AMD Ryzen (Summit Ridge) Benchmarks Thread (use new thread)

Page 149 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Status
Not open for further replies.

krumme

Diamond Member
Oct 9, 2009
5,956
1,595
136
Man that Intel guy at the end, didn't even say thank you. Kind of rude if you ask me.
Naa. I dont hink so. First they are all a little bit nervous. You can also tell clark is by his clearing his throat often. Secondly I think its more just because they are focusing on the question and the answer and right after the answer starts to analyze it. Thats how interprete the situation.
 

richaron

Golden Member
Mar 27, 2012
1,357
329
136

Dude you are "No"-ing my claims that someone should not be dissapointed with, or yell "fail" at, a CPU with roughly 40% IPC gains from AMD's past architecture. That is: I'm saying no-one should be surprised if leaked benchmarks perform in that range. Since everyone has been talking about this forever.

Unfortunately, that is a problem when the person to be ignored happens to be the OP and an active participant in a "main benchmark thread".

The person/OP might also have an obvious history of starting threads and promoting topics which paint AMD in a bad light. And even though their comments appears at first glance to be objective and informative, on the whole one can easily see the selective and biased nature of these posts.

Personally I find it hard to fulfill my moral obligations of informing newcomers/casual readers of incorrect reasoning and serial biased posters whilst also respecting the "no call-outs" rules on this forum. But of course I will try to fulfill both.
 
  • Like
Reactions: coffeemonster

coercitiv

Diamond Member
Jan 24, 2014
6,630
14,061
136
Naa. I dont hink so. First they are all a little bit nervous. You can also tell clark is by his clearing his throat often. Secondly I think its more just because they are focusing on the question and the answer and right after the answer starts to analyze it. Thats how interprete the situation.
That was also his second question, he came back back for a clarification. The first time he did say "ok, thank you" in a polite manner.
 

krumme

Diamond Member
Oct 9, 2009
5,956
1,595
136
That was also his second question, he came back back for a clarification. The first time he did say "ok, thank you" in a polite manner.
Yeaa it was actually a nice clarification.
But i can understand its difficult as its kind of a delicate situation :)

Unfortunately i would have liked a better answer to the first question about power and zen adressing both big and little core market. Anyway its great there is culture for those q&a. But engineers and technicians generally have a very cool working culture as i can tell as an outsider.
 

bx3

Junior Member
Dec 20, 2016
1
0
1
Yeaa it was actually a nice clarification.
But i can understand its difficult as its kind of a delicate situation :)

Unfortunately i would have liked a better answer to the first question about power and zen adressing both big and little core market. Anyway its great there is culture for those q&a. But engineers and technicians generally have a very cool working culture as i can tell as an outsider.
The only conclusion of this debate is that the cpus is not to be released before q3, othewise we should have seen some serious leaks..

bx3
 

unseenmorbidity

Golden Member
Nov 27, 2016
1,395
967
96
We have to remember they say a quarter of the performance uplift in zen is due to the prediction...;)
Where was that said? I know I saw something like that somewhere, but I could not remember the source.

The person/OP might also have an obvious history of starting threads and promoting topics which paint AMD in a bad light. And even though their comments appears at first glance to be objective and informative, on the whole one can easily see the selective and biased nature of these posts.

Personally I find it hard to fulfill my moral obligations of informing newcomers/casual readers of incorrect reasoning and serial biased posters whilst also respecting the "no call-outs" rules on this forum. But of course I will try to fulfill both.

Why can't they just ban them!?
 
  • Like
Reactions: coffeemonster

TheELF

Diamond Member
Dec 22, 2012
4,027
753
126
Where was that said? I know I saw something like that somewhere, but I could not remember the source.
At the new horizon presentation,the first thing lucy lu talked about was senseMI,which is comprised of the neural network and smart prefetch* and together those two account for 1/4 of the IPC uplift.

* and pure power and xfr
 

superstition

Platinum Member
Feb 2, 2008
2,219
221
101
Today I spent (too) many hours while testing Blender.

