQ]

I remember that I mentioned AMD to have more bugs, and got attacked hard. Someone even believed me to have a personal vendetta against anything AMD... I was merely forwarding points by my friend T who happens to have worked for Baidu and Tencent.

Translation of conversation between me and my friend T below:

Me: Baidu has started using AMD (Ryzen)
T: Yes, my colleague has fallen for the sales pitch
T: AMD's instruction sets are not so compatible
Me: What instruction sets?
T: If they set high optimisations for GCC's compiling, then certain programs would crash
T: The market is like this. The ecology has been broken for a long time (due to AMD's hardware staying behind for too long).
T: It's of no use if there's no software support, no matter how cheap the hardware is.
Me: But the price performance ratio for Ryzen is already halved
T: Yes, that's why I said for home usage it's totally fine (to pick AMD)
T: I wouldn't risk using AMD for industry

Um so GCC optimisations can cause crashes on AMD CPUs. Not entirely sure how you leap from there to "AMD has more bugs". GCC has been optimised for Intel for years, it doesn't surprise me at all that there are issues with the optimisations it does. I'm sure Ryzen cores have bugs, just like all of Intel's ones do, but I bet some of this is just GCC incompatibility.

 

[SiDeards73]

Um so GCC optimisations can cause crashes on AMD CPUs. Not entirely sure how you leap from there to "AMD has more bugs". GCC has been optimised for Intel for years, it doesn't surprise me at all that there are issues with the optimisations it does. I'm sure Ryzen cores have bugs, just like all of Intel's ones do, but I bet some of this is just GCC incompatibility.

+1

Its laughable the lengths people will go to just to put shadow on AMD right now, people have small memories and narrow minds... all this stuff has been optimized and only run on Intel hardware for how many years? and because you can drop in another Intel chip with the same instruction sets etc you overlook these things.

The minute AMD bring competition to the marketplace, its like people lose their minds and cant comprehend another company can be competitive and go to lengths to find reasons to deny any positives etc

 

[voidshatter]

Ryzen and TR have both been reported to have relatively weak performance for DAW usage when the workload is tuned for low latency audio.
Latency is much more important for audio than gaming as a few audible glitches are more distracting than the odd slow frame.
I wonder if the variable latency caused by the dual module approach of the Zeppelin die is the culprit here?
As good as the Zeppelin die is it does seem to have an Achilles heal albeit a relatively small one.
This downside is one reason why people are interested in CL even though in many ways AMD are more appealing.
It's a case of looking to see if the downside is worth it for the upside and that varies by usage.
As a DAW user CL is a slam dunk from what I've seen so far.

Wait a moment - Even 8-core Ryzen (not TR) has a Zeppelin die design? I thought it's just the TR and EPYC having dies stitched together...

 

[DragonQ]

Wait a moment - Even 8-core Ryzen (not TR) has a Zeppelin die design? I thought it's just the TR and EPYC having dies stitched together...

No, AM4 Ryzen is one die with two quad-core CCXs.

 

[voidshatter]

Just how is Kaby Lake or Coffee Lake's die to be more consistent regarding latency between cores? Via the shared L3 cache?

 

[smilingcrow]

Wait a moment - Even 8-core Ryzen (not TR) has a Zeppelin die design? I thought it's just the TR and EPYC having dies stitched together...

Zeppelin is a single chip so they are all Zeppelin based CPUs just with different number of Zeppelin chips.

 

[Hagar]

Wait a moment - Even 8-core Ryzen (not TR) has a Zeppelin die design? I thought it's just the TR and EPYC having dies stitched together...

use the correct terminology Glued together

 

[voidshatter]

Zeppelin is a single chip so they are all Zeppelin based CPUs just with different number of Zeppelin chips.

If AM4 Ryzen is just a single die, then the communication latency between two CCX groups shouldn't be that high, not as high as between dies. How is Kaby Lake better in this regard? Make the following assumption:

latency between dies > latency between CCX groups in the same die > latency between Kaby Lake cores?

 

[gavinh87]

If AM4 Ryzen is just a single die, then the communication latency between two CCX groups shouldn't be that high, not as high as between dies. How is Kaby Lake better in this regard? Make the following assumption:

latency between dies > latency between CCX groups in the same die > latency between Kaby Lake cores?

This probably shows it better



https://www.pcper.com/reviews/Proce...Core-i5/CCX-Latency-Testing-Pinging-between-t

Notice that going from 2133 to 3200 reduces the latency from 140ns to 110ns, a fair amount but still a far cry away from kaby. I suspect this is where the issue lies with games that use more than 4 cores.

 

[DragonQ]

This probably shows it better



https://www.pcper.com/reviews/Proce...Core-i5/CCX-Latency-Testing-Pinging-between-t

Notice that going from 2133 to 3200 reduces the latency from 140ns to 110ns, a fair amount but still a far cry away from kaby. I suspect this is where the issue lies with games that use more than 4 cores.

Yes, somewhere along the line, whether it be in the game engine, DirectX/Vulkan, the OS, or the CPU driver, something has to prioritise keeping things on the same CCX where possible. If not, you can end up with a performance hit due to the high inter-CCX latency. That's where the talk of "optimisations" comes in.

 

[gavinh87]

Yes, somewhere along the line, whether it be in the game engine, DirectX/Vulkan, the OS, or the CPU driver, something has to prioritise keeping things on the same CCX where possible. If not, you can end up with a performance hit due to the high inter-CCX latency. That's where the talk of "optimisations" comes in.

Whilst that's achievable I don't see it happening for many titles. I'm getting flashbacks of the 8350 optimisations that never arrived. Yes I bought into that hype train too and got a fx6300.

 

[voidshatter]

Whilst that's achievable I don't see it happening for many titles. I'm getting flashbacks of the 8350 optimisations that never arrived. Yes I bought into that hype train too and got a fx6300.

I get it... You are just generally itchy to upgrade every time something new comes out, enjoying the process of tying out shiny new stuff, and you are definitely not financially-bound, not one of those with a budget limit

Multi-thread programming is difficult enough - with huge maintenance cost taken into account (debugging, patching, optimisation etc). Nowadays game engines are generally developed with the architecture of consoles in mind, and it is probably not that easy to optimise for communication latency intelligently for more complicated topology like inter-CCX communications, inter-die communications and inter-socket communications.

Does the above explain why Skylake-X's mesh may not be a good choice for gaming?

 

[gavinh87]

I get it... You are just generally itchy to upgrade every time something new comes out, enjoying the process of tying out shiny new stuff, and you are definitely not financially-bound, not one of those with a budget limit

Multi-thread programming is difficult enough - with huge maintenance cost taken into account (debugging, patching, optimisation etc). Nowadays game engines are generally developed with the architecture of consoles in mind, and it is probably not that easy to optimise for communication latency intelligently for more complicated topology like inter-CCX communications, inter-die communications and inter-socket communications.

Does the above explain why Skylake-X's mesh may not be a good choice for gaming?

I wouldn't say I like upgrading but I do seem to fall for the hype of AMD products. Before that my 4770 served me for a good while. Whilst not financially bound I don't like going past the point of diminishing returns for my money.
The x299 is not a good choice for gaming, as I have said all along. The price/performance ratio isn't good for gaming.

We've heard this consoles using more cores for years now, it was going around with the 360 and PS3 and it didn't do the fx series any good back then.