Discussion Cinebench R23 renders slower than R20

First, I know we can't compare R20 and R23 scores -just making clear on that.

In R23, render test is set to 10 minutes by default. And if single rendering is finished sooner, cycle will repeat until 10 minutes pass. That is reasonable time at which CPU temperature stabilizes enough and we get real life scores. R20 can do the same if we set time manually to 600 seconds -which I did.
I won't talk about scores for my CPU. Let me show how long it took to finish each test on both Cinebench versions:

R23 MultiCore: 632sec for 17 passes
R20 MultiCore: 634sec for 18 passes (17 passes would be 634x17/18=599sec)

R23 SingleCore: 1002sec for 2 passes
R20 SingleCore: 935sec for 2 passes

As we can see, R20 is in both cases (Multi & Single core) almost exactly 6% faster.
I must say that was quite a surprise for me.. because a while back (after R23 was introduced) there was mentioned that R23 is using more recent compiler (which assumes better/faster). But as it seems, that's not the case -newer "better" compiler (that Maxon is using) is slower.

Now I wonder.. is that only the case for AMD CPU's (which I have), or is that also the case for Intel CPU's? Anyone interested to fire up Intel CPU on Cinebench R20 and R23, to confirm?
 
  • Like
Reactions: usiname
My guess would be that the newer version of the render engine is more accurate than the one used in R20, so whilst the newer engine is complied to take better advantage of the hardware the developers have used that headroom to increase output image quality rather than increase speed. In the actual rendering suite you would likely be able to extract a performance boost by adjusting settings down to match the image quality of the older version.

I don't think this is an issue for AMD Cpu's- when you compare scores of AMD vs Intel hardware in R23, the AMD hardware is scaling as expected vs Intel so this regression in speed is likely universal. It's also possible that the developers have intentionally increased the time required slightly to make the test more challenging for current and future cpu's.
 
My guess would be that the newer version of the render engine is more accurate than the one used in R20
-yes, that's possible... I've compared both R20 & R23 resulting images at pixel levels and differences are not even worth to mention. Ok, R23 rendering might be "better" in case of different scene.
But that's not the reason why I started this thread.
Cinebench is benchmarking tool, where we expect it to be fair/objective for all supported CPU's. And in that sense, I was hoping someone would confirm that R23 is slower on Intel CPU's as well.

It's also possible that the developers have intentionally increased the time required slightly to make the test more challenging..
No, benchmark time is the same in both cases (10 minutes) -it's just, R23 is rendring slower during that time.
 
Cinebench is benchmarking tool, where we expect it to be fair/objective for all supported CPU's. And in that sense, I was hoping someone would confirm that R23 is slower on Intel CPU's as well.

I get where you are coming from - looking at the scores in R23 there doesn't appear to be any particular bias towards Intel.

No, benchmark time is the same in both cases (10 minutes) -it's just, R23 is rendring slower during that time.

From what you are saying, R23 completes less runs in the same time than R20 - that means that it takes longer to render a frame in R23 than in R20. Cinebench is just a time based test - lower time to complete a scene = higher score.

In terms of how Intel and AMD stack up against each other in the two versions, you can compare scores here:
https://www.cpu-monkey.com/en/cpu-intel_core_i9_10900k-1139

I am specifically looking at the i9 10900K - whilst you cannot compare scores directly between R20 and R23, you can compare the delta between scores to see if there is any relative change.

Comparing the multi core scores between the Ryzen 9 3900X and the i9 10900K gives the following:
Core i9 10900KRyzen 9 3900X% Diff
Cinebench R2063997178R9 12% Faster
Cinebench R231687518424R9 9% Faster

It looks like R23 tends to run comparatively slightly better on Intel - although it doesn't appear to change the overall positioning between the cpu's in any meaningful way. The 3900X is still faster than a 10900K.

Interestingly, looking at some other parts it appears that the older Threadripper parts (e.g. 1950X and 2950X) appear to gain quite a bit of ground in R23 vs R20 and perform closer to where you would expect them based on core counts.
 
Thank you again for participating.
Now we're getting closer to what I had in mind. Scores aside.. my question would be (from your post): if someone ask you, how much faster is 3900X compared to 10900K, what is the answer? 12% or 9%?

Or if I measure times to complete benchmark, why is my CPU 6% slower on R23? Yes, because benchmark code has changed. I only wish to know if Intel CPU's are also 6% slower on new code -by measuring time to complete benchmark.
 

usiname

Prominent
BANNED
Feb 27, 2020
92
20
535
There is not difference on pixel level, the image is same in every pixel and color. Intel is too slowed down, less than AMD, but still slowed. The "new" algorithms are just extra weight for x86 while the real optimisations are only for crapple's m1. M1 now is 30% slower than AMD in same power budget, imagine +6% and slower algorithms of R20 on m1 on 7nm, then you will see how the power "effective" arm can't pull same score with double the power. Now cinema4d has new sponsor so get used to it.
 
There is not difference on pixel level, the image is same in every pixel and color.
While scene is 100% equally rendered for R20 and R23, there are differences in pixel colors. One can see where differences are (at which pixels and amount of difference) by stacking R20 and R23 resulting images as layers into image editor and apply "Difference" blend mode on top layer -I'm using Corel PSP (Adobe PS should have similar option).
Differences are VERY small.. pixel example: R20 RGB=238 224 218 and R23 RGB= 239 224 219 -and don't occur on entire image. To me, that difference is more within margin of error -but is there.

Intel is too slowed down, less than AMD, but still slowed.
-thank you for confirming!
It really isn't about which CPU is faster or slower.. I'm only curious about R23 as benchmarking tool.