I made few customized builds including the standard MSVC 2013 compiled version, with and without AVX2 and then a similar set of builds with Cycles compiled with ICL. It soon became obvious that despite Blender does have AVX2 kernel present, it does absolutely nothing (at least speed wise). The AVX2 kernel can be disabled easily during build time (CCX_HAS_AVX2 = False) and disabling it didn't change the performance even the slightest neither on MSVC or ICL. The funniest thing is that compiling Cycles with ICL generic arch tunings (SSE2, SSE3, SSE4.1 & AVX) improved the performance on Piledriver by ~ 39%. Meanwhile the performance on Haswell-E improved by < 9%. So much for the ICL making AMD CPUs look worse than they actually are ;)

Even when compiled with ICL, the SMT yield in Blender is abnormally high. Combined with the fact that the AVX2 kernel seems to do nothing, I don't think Blender manages to extract all of the performance potential out of Haswell and newer Intel parts.

Tomorrow morning I'll upload all of the builds I compiled today, so other people can test them too.
Any chance you could repost these? A 39% improvement for Piledriver explains a lot, since it was getting crushed by 4 core 4 thread Lynnfield.

https://forums.anandtech.com/threads/summit-ridge-zen-benchmarks.2482739/page-24#post-38436728
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
26,129
15,275
136
Where was that said? I know I saw something like that somewhere, but I could not remember the source.



Why can't they just ban them!?


This is posting as a mod. Unless someone posts an extremely provocative post, everyone is allowed an opinion, and mod action is not allowed. You have to step out of the rules.
Please stay within the post discussion in a technical manner,.

Markfw
Anandtech Moderator.
 

superstition

Platinum Member
Feb 2, 2008
2,219
221
101
What??? For a fp bm ! Lol. This then goes to show the information gained by CB R15 is either borked or bend to a purpose making bw shine vs prior gen. Such a ipc result to me - beeing of by aprox 300% vs the results eg AT get over a broad portfolio of apps - just points to this beeing a marketing tool for sellig the new generation of cpu.

But yeaa your post made me look at
http://images.anandtech.com/doci/9482/IPC Over Sandy.png?_ga=1.59123949.1968553451.1464761965

Cb r15 shows by far the biggest gain of all.

Unfortunately marketing can be dangerous because it the pull down the performance of your next gen skylake :)

http://images.anandtech.com/doci/9483/01 - Gains over Sandy.png?_ga=1.132277069.258125519.1478557456

http://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/9

But i guess thats what a new CB R18 should adress. Lol. There is work for all; Rinse and repeat.
Interesting.
 

The Stilt

Golden Member
Dec 5, 2015
1,709
3,057
106
Any chance you could repost these? A 39% improvement for Piledriver explains a lot, since it was getting crushed by 4 core 4 thread Lynnfield.

https://forums.anandtech.com/threads/summit-ridge-zen-benchmarks.2482739/page-24#post-38436728

Unfortunately I've already deleted them.
However I made a new, similar build few days back.
https://1drv.ms/u/s!Ag6oE4SOsCmDhFJcooVchjDI9SFD
https://1drv.ms/u/s!Ag6oE4SOsCmDhFJcooVchjDI9SFD

Password: "SIMD"

That is the same 2.78.4 build, with Cycles kernels compiled with MSVC 2015 + ICL 2017 (instead of just MSVC).
The vendor dependant dispatcher has been removed, however the instructions dependant dispatcher is still present (for obvious reasons).
On the newer build using ICL on Cycles kernels boosts performance on Haswell-EP by ~5%, but I haven't tried it on any AMD CPUs.
 

superstition

Platinum Member
Feb 2, 2008
2,219
221
101
Thanks for the reply. Is that the Blender_278_1412GIT_AVX2_MSVC file or a different one? I already started testing the one I mentioned. It got vastly better scores than the stock Blender did on Piledriver. I haven't had a chance to test with Lynnfield yet:

update: I see that it is a different build and am updating my results as they come in.

update 2: Thanks for the builds. My machine froze even after I upped voltages. The VRM system in this board is still a problem. I'm going to go back to a lower clock and test again.

