16C/32T CPUs helpful for games? M&B2 says yes.

[Mr Evil]

CPUs with many cores have been a thing for a long time, but games haven't really caught up - until now. Some time ago I noticed that Mount & Blade 2: Bannerlord uses a lot of CPU. And I mean a lot; it will fill my 5950X up with 32 threads, peaking at 90% total CPU use. It's easy to see how it could use that much, with 1000 combatants at once, all with AI and physics, and the game looks nice too. So, with a properly multi-threaded game available, I thought I would do a bit of benchmarking to see how it scales.

The game has a handy built-in benchmark that can record the FPS of every frame and save it to a file, allowing me to produce some confusing, vomit-coloured charts. The benchmark records "CPU FPS", which presumbaly is the FPS you would get if the GPU was infinitely fast. This is the number that we want here. It records "GPU FPS" too, which I ignored. I ran the benchmark at 4k, "Very High" settings. Full settings and system specs at the end.

I used the start /affinity command to limit the game to specific cores. As far as I can tell, the game actually spawns enough worker threads to fit the number of logical cores, not affected by CPU affinity, so running the game like this will be slightly slower than a real low-core-count CPU due to thread overhead. I can't easily tell how much the overhead is, but it's probably not enough to change the results.

The hardest part of this was trying to visualize the results in a way that makes the difference clear. I ended up with multiple charts, starting with frame times for every frame:

It's pretty obvious that 4 threads isn't enough, with much of the plot well off the top of the chart, which ends at 20ms (50fps). 8 threads is playable, staying below 16.7ms (60fps) most of the time. There continue to be improvements all the way up to 32 threads, especially in the consistency of frame times. That's easier to see when the chart is zoomed in to show only 1s of frames:

Now you can see that it takes 24 threads to get something that looks like a straight line.

Plotting every frame makes it hard to see the overall difference because there are so many points, so I also plotted FPS with a moving average window of 30 frames:

Now you can see that the scaling from 4 to 8 threads is really good, and there's still a fairly large jump from there to 12 threads. Returns quickly diminish above that, but performance is still increasing up to 32 threads, and it looks like it would keep scaling even further. It's really impressive; we need more devs that can write game engines like this.

And just for completeness, I also did some benchmarks with the CPU affinity set to alternating cores, to simulate a CPU without hyperthreading:

Physical cores perform better, but SMT does help quite a bit.

Conclusion: Moar cores is moar betterer.

Setup:
Game version 1.0.3.9860. Settings, as recorded by the benchmark itself:

Spoiler

Battle Size, 360 vs 640
Antialiasing Technique, Very High
Character Detail, Very High
Decal Quality, Very High
Enable Cloth Simulation, Very High
Enable Flora Sway, Very High
Environment Detail, Very High
Foliage Quality, Very High
Lighting Quality, Very High
Number of Ragdolls, Very High
Occlusion Technique, Very High
Particle Detail, Very High
Particle Quality, Very High
Postfx Bloom, Very High
Postfx Chromatic Aberration, Low
Postfx Dof, Very High
Postfx Grain, Very High
Postfx Hexagon Vignette, Custom
Postfx Lens Flares, Low
Postfx Motion Blur, Very High
Postfx SSR, Very High
Postfx SSSSS, Very High
Postfx Streaks, Low
Postfx Sun shafts, Very High
Postfx Vignette, Very High
Shader Quality, Very High
Shadow map Filtering, Very High
Shadow map Resolution, Very High
Shadow map Type, Very High
Terrain Quality, Very High
Tesselation, Very High
Texture Budget, Very High
Texture Filtering, Very High
Texture Quality, Very High
Water Quality, Very High

Not mentioned in that list is that the FPS limit was set to the maximum 360 (not sure if that has any effect on the benchmark), and the resolution was 3840x2160, 165Hz.

System specs:

Spoiler

CPU: AMD Ryzen 5950X. Undervolted with core optimizer.
RAM: 64GB DDR4 3600
GPU: Sapphire 7900XTX Nitro+ (stock)
OS: Windows 10 LTSC

 

[Zarax]

Now this would be an interesting benchmark to try with the newer Intel CPUs as well as 5800x3d.

 

[Cyber-Mav]

good it uses all 16 cores and 32 thread. does it matter? no because its a crap game.

 

[Zarax]

good it uses all 16 cores and 32 thread. does it matter? no because its a crap game.

Reviews have been generally good, why would you say it's crap?
I didn't buy it yet but I deeply enjoyed the first one.

 

[Mr Evil]

Now this would be an interesting benchmark to try with the newer Intel CPUs as well as 5800x3d.

Indeed. I would be particularly interested to see E-cores vs. P-cores.

good it uses all 16 cores and 32 thread. does it matter? no because its a crap game.

Au contraire, it is a superb game. The first one was great, and now they've improved it by adding things like seige weapons - sending a projectile from a trebuchet smashing through enemy troops is just so very satisfying.

 

[HangTime]

I've noticed similar (albeit only 12C/24T) with Ultimate Epic Battle Simulator and, surprisingly, Assassin's Creed Unity (yes, a 2014 game that utilises many cores).

I do think it's quite niche though, I'd say these sort of large battle sims with many units lend themselves well to parallelisation, but still for the majority of games I expect the scaling beyond 8c16t is minimal.

 

[Freddie1980]

Looks like it does a good job for delivering consistent frame timings rather then delivering substantially better averages.

Does your 5950x get better boost clocks when disabling cores?

 

[Mr Evil]

...Does your 5950x get better boost clocks when disabling cores?

Just did a quick test using CPU affinity again, and with 4 threads I get 2 cores running at around 4.9GHz. With 8 threads, 4 cores at about 4.8GHz. With all 32 threads, all 16 cores run at about 4.7GHz.

 

[Zarax]

Another game where this could be interesting is Ashed of the Singularity, the poster child of DX12.