Discussion Cinebench 2024 Released

[csbin]

Evaluate your computer's hardware capabilities | Cinebench from Maxon

Cinebench is a free cross-platform test suite that evaluates your computer's hardware capabilities. Get a more accurate measurement of Cinema 4D's ability to take advantage of multiple CPU cores and modern processor.

www.maxon.net

System Requirements​

Operating Systems

• Windows 10 Version 20H2 or higher or Windows 11 for x86/64 hardware
• Windows 11 for arm64 hardware
• Apple macOS 12.6+ (Monterey) or 13.3+ (Ventura) for Redshift GPU support
• macOS 11.7.7+ (Big Sur), if you want to run CPU tests only

Minimum Requirements (Windows)

• 16 GB of RAM
• 64-bit Intel or AMD cpu with AVX2 support or Windows 11 on Snapdragon compute platform or ARM v8.1 64-bit CPU
• NVIDIA GPU with CUDA compute capability 5.0 or higher and 8 GB VRAM, or
• AMD "Navi" or "Vega" GPU or later with HIP capability and 8 GB VRAM or more (see GPU list below)

Minimum Requirements (macOS/Intel)

• 16 GB of RAM
• 64-bit processor with SSE4.2 support
• AMD "Navi" or "Vega" GPU or later with 8 GB VRAM or more (see GPU list below)

Minimum Requirements (macOS/Apple Silicon)

• 16 GB of unified memory; CPU rendering works on Apple Silicon machines with 8 or 12 GB of memory too, but OS memory paging can degrade the performance compared to machines with 16 GB of memory
• Apple “M” Series (M1, M1 Pro, M1 Max, M1 Ultra, M2, M2 Pro, M2 Max, M2 Ultra)

 

[Markfw]

My first run, 7950x is 2004. Now its installing on my 9554. How do you remove the EVGA precision X crap ?

 

[Markfw]

Not done yet, but already several passes and scores about 5000 for my 9554 64c/128t. But why 2 boxes for 128 threads ?

 

[JoeRambo]

Not done yet, but already several passes and scores about 5000 for my 9554 64c/128t.

That's suprisingly good. 4090's at fixed 2400Mhz is scoring 33000, so CPU being slower just 6 times is quite an achievement. What is the power usage during the run, 4090 does it with sub 200W.

 

[Markfw]

That's suprisingly good. 4090's at fixed 2400Mhz is scoring 33000, so CPU being slower just 6 times is quite an achievement. What is the power usage during the run, 4090 does it with sub 200W.

Here is the multi, after the single core run.18.39 multi

 

[eek2121]

Still waiting on Single Core, will edit this post when complete:

Initial Run:

SC: 123
MC: 2055
GPU: 37332

Ryzen 7950X, 64gb DDR5 6000 32-38-38-38, RTX 4090 FE.

RAM set to DOCP tweaked and some timing values changed. FCLK is 2133mhz. 1:1 mode. GDM off. Will play with it more later. I also have other systems I can run it on (2x Cezanne, Macbook Pro, later on a 5950X, possibly a 3600XT or 3900X)

Screenshot:

 

[Markfw]

The multi. 67.87 speedup. on a 128 thread box.

 

[Markfw]

3080TI 21,498. This seems more about a GPU rendering score than a CPU score.

 

[Det0x]

7950x3d @ 24/7 daily clocks
8000MT/s CL32 stable memory settings


And a other run with hwinfo open
CPU did manage to hit a whopping 160watt lol

Now back to starfield

 

[Brunnis]

Asus Zephyrus G14 with Ryzen 9 7940HS (Phoenix) + RTX 4070 mobile:

GPU: 11568
CPU (multi): 957
CPU (single): 108

 

[ondma]

8700k @ 4.9:
Multi: 532
Single: 78
GPU (1060 6gb); 2502

So looks like Zen 4 is about 65% increase performance per core, compared to Coffee Lake. Hard to read from the screen shots, what frequency were the 7950s running at?

 

[Hitman928]

Here's my 5900x for comparison (all cores run at ~4.1 GHz during the multi core test and single core runs 4.8 GHz - 4.9 GHz).

 

[Gideon]

I'm actually surprised that Cinebench 2024 actually benefits from 3D cache, unlike previous versions.

(usually my 5800X3D is slightly slower than 5800X) due to ~200 Mhz lower Turbo speed, bu it was consistentl faster here (i wonder if the same is true for the 7950X vs 7950X3D scores)
This is probably because the dataset is larger and finally spills over the 32MB L3 cache:

https://twitter.com/x/status/1699171657573884175

BTW, I made the screenshot slightly too soon, the final score was 875, but I don't have the correct screenshot at hand on the work machine.


Speaking of for the work machine, - gotta give credit to apple for their 0 variability (Pro and Max have the exact same CPU)

I'd say it's even a tad slow for a 10 Core laptop (8+2) , but totally understandable, considering no SMT and excellent perf/watt (It was essentially silent on the whole 10 minute run).

Next I'll try my alternative work-machine (for windows-only projects) It's a a Lenovo T14s wih a Ryzen 6850U. My guess it, it edges out the 9880H, but won't quite catch the M1 Pro.