Windows 8.1 64-bit, 8370E 5 GHz, 1814 speed CAS 9 16 GB RAM, 238 BCLK, UD3P 2.0 board FC BIOS

Blender 2.78.4 — The Stilt "SIMD" build (Blender_278_GIT_SIMD_INTRIN)

100 samples: 0:35.62
150 samples: 0:53:38
200 samples: 1:12.47

Blender 2.78.4 — The Stilt "MSVC" build (Blender_278_1412GIT_AVX2_MSVC)

100 samples: 0:39.67
150 samples: 0:59.21
200 samples: 1:18.65

Blender 2.78a

100 samples: 1:08
150 samples: 1:41.5
200 samples: 2:14.5

2.77a

100 samples: 1:07
150 samples: 1:40
200 samples: 2:13

2.76b

100 samples: 1:06
150 samples: 1:39
200 samples: 2:11

2.75a

100 samples: 1:04
150 samples: 1:35
200 samples: 2:06

-------------------------------------------------------

CPU-Z 1.77.0.x64

single: 1471
multi: 9561

CB R15

single: 120
multi: 797

CB 11.5

single: 1.32
multi: 8.64
mp ratio: 6.53x
 
Last edited:

BeepBeep2

Member
Dec 14, 2016
86
44
61


So if those results are fake would the OP please remove them from his postings in this thread.

Not so fast on calling out the CB15 guys. ;)
CortexA99 @ Overclock.net said:
Great find! I'll share your work to my friends.
thumb.gif


I'm not surprised if cinebench score is closed to reality, if it just perform a little bit better it should fits my expectation that Zen perform in cinebench. And it's not contradict with what AMD demo us using Blender.

The fritzchess looks quite unrealistic though, I don't think it's benched with same CPU.

http://www.overclock.net/t/1617227/...eview-on-12-13-at-3-pm-cst/1600#post_25723557
 

unseenmorbidity

Golden Member
Nov 27, 2016
1,395
967
96
This is posting as a mod. Unless someone posts an extremely provocative post, everyone is allowed an opinion, and mod action is not allowed. You have to step out of the rules.
Please stay within the post discussion in a technical manner,.

Markfw
Anandtech Moderator.
A blatant and obvious anti-amd troll thread to incite people isn't an extremely provocative post? The WHOLE point of the thread is to be provocative!





Mod callouts are not allowed. You have an issue with the moderation?
Make an MD thread.


esquared
Anandtech Forum Director
 
Last edited by a moderator:
  • Like
Reactions: coffeemonster

rvborgh

Member
Apr 16, 2014
195
94
101
Hi Stilt, that new build drops performance on K10 at least.

24 cores at 3.0 Ghz 150 samples: 41.70 seconds
48 cores at 3.0 Ghz 150 samples: 21.80 seconds

previous official 2.78a build:

24 threads at 3.0 Ghz 150 samples: 35.67 seconds
48 threads at 3.0 Ghz 150 samples: 18.70 seconds (i reran this one after i reverted)

Unfortunately I've already deleted them.
However I made a new, similar build few days back.
https://1drv.ms/u/s!Ag6oE4SOsCmDhFJcooVchjDI9SFD

Password: "SIMD"

That is the same 2.78.4 build, with Cycles kernels compiled with MSVC 2015 + ICL 2017 (instead of just MSVC).
The vendor dependant dispatcher has been removed, however the instructions dependant dispatcher is still present (for obvious reasons).
On the newer build using ICL on Cycles kernels boosts performance on Haswell-EP by ~5%, but I haven't tried it on any AMD CPUs.
 

superstition

Platinum Member
Feb 2, 2008
2,219
221
101
8370E at 5 GHz and CMT off (4M/4T, 1 integer and 1 FPU per module, Gigabyte BIOS option), Windows 8.1, 1814 CAS 9 RAM

Blender 2.78.4 — The Stilt "SIMD" build (Blender_278_GIT_SIMD_INTRIN)

100 samples: 0:59.59
150 samples: 1:29.36
200 samples: 1:58.46

Blender 2.78.4 — The Stilt "MSVC" build (Blender_278_1412GIT_AVX2_MSVC)

100 samples: 1:06.21
150 samples: 1:39.53
200 samples: 2:11.45

Blender 2.78a

100 samples: 2:04.66
150 samples: 3:06.01
200 samples: 4:08.26

Blender 2.75a

100 samples: 1:56.83
150 samples: 2:54.58
200 samples: 3:52.77

CB R15

single: 119
multi: 467

With the R15 OpenGL test I interestingly got a higher FPS rate than with even 5.12 GHz and the fully-enabled CPU.

