【Maxon】Cinebench R20 Benchmark Thread

Page 4 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

SPBHM

Diamond Member
Sep 12, 2012
4,843
69
126
#77
yes your score is more in line with normality, the MT is a little lower from the one on the list but the ST score much better, so the scaling looks closer to the other 2c/4t CPUs with fixed max clock.
 

Abwx

Diamond Member
Apr 2, 2011
8,939
298
126
#78
With 256 bit ops Intel hardware has double load/store bandwith to L1 cache vs 128 bit ops. Obviously L1 bandwith isn't problem for one thread as Zen can keep up but for SMT scaling extra L1 bandwith sure helps.
But with such higher LSU bandwith, not counting the alleged more FP ops/instruction, Intel score in MT should be vastly better clock/clock, wich is simply not the case, MT comparative scores are close to R15, wich say that your argument is BS.

What would be of interest is to measure power comsumption at fixed frequency/voltage for both R20 and R15 with say 8C/8T and then with SMT enabled.

Edit : Computerbase added a ton of CPUs at various frequencies in their chart :

https://www.computerbase.de/2019-03/cinebench-r20-community-benchmarks/
 
Last edited:

majord

Senior member
Jul 26, 2015
316
6
106
#79
Zen+ vs CFL st IPC and throughput/clock has definitely gone up for Intel.

at iso clocks AMD used to tie in ST and have a slight lead in MT typically on r15.. Now Intel has a clear +5-10% lead. It's not much but's there.

Conveniently the score's are all roughly identical to the clockspeed for an 8core16t Ryzen 2xxx, so this can be seen clear as day

Will be interesting to see how Zen 2 fairs with it's beefed up AVX perf
 

Markfw

CPU Moderator, VC&G Moderator, Elite Member
Super Moderator
May 16, 2002
18,006
1,835
136
#80
But with such higher LSU bandwith, not counting the alleged more FP ops/instruction, Intel score in MT should be vastly better clock/clock, wich is simply not the case, MT comparative scores are close to R15, wich say that your argument is BS.

What would be of interest is to measure power comsumption at fixed frequency/voltage for both R20 and R15 with say 8C/8T and then with SMT enabled.

Edit : Computerbase added a ton of CPUs at various frequencies in their chart :

https://www.computerbase.de/2019-03/cinebench-r20-community-benchmarks/
WOW, I would have to say for this type of load, the 2990wx really beats them all for perf/$$$. Stock is 4th place($1730), OC'ed is 3rd place, and 2nd is dual $5000 EPYC 7601, and first place is dual $14000 8180. Now the 7980xe is only a little more than the 2990wx at $1850, but even OC'ed to 4.6, it still looses, and by a fair amount. When the 2990wx is OC'ed, its not even close.
 

naukkis

Senior member
Jun 5, 2002
213
39
101
#82
But with such higher LSU bandwith, not counting the alleged more FP ops/instruction, Intel score in MT should be vastly better clock/clock, wich is simply not the case, MT comparative scores are close to R15, wich say that your argument is BS.

What would be of interest is to measure power comsumption at fixed frequency/voltage for both R20 and R15 with say 8C/8T and then with SMT enabled.

Edit : Computerbase added a ton of CPUs at various frequencies in their chart :

https://www.computerbase.de/2019-03/cinebench-r20-community-benchmarks/
Isn't that R20 speedup over R15 measured by users in thread be about 2.3?

https://software.intel.com/en-us/articles/how-embree-delivers-uncompromising-photorealism

Seems to working just like Intel says it should perform.

Zen just keeps up with more 128-bit SIMD resources but there just isn't enough free FPU time to have as good SMT scaling as with R15. But multi-core performance with R20 is still much better.
 

TheELF

Platinum Member
Dec 22, 2012
2,864
128
126
#84
Interesting. CBR20 is so much better-optimized than previous versions that SMT matters a lot less.
What do you mean matters?
It is a lot less,yep r15 used a lot less of the ryzen cores so it had a lot more to use on SMT,there where quite a few people that where saying just that.
Also 514 times 6 is 3084,holy crap at that 2318....just a small difference of 33% in ST when running all threads...well you got to do something to claim that sweet low TDP.
Yeah right, certainly a great option for multitasking when your individual threads drop 33% of the performance...
 

jpiniero

Diamond Member
Oct 1, 2010
6,500
295
126
#85
The 514 is probably a typo. Computerbase got 416/3034 on their 2600X.
 
Jan 12, 2005
15,829
3,968
126
#86
2996, 2600X on stock clocks with lowly 2133 ram speeds.

*edit - 3010 with ram at 2666. Small bump.
 
Last edited:

Thunder 57

Senior member
Aug 19, 2007
724
228
136
#87
The 514 is probably a typo. Computerbase got 416/3034 on their 2600X.
I'm going to go with typo. I ran it again and got 408 and 410 the second time. Sounds more realistic.
 