 

[Gideon]

Next I'll try my alternative work-machine (for windows-only projects) It's a a Lenovo T14s wih a Ryzen 6850U. My guess it, it edges out the 9880H, but won't quite catch the M1 Pro.

Well I was too optimistic. The 15W TDP massively limits the CPU Clocks, the end result is 472 points.

So the M1 chip really is quite impressive in perf/watt, considering it reaches 791 in a similar power envelope.

Now obviously there is a rather large process node deficit as well. Zen 4 should be a lot more competitive, even in "U" form, but alas I don't have access to any (and they were impossible to find a couple of months back when the aforementioned laptop was bought)

 

[HoveringStyle]

If that's really just 15W that's very impressive, the plain M1 has a "TDP" of 20W and the M1 that scores 791 is the Max which used 34W of package power when running CBR23 MT.

On the other hand, here's my laptop (Lenovo Legion i5 g7), it's not limited to 45W, more like 80W+ during the whole duration of the benchmark, not very efficient.

 

[eek2121]

If that's really just 15W that's very impressive, the plain M1 has a "TDP" of 20W and the M1 that scores 791 is the Max which used 34W of package power when running CBR23 MT.

On the other hand, here's my laptop (Lenovo Legion i5 g7), it's not limited to 45W, more like 80W+ during the whole duration of the benchmark, not very efficient.

View attachment 85425

Unsure where you got those M1 TDP numbers from, but they aren’t correct. First, Apple hasn’t released any official TDP, but they have listed power limits for some of the chips:

M1 - 39W
M2 - 50W
M2 Pro - 100W

Anandtech and other outlets confirm the following:

M1 Pro - ~70W
M1 Max - ~100W

Note these are maximums, not the numbers CB consumes. Also, power limits appear to vary by device. The Macbook Air has a lower power limit than a mini, for example.

I am also not claiming they aren’t efficient. I use a Macbook Pro with an M1 Pro and it is very efficient.

 

[HoveringStyle]

Those power limits are for the whole soc (CPU+GPU), the numbers I mentioned were measured on multi-thread cpu workloads by this site.

M1 max on Cinebench R23 MT
M1 on a mac mini

 

[igor_kavinski]

Cascade Lake beating a 2990WX in MT but failing to overtake it in ST.



This task manager screenshot shows why Cascade Lake posted a lower ST score:



Stupid Windows Server 2019 scheduler bouncing the CB thread between the first real core and the last core's HT core. How moronic is that?

 

[Abwx]

Cascade Lake beating a 2990WX in MT but failing to overtake it in ST.

View attachment 85449

This task manager screenshot shows why Cascade Lake posted a lower ST score:

View attachment 85450

Stupid Windows Server 2019 scheduler bouncing the CB thread between the first real core and the last core's HT core. How moronic is that?

If there s no more than 1 thread the last HT core is still a full core since all ressources of this core are dedicated to a single thread.

 

[igor_kavinski]

If there s no more than 1 thread the last HT core is still a full core since all ressources of this core are dedicated to a single thread.

Hmm...May have to test that by turning HT off.

 

[Mopetar]

I'd say it's even a tad slow for a 10 Core laptop (8+2) , but totally understandable, considering no SMT and excellent perf/watt (It was essentially silent on the whole 10 minute run).

I'm running it on my MBP (M1 Max) right now and it's eerie at how quiet it's staying. Even the multi core test wasn't spinning up the fans and I'm struggling to hear anything from it over the top of the humidifier across my room that's running on its nighttime quiet setting. The results aren't that interesting since it finishes about where its expected to. Single core was an even match, multi core about 4.5% better than the listed score.

I wish that Activity Monitor was a bit more informative with regards to power use. For reference the power usage for the single core run is about half of what Final Cut Pro is listed as in the last 12 hours which I was using to edit and render a 30 minute video over the course of about an hour. Of course that's using more of the dedicated hardware than the CPU cores on the SoC so I don't know how fair it is. The multi core test used about 4.6x the energy as the single core test, but I don't know how precise the reporting is.

I don't know if the graphics test will show anything useful, but one of the few things that gets the fans to kick on are running games.

 

[Nothingness]

I worked on a program that heavily uses NEON SIMD instructions and was multithreaded. Within seconds the fans of my MBP M1 (from 2020) turned on. So even CPU-only tasks need to have fans on

This makes me wonder how aggressive is CB in using SIMD on Arm (and on x86).

 

[JoeRambo]

This makes me wonder how aggressive is CB in using SIMD on Arm (and on x86).

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.

 

[Mopetar]

The graphics test didn't get the fans to ramp up either. The CPU use was only ~5% across the system during the run so there was probably plenty of thermal headroom.

 

[igor_kavinski]

The graphics test didn't get the fans to ramp up either. The CPU use was only ~5% across the system during the run so there was probably plenty of thermal headroom.

Maybe in the next CB, they decide to include a heterogeneous test using all available CPU/GPU cores. The decision to use a highly parallel algo to maximize GPU parallelism has the unfortunate side effect of making all existing CPU cores look really wimpy.