4M/4T 5 GHz PD, 7870 GPU: 106.69 fps
4M/8T 5.12 GHz PD, 7870 GPU: 105.81 fps
4M/8T 4.92 GHz PD, 7870 GPU: 102.88 fps

CB R11.5

single: 1.32
multi: 4.90
mp ratio: 3.70x

OpenGL test showed no significant change from 4M/8T test (83.x fps).

Interesting! It looks like CPU-Z's test is not optimized for 4M/8T Vishera. The multi score is higher than I ever got with 5 GHz in a full-enabled CPU and the single score is within the margin of variation. So much for this test being superior to Cinebench. At least Cinebench's multicore test (although not OpenGL) scales rather than declines.

CPU-Z 1.77.0 x64

single: 1470
multi: 5670
 
Last edited:
  • Like
Reactions: BeepBeep2

Nothingness

Diamond Member
Jul 3, 2013
3,072
2,066
136
If you have to ask for that to know the answer you need a reboot and change the predictor. Thats for sure. Take care the next 20 years is not 20 times 1 year then ;)
And if instead of trying hard to look like a fool, you'd also bet on branch prediction? I placed my bet despite an ill-defined context and tried to explain why I did bet on Intel I (though admittedly the reason is scientifically poor).

Do you think Ryzen will do better than Intel on branch prediction? And why? How do you intend to measure that?
 

krumme

Diamond Member
Oct 9, 2009
5,956
1,595
136
And if instead of trying hard to look like a fool, you'd also bet on branch prediction? I placed my bet despite an ill-defined context and tried to explain why I did bet on Intel I (though admittedly the reason is scientifically poor).

Do you think Ryzen will do better than Intel on branch prediction? And why? How do you intend to measure that?

I think your explanation is perfectly fine. Actually its the kind of explanation i can understand as the finer technical details is impossible for me to interprete anyways.

As for how to meassure it i will leave it to the experts; meaning folks like you. I understand why its difficult but the same can be said about many other parts. And ofcource it will still be an aproximation.

As i wrote some pages before i think the branch prediction is nearly bwe level. Not due to eg the correlation Dresdenboy have shown between cb and blender for a lot of archs. Not because its needed for the cloud server market.

But because i asume its far less process related, demands the right ideas and knowledge - excactly as shown by the intel compettitions you mention - and because that knowledge is more transferable than eg loads of man hrs optimizing eg cache or layout. I then asume looking at apple core perf that keller could bring in extra knowledge from Apples work to amd and how best to organize work with it. So i asume you were not so dependant on ressources as many other parts.

And then we have Lisa Su stating on stage that a quarter of the performance uplift - and i notice it wasnt ipc - can be related to branch predictor and prefetch.

And btw i am not afraid to look like a fool. Its not important at all. Without mistakes we dont go anywhere.
 
  • Like
Reactions: Nothingness

rvborgh

Member
Apr 16, 2014
195
94
101
Rather interesting... i went through the submissions on the Ryzen 150 thread and converted them as well.

i only found one submission with a higher value... a Skylake 6700K sample. Its too bad that how good AMD's SMT implementation is not known... but typically (R15 score / 1.25 / # cores) gets in the ball park (with the newer Intel chips).

All units in samples/s/core/GHz (converted from 150 sample submissions)

FX-8350 stock 23,668.64
FX-8370E (@ 5GHz) 23645.32
FX-8370E (@ 5Ghz) single thread per module 25,805.06
Phenom 965 BE@ 3.4 28,811.52
A8 3850 32,345.01
Athlon X4 845 34,873.58
Phenom 2 X6 1100 @ 3.4 36,038
K10 (48 core Opteron at 3 GHz) 39,608 (at best when running 4 cores), 35,461 (at worst 48th core)
Sandy Bridge 2500K @ 4.5 Ghz 51,282.05
i5 750 (“Lynnfield” 4C/4T, 3.8 GHz) 51,306.17
Sandy Bridge I5 @ 3 Ghz 52,287.58
Xeon E5 2670 x2 (3.0 GHz turbo) 69135 (first core) 63,694 (16th core)
Xeon 5660 68356.78
3770k 68982.76
4770K 64,205.46
SkyLake i5@3.63GHz 67,838.8
SkyLake 78,947.37
3770 (non K) at 3.4 GHz 83,457.36
5960X 91,427.18
5820K @ 4.4GHz 92,528.34
6950X 93,389.76
5930K 93,811.38
4790K 94,861.66
Ryzen 98,039.22
i7-6700K "SkyLake" 105,207.79

