How Well Modern Linux Games Scale To Multiple CPU Cores

[mlau]

Quite shocking how bad the scaling of even modern games (escp. Talos!) is.

The Talos results are probably the nvidia drivers' threaded optimizations at work, which are default enabled with 378.13.
The whole test should have been run with __GL_THREADED_OPTIMIZATIONS=0 set, to get the driver interference out of the way. As is, the results should be taken with a large grain of salt.

[duby229]

Doing these kinds of benchmarks at resolutions like 1440p with high quality settings? Michael you do realize that unless you're using really low end CPUs these are essentially GPU bound scenarios? While being rather unrealistic, most CPU benchmarks for games are run at 1080p or below with low quality settings specifically to make sure that the games are CPU and not GPU bound during testing. I'd recommend that you re-do these tests at 720p and low settings to see how well they scale when they're properly CPU bound.

I -only- agree with you in cases where the actual gameplay results(1) are literally right next to the artificial results. That way folks will easily be able to tell what they will actually get vs what an artificial configuration nobody plays at shows.

(1) results from a configuration like you intend to play from.

[indepe]

Michael, thanks also for the CPU and GPU usage graphs, even if I'm not currently commenting on those: they require more thinking.

[R00KIE]

Now you only have to test if (for that board+bios) it makes any difference to leave everything enabled on the bios and bring cores offline on linux (# echo 0 > /sys/bus/cpu/devices/cpu?/online)

[Luke_Wolf]

Doing these kinds of benchmarks at resolutions like 1440p with high quality settings? Michael you do realize that unless you're using really low end CPUs these are essentially GPU bound scenarios? While being rather unrealistic, most CPU benchmarks for games are run at 1080p or below with low quality settings specifically to make sure that the games are CPU and not GPU bound during testing. I'd recommend that you re-do these tests at 720p and low settings to see how well they scale when they're properly CPU bound.

No, they do it because they are stupid or are trying to play games with the numbers. Games are absolutely not valid synthetic benchmarks, do not pretend they are. What games are when benchmarked using realistic profiles are completely valid real world benchmarks. However you need to understand that while yes games become GPU bound when at extreme settings in some games as long as the CPU is fast enough to feed it, that does not mean games become magically CPU bound when you run at minimum settings. What being bounded means is that component X is taking so long to complete it's task that for the entire run of the benchmark all of the other components are waiting on this one component to finish, and thus aren't effecting performance. When you un-bound it, it does not magically become bounded by another component as games are very complex programs that touch everything, and it is much more likely to have a cascade of things slowing it down (See the effect of memory on performance) as opposed to one thing when you're not intentionally increasing load to isolate out a single component.

Furthermore just because you get zOMG Frames in one game, does not mean that this performance translates to other games, Doom for example has an extremely different load profile when compared to Deus Ex Mankind Divided, and the besides of which is there's plenty of better ways to actually test CPU performance and how well the processor scales... like code compiling.

[chimpy]

On current steam survey 46.19% of people have 2 cores and 47.74% have 4 cores.

So games are optimized for the mass and that ~94% of userbase, which is kind of normal and expected

Agreed. By just targeting 2 and 4 cores developers can satisfy 93.93% of gamers on steam. But yea since most of us do more than just gaming (future games will use more cores by the way), Ryzen is an awesome processor

[Luke_Wolf]

The question is: what do you hope to learn from these tests?
IF your goal is to learn something about how well the CPU code parallelizes over multiple CPUs, what you suggest makes sense. But if your goal is to learn "what's the best 'CPU/dollar' value for games played the way I like to play them", Michael's tests may well make more sense.
Since I don't play games, I've no opinion on the subject --- but my guess is that more people play games at 1440P HiQ than at 720P LoQ?

For the set of people who would spend $330+ on a processor (which is not all gamers by far), the only people likely to be running them at LoQ at 720p as opposed to 1080p HiQ are "Pro Gamers" who think that running their games at 300fps grants them some sort of advantage (Hint: It doesn't)

Agreed. By just targeting 2 and 4 cores developers can satisfy 93.93% of gamers on steam. But yea since most of us do more than just gaming (future games will use more cores by the way), Ryzen is an awesome processor

Doom currently uses between 12-16 Threads depending upon the zone on my Ryzen 1700, so... it's not that far out from that being the norm.

[smitty3268]

For the set of people who would spend $330+ on a processor (which is not all gamers by far), the only people likely to be running them at LoQ at 720p as opposed to 1080p HiQ are "Pro Gamers" who think that running their games at 300fps grants them some sort of advantage (Hint: It doesn't)



Doom currently uses between 12-16 Threads depending upon the zone on my Ryzen 1700, so... it's not that far out from that being the norm.

Using threads doesn't mean the same thing as needing cpu cores.

16 threads at 10% cpu use per core is the same thing as just needing 4 cores at 40%, after all, and it's possible there would be no performance change from that.

[labyrinth153]

I was pretty psyched when my imac came with i5-6600. Its the cpu I wanted for single core performance.

[Luke_Wolf]

Using threads doesn't mean the same thing as needing cpu cores.

16 threads at 10% cpu use per core is the same thing as just needing 4 cores at 40%, after all, and it's possible there would be no performance change from that.

Well when we're talking pure raw FPS sure, however... for gaming usage more core at vastly lower utilization for the same number of threads is better than fewer cores at higher utilization because games have very hard latency requirements (Which again has nothing intrinsically to do with FPS). The more you can thread out to cores, the less threads get in each other's way, and therefore the less microstutters you have to deal with.

Also Doom is like 30-50% per core (Max Settings without Motion Blur or Depth of Field on a R9 285)