IEC

Super Moderator
Super Moderator
Jun 10, 2004
13,695
770
136
#88
The 514 is definitely a typo. It must be 414 instead.

I get 541 on an OC'd i7-8700K @ 5.2GHz AVX load. There is no way a Ryzen 2600X scores in the ballpark of a 5GHz 8700K in ST performance.
 

Thunder 57

Senior member
Aug 19, 2007
724
228
136
#89
The 514 is definitely a typo. It must be 414 instead.

I get 541 on an OC'd i7-8700K @ 5.2GHz AVX load. There is no way a Ryzen 2600X scores in the ballpark of a 5GHz 8700K in ST performance.
It was. I corrected it in the original post and included the two results I got just above your post, 408/410.
 

Thunder 57

Senior member
Aug 19, 2007
724
228
136
#90
Interesting. CBR20 is so much better-optimized than previous versions that SMT matters a lot less.
That's still nearly 29% better with SMT enabled. I wonder if the increase is similar with Intel. I would expect it would be a bit less, as Intel seems to have better IPC but benefit less from SMT. Anyone want to give it a shot?
 
Sep 4, 2016
82
49
61
#91
Ryzen 1700 @ 4000MHz RAM 3200MHz CL14: 3994 MT/403 ST
1700_4000_3200c14.png

RAM @ 3400 CL14 same sub-timings: 4008 MT/ 405 ST
1700_4000_3400c14.png
 
Feb 19, 2017
95
42
76
#92
Threadripper 1920X with 32GB DDR4 3333MHz 14-15-15
CinebenchR20_TR1920_4.1.png

5873 was at 4GHz
 
Apr 27, 2000
12,376
1,325
126
#93
What do you mean matters?
SMT on versus SMT off was a pretty small delta. That tells me that it kept the instruction pipelines packed well enough per thread that SMT had very little space to operate. Still not in the same league as something like linpack, but it got closer.

That's still nearly 29% better with SMT enabled. I wonder if the increase is similar with Intel. I would expect it would be a bit less, as Intel seems to have better IPC but benefit less from SMT. Anyone want to give it a shot?
Agreed, I would like to see someone with a 9900k or (even better) Skylake-X take a shot at the bench with and without SMT.
 
Last edited:
Apr 16, 2014
186
32
101
#94
Here's my SuperMicro H8QGi-F quad Opteron setup (its 48 K10 cores 3.2 GHz, DDR3-1333 memory)

7459cb multi (will run the single thread later)

199cb single core (3.8 GHz)

This new benchmark is a bit flaky... as it sometimes blows up halfway through the benchmark in the embree library. i'm going to clock down to 2.1 GHz to see if its hardware related (but this system is IBT rock solid through hundreds of runs).

For reference with previous benchmarks it scored 39.4 in Cinebench R11.5, and just over 3229cb in R15 at exact same settings. The previous benchmarks don't have this crash problem however.

Observations (vs R15 and 11.5):

1) doesn't seem to heat the processors up as much at the same GHz
2) is most definitely not as bandwidth sensitive on my multi socket system. Previously, if i had node interleaving off... scores would drop precipitously with R15. Now they only drop around 200 points or so. Perhaps R20 is properly using NUMA nodes?


Cinebench R20 at 3.2 GHz with node interleaving on.jpg
 
Last edited:
Oct 9, 1999
13,282
61
126
#95
This new benchmark is a bit flaky... as it often blows up halfway through the benchmark in the embree library. i'm going to clock down to 2.1 GHz to see if its hardware related (but this system is IBT rock solid through hundreds of runs).
It most likely could be your system is flaky. Not the benchmark.
 
Apr 16, 2014
186
32
101
#96
My system is unusual (most folks do not run a server grade $800 motherboard with quad opterons for their personal desktop PC), but its most definitely not flaky... i've run it through hundreds of runs of Intel Burn Test with no issues (with it soaking up over 1000 watts through the wall socket for hours and hours).

One thing i did notice about this benchmark vs the previous... it doesn't seem to heat the processors up as much as say R15 did...

[Edit - upped voltage from 1.175v to 1.20v now all stable - evidently this benchmark found an instability]

It most likely could be your system is flaky. Not the benchmark.
 
Last edited:

CHADBOGA

Golden Member
Mar 31, 2009
1,829
102
136
#97
Chill out man..... No need to get all bold with a larger font size all you had to do was ask normally. Anyway......here it is.
The large and bolded font was so other people could see it and it would act as a reminder.
 

Markfw

CPU Moderator, VC&G Moderator, Elite Member
Super Moderator
May 16, 2002
18,006
1,835
136
#99
The large and bolded font was so other people could see it and it would act as a reminder.
And what if I don't WANT to run it ? I started to on one of my 2700x's, and it was taking forever. We don't need to big font and bolded.
 
Feb 14, 2005
10,295
27
126
This thread is fail without a leader board for the different CPU classes in the OP.
 


ASK THE COMMUNITY