Stock R15 figures:
135 672 2600K 3.4 GHz
143 708 3770k 3.5 GHz
156 791 4770k 3.5 GHz
181 894 4790K 4.0 GHz
140 1337 5960X 3.0 Ghz
153 1547 6900K 3.2 GHz
182 919 6700K 4.0 GHz

Looks like Ryzen sits right in between Haswell refresh and 6700K. 1519 would be the 8 core Haswell refresh score at 3.4 GHz, and 1562.3 would be the 8 core Skylake score. Split the difference... you end up with a projected Ryzen score of 1540 (actually slightly below that since not exactly half). Do the usual formula (1.25 for SMT when running Cinebench), and you end up with 154 for single thread at 3.4 GHz and 176.6 if they can turbo single core up to 3.9 GHz.

So basically we are looking at Ryzen CB R15 of 1540, and 154 ST assuming 1.25 SMT.

Interesting that Athlon X4 845 still has not caught up with K10 all these years (almost equal) in Blender. In CB R15 its just very slightly ahead in ST.

Here's my Ryzen.blend results:

Xeon E5 2670 x2 (3.0 GHz turbo)
FreeBSD 11.0
Blender 2.76, 150 samples (total samples = 800*800*150)

Code:
1 core:   7 minutes 42.86 seconds     69135 samples/s/core/GHz
2 cores: 3 minutes 52.16 seconds      68918 samples/s/core/GHz
4 cores: 1 minute 57.42 seconds       68131 samples/s/core/GHz
8 cores: 59.99 seconds                66678 samples/s/core/GHz
16 cores: 31.40 seconds               63694 samples/s/core/GHz

1 core, 2 threads: 6 minutes 14.92 seconds   85352 samples/s/core/GHz
16 core, 32 threads: 26.31 seconds           76017 samples/s/core/GHz



I took the liberty of adding samples/s since that makes quantification easier. Looks like our systems have about 10% decrease in IPC from low cores to max cores.
 
Last edited:
  • Like
Reactions: psolord and blublub

superstition

Platinum Member
Feb 2, 2008
2,219
221
101
Which Ryzen thread is that? You don't have the data from my Lynnfield, my i7 Macbook Pro, or my 8370E at 5 GHz.
 

DrMrLordX

Lifer
Apr 27, 2000
22,031
11,614
136
Dude you are "No"-ing my claims that someone should not be dissapointed with, or yell "fail" at, a CPU with roughly 40% IPC gains from AMD's past architecture. That is: I'm saying no-one should be surprised if leaked benchmarks perform in that range. Since everyone has been talking about this forever.

I have never expected +40% IPC vs. XV would yield performance equivalent to Sandy/Ivy Bridge in an MT workload. Or rather, I expected that if it did, that AMD would have failed miserably. Many expected that but not everyone.

The WHOLE point of the thread is to be provocative!

It doesn't have to stay that way.
 

rvborgh

Member
Apr 16, 2014
195
94
101
just added your 8370E result. scores just the same per GHz as the FX8350.

edit: also added your results with a single thread running through each module.
 
Last edited:

superstition

Platinum Member
Feb 2, 2008
2,219
221
101
just added your 8370E result. scores just the same per GHz as the FX8350.

edit: also added your results with a single thread running through each module.
Thanks. Here is my Macbook Pro at 150 samples:

0:55.51
i7-4960HQ, late 2013 Macbook Pro, 16 GB DDR3L 1600, 10.9.5 "Mavericks", Blender 2.78a

https://s27.postimg.org/v20uikmqb/time.png

New Lynnfield results are coming, including The Stilt's two builds.

update:

Here are my CB R15 scores for the i7-4960HQ

multi: 668
single: 138
OpenGL (with Iris Pro): 29.75 fps
OpenGL (with GT 750M): 54.18 fps

CB R11.5

multi: 7.09
single: 1.59
mp ratio: 4.46x
Open GL (with Iris Pro): 43.28 fps
Open GL (with GT 750M): 48.94 fps
 
Last edited:
Status
Not open for further replies.