Discussion Cinebench 2024 Released

[igor_kavinski]

Right-click and run as administrator? Usually runs fine just by double clicking on the EXE. Maybe Windows needs repairing with DISM and SFC commands?

 

[Muadib]

Thanks for the reply! Running as administrator was the first thing I did after it didn't run. My Windows 11 pro is relatively a new install. The only software I have on it is office, bitdefender, and printer crap. I will try DISM and SFC though. Thanks!

 

[AdamK47]

Check to see if the process is listed and hung in Task Manager. Check Event Viewer. The usual troubleshooting.

 

[Muadib]

Turned out to be the software to my new keyboard. I just started removing programs until I found the problem child. Thanks to you both!

 

[igor_kavinski]

Turned out to be the software to my new keyboard.

Which one so we may avoid it like the plague.

 

[Muadib]

Which one so we may avoid it like the plague.

The keyboard is the Steelseries APEX PRO TKL (2023). They call their software GG. It controls the RGB lighting. That's a problem for another day.

 

[Nothingness]

For those interested in knowing technical details about the benchmark (instruction mix, resource stalls, etc.): https://chipsandcheese.com/2023/10/22/cinebench-2024-reviewing-the-benchmark/

 

[alasdairvfr]

This one has my best GPU score




Best CPU score


Some folks are getting substantially higher CPU on the 7950x3d, not sure if eclk is being used, chiller, etc.

 

[Rigg]

Daily Tune

13700k - 6800c32, Fuma 2 w/Noctua fans, 90c limit, v/f point tuning, core usage 57/57/57/57/57/57/57/56, P cores 0-3 capped @ 56 4-7 @ 57, 8 core TVB -1 @84c -2 @88c, E cores @ 41

6900xt - 80 mV undervolt, 2150 memory, 2500 max clocks, -15% power

 

[StefanR5R]

For those interested in knowing technical details about the benchmark (instruction mix, resource stalls, etc.): https://chipsandcheese.com/2023/10/22/cinebench-2024-reviewing-the-benchmark/

Thank you for this link to useful information.

Here is some higher-level info from Maxxon themselves which I for one missed in @csbin's post #1:
https://www.maxon.net/en/tech-info-cinebench (Cinebench Technical Information)
Cinebench 2024 uses a new test scene with...

• 6 times the required computational effort compared to R23,
• 3 times the memory footprint: 6.5 - 8.5 GB RAM by the CPU renderer, 8 GB VRAM by the GPU renderer.
• Maxxon provide a separate Redshift GPU renderer benchmark for multi-GPU computers, with a considerably larger workload than the one in Cinebench.

What I still am missing on Maxxon's web site is the info on the maximum number of logical CPUs which the CPU renderer benchmark is able to utilize. Previous versions of Cinebench are known to be unable to use all hardware threads of modestly modern dual-socket computers with higher core count CPUs. I also quickly looked around the Cinema 4D and Redshift sections on Maxxon's web site and haven't found the equivalent information there either.

Well, I suppose somebody who is interested in high rendering throughput is not considering to render on CPUs in the first place. It seems there aren't any features missing if you go to GPU, are there? (The stepping stones to really high rendering throughput would be GPU rendering -> multi GPU rendering -> GPU rendering on multiple networked render nodes.) Since nobody (?) seriously renders on CPUs, there is no serious need for a high core count CPU renderer benchmark (in the Maxxon Cinema 4D context), is there?

Supposedly there is INi file that can turn off some SIMD stuff:

https://twitter.com/x/status/1699307923929784787

So playing with it could reveal "speed up" factor.

(Side note, I can't see twitter content.) Starting with Cinema 4D 2024, Maxxon made AVX2 support on CPUs mandatory. Likewise, they list AVX2 as a requirement for Cinebench 2024. (On Windows, not on Mac OS.) Therefore it is either impossible, or at least won't make sense, to switch off AVX2 usage in Cinebench 2024.

Edit:
And there is another thing which I have been wondering about: Does Cinebench have error checking built in? That is, are testers being informed whether their tested system actually computed as designed, or merely spat out random numbers? (You may not be able to tell just by looking at the rendered image.)

 

[JoeRambo]

(Side note, I can't see twitter content.) Starting with Cinema 4D 2024, Maxxon made AVX2 support on CPUs mandatory. Likewise, they list AVX2 as a requirement for Cinebench 2024. (On Windows, not on Mac OS.) Therefore it is either impossible, or at least won't make sense, to switch off AVX2 usage in Cinebench 2024.

It throws an error if cfg file contains

g_disableAVX2=true

So yeah, no longer possible to disable AVX2. Tho still possible to disable FMA, AES, PCLMULQDQ tho i doubt they are being used much.

 

[Markfw]

It throws an error if cfg file contains

g_disableAVX2=true

So yeah, no longer possible to disable AVX2. Tho still possible to disable FMA, AES, PCLMULQDQ tho i doubt they are being used much.

Is that true for Intel and AMD 2023 chips ? ie, with Intel having a hybrid approach and ecores may not support certain functions, maybe its only Intel that errors out ??

Asking, not saying.

 

[Timmah!]

This one has my best GPU score


View attachment 87696

Best CPU score
View attachment 87698

Some folks are getting substantially higher CPU on the 7950x3d, not sure if eclk is being used, chiller, etc.

I dont know how you got even that 2188 score, i got only 2119 with my 7950x. And i actually meddled with it, increased the MT clocks to 5,25/5,175 on CCD0/1 - it was less by default. Perhaps CB24 actually takes advantage of v-cache, this time around?

 

[randomhero]

I dont know how you got even that 2188 score, i got only 2119 with my 7950x. And i actually meddled with it, increased the MT clocks to 5,25/5,175 on CCD0/1 - it was less by default. Perhaps CB24 actually takes advantage of v-cache, this time around?

Yes it does. Go on chips and cheese site, they say that there is quite bit more trips to RAM, and that workloads can go over 8MB of cache per thread.

 

[StefanR5R]

Go on chips and cheese site, they say that there is quite bit more trips to RAM, and that workloads can go over 8MB of cache per thread.

I presume you took the 8 MB figure from this part:

In the end, Cinebench 2024 poses decent challenges to the frontend and backend. It has a more realistic instruction footprint than SPEC2017, which has no subtest with more than 12 L1i MPKI. Maxon has also addressed Cinebench R15 and R23’s small data-side footprint, which could be mostly contained by a 8 MB last level cache. High core count systems could be constrained by memory bandwidth, which happens across a lot of other well-threaded applications.

I am not so sure if this is meant to be per thread.
With CB R15, CnC measured…
…0.23 L3$ MPKI on i5-6600K (4c/4t, 6 MB L3$, 1.5 MB L3$/t)
…0.10 L3$ MPKI on R9 3950X (4x 4c/8t, 4x 16 MB L3$, 2 MB L3$/t)

With CB 2024, he measured…
…0.30 L3$ MPKI on i7-7700K (4c/8t, 8 MB L3$, 1 MB L3$/t)
…0.31 L3$ MPKI on R9 3950X (4x 4c/8t, 4x 16 MB L3$, 2 MB L3$/t)
…0.29 L3$ MPKI on R9 7950X3D (2x 8c/16t, 32+96 MB L3$, 2 or 6 MB L3$/t)

MPKI = misses per 1000 instructions

According to the R15 figures, going from 1.5 to 2 MB L3$/thread already halved the L3 cache misses. But he pointed out in the R15 article that "L3 misses are rare on both architectures, so cache performance is more important than DRAM performance here."

BTW, note the little difference between 1 vs. 2 vs. "2 or 6" MB L3$/thread figures of the CB 2024 measurements of cache miss rate vs. instruction rate. CnC says about CB 2024: "If an access is difficult enough to not be caught by L2, there’s a good chance it’ll miss L3 as well."

One could perform the following experiments in order to examine the influence of 3D cache on CB 2024 performance of Ryzen 9 7950X3D:
– Test one and the same 7950X3D computer twice, once stock, once with 3D cache disabled in the BIOS.
– Test one and the same 7950X3D twice but with CB restricted to 16 threads, once with CPU affinity to the first CCX, once with CPU affinity to the second CCX. (Caveat, this test setup would obviously double the available DRAM bandwidth per thread, compared to 32-threaded tests.)

 

[TESKATLIPOKA]

What is interesting is that M3 Max performs much better in CB 2024 than in CB R23.

59% higher score than 7840S in CB R23.
89% higher score than 7840S in CB 2024.
That's a big improvement. Why?

Now that I look at It, 13620H was on par with 7840S, now It scores 19% less in comparison.

 

[igor_kavinski]

That's a big improvement. Why?

Could be the higher RAM bandwidth that helps the CB2024 renderer.

 

[Timmah!]

Could be the higher RAM bandwidth that helps the CB2024 renderer.

Oh, you see, that may be other reason why i scored less, i run only 6000MHz RAM, when other people may run 6400+ or whatever is even possible nowadays with those latest BIOSes.

 

[Nothingness]

What is interesting is that M3 Max performs much better in CB 2024 than in CB R23.
View attachment 89040View attachment 89041
59% higher score than 7840S in CB R23.
89% higher score than 7840S in CB 2024.
That's a big improvement. Why?

Now that I look at It, 13620H was on par with 7840S, now It scores 19% less in comparison.

Part of it could be due to poor NEON implementation in Cinebench R23 (which was done with a tool to convert SSE/AVX to NEON it seems).

 

[Rigg]

R5 3600 @4.5 Ghz

Termalright Assassin King 120SE w/ Noctua NF-F12

 

[Thunder 57]

Well I redid mine and for some reason I got a 39% higher score on the RX 5700. 4515 vs 3246. I must have had it minimized and was doing other stuff while it was running last time. CPU score seems a bit low this time though.

 

[nicalandia]

What I still am missing on Maxxon's web site is the info on the maximum number of logical CPUs which the CPU renderer benchmark is able to utilize. Previous versions of Cinebench are known to be unable to use all hardware threads of modestly modern dual-socket computers with higher core count CPUs. I also quickly looked around the Cinema 4D and Redshift sections on Maxxon's web site and haven't found the equivalent information there either.

Cinebench 2024 Can not use more than 256 Threads. It used half of the threads on a 2S Bergamo.

 

[tamz_msc]

This edition, even though memory bound, is a much better benchmark among renderers. Regardless of single core or multi core, all of the CPU works in tandem to render a SINGLE tile, unlike the previous versions. It is closer to how MT scaling works in the real world. Kind of like Geekbench 6.

 

[tamz_msc]

According to Techspot/HUB, scaling is 48x for the 7980X in CB 2024. Anandtech shows 52x scaling in CB R23.

 

[Timmah!]

This edition, even though memory bound, is a much better benchmark among renderers. Regardless of single core or multi core, all of the CPU works in tandem to render a SINGLE tile, unlike the previous versions. It is closer to how MT scaling works in the real world. Kind of like Geekbench 6.

It is better simply on basis of including GPU and shows how much faster those are at this kind of tasks.