Test: CPU Core count and RAM scaling in ACC, AC1 and R3E

Initial testing. 720p. Using the same format as Dan. His replay file. Looking back from lead car for a full lap at start of the race.

9900k @ 5.2ghz all core. 47x core HT off.
15-15-15-32 CR1 3600mhz DDR4
Stock 2080ti

22-08-2019, 09:29:00 AC2-Win64-Shipping.exe benchmark completed, 16783 frames rendered in 110.672 s *CPU impact maxed out but GPU impact minimized by disabling or setting LOW to gpu hogging resources. I'll add screenshots later*
Average framerate : 151.6 FPS
Minimum framerate : 104.3 FPS
Maximum framerate : 208.5 FPS
1% low framerate : 104.4 FPS
0.1% low framerate : 99.3 FPS

22-08-2019, 09:34:54 AC2-Win64-Shipping.exe benchmark completed, 12793 frames rendered in 109.984 s *EVERY SETTING AT EPIC OR MAX POSSIBLE*
Average framerate : 116.3 FPS
Minimum framerate : 87.6 FPS
Maximum framerate : 148.2 FPS
1% low framerate : 87.4 FPS
0.1% low framerate : 76.5 FPS

This game seems heavily GPU bound and could likely use a fair amount of optimization. Whether it's design decisions or a engine limitations, there's something not right here.

For fun, I ran it in 3440x1400 with max settings:

22-08-2019, 09:42:20 AC2-Win64-Shipping.exe benchmark completed, 8526 frames rendered in 125.015 s This one went on for a few seconds longer as I got distracted but it's not going to change the results much.

Average framerate : 68.1 FPS
Minimum framerate : 52.8 FPS
Maximum framerate : 82.4 FPS
1% low framerate : 52.2 FPS
0.1% low framerate : 30.5 FPS

I can see why they walked away from RTX. There's little to no chance of making it run at any level of acceptable performance. It'd be just another PR mess to deal with.

What @kunos should do is provide a detailed optimization guide if they're going to limited to these levels of performance. Knowing unreal engine from other games, I'm sure certain settings have a much higher performance tax than they're worth. It's upto the dev to share those details.
 
Last edited:
We shouldn't correlate your scores with mine, but your first test is way too close to my results. It would suggest there is a cpu bottleneck after all. The 2080TI should be more than 5% faster than my 1080ti if there actually was a GPU bottleneck. Or maybe there was in my case, but it only needed a 5% stronger GPU to stop being gpu bound.

Only you can find out now Robert. By underclocking your 2080ti
 
We shouldn't correlate your scores with mine, but your first test is way too close to my results. It would suggest there is a cpu bottleneck after all. The 2080TI should be 25~30% faster in your first test.

Only you can find out now Robert. By underclocking your 2080ti

My first test settings are quite a mixed bag with lots of things being set to low. I'll put up screenshots tonight on the custom settings.
 
This game seems heavily GPU bound and could likely use a fair amount of optimization. Whether it's design decisions or a engine limitations, there's something not right here.

So that we're clear, you're saying that you think ACC is GPU-bound running at 720p with a 2080 Ti? I must be misunderstanding something here as I can run comfortably at 1440p with mainly high or epic settings and a full grid of AI with my GTX 1080.
 
So that we're clear, you're saying that you think ACC is GPU-bound running at 720p with a 2080 Ti? I must be misunderstanding something here as I can run comfortably at 1440p with mainly high or epic settings and a full grid of AI with my GTX 1080.
Well to be fair his CPU limit seems to be between 150-170 fps but before that, his 2080 ti, even at 720p can't reach more than these fps numbers.
I doubt you can reach 160+ fps on your 1080 at 1440p?
 
So that we're clear, you're saying that you think ACC is GPU-bound running at 720p with a 2080 Ti? I must be misunderstanding something here as I can run comfortably at 1440p with mainly high or epic settings and a full grid of AI with my GTX 1080.

I'd recommend downloading the replay above, using the same parameters to come close to a like for like scenario.

As you can see in my 3rd example, that's a 3440x1440 output. The reason I asked for Kunos to do an optimization guide is because unless someone goes through each iteration, there's no way of knowing what the hit in performance is between "high and epic" and so on.
 
I doubt you can reach 160+ fps on your 1080 at 1440p?

Definitely not! However, it's not all about framerates. Being able to race with a full grid, great graphics and no stuttering or lag is more important to me than hitting specific numbers. Any game is going to reach a limit if pushed hard enough - it's all about whether or not that limit still produces acceptable gameplay on your system. I was just a little surprised to see RobertR1 describing his results as "heavily GPU bound". Apart from things like flight sims, there are very few modern games which are not ultimately GPU bound. It's just a natural consequence of the desire for ever more realistic graphics.

The reason I asked for Kunos to do an optimization guide is because unless someone goes through each iteration, there's no way of knowing what the hit in performance is between "high and epic" and so on.

Whilst I agree in principle, apart from stating the obvious (that epic is going to be more of a hit than high), it's very difficult to quantify what effect individual settings will have on a specific system with the wide variety of hardware, software and driver combinations that are possible. Over on the official forum, I seemed to be getting better performance (with higher settings) than someone with much better hardware than mine - optimization is as much about the system as it is about the game itself. Apart from Kunos stating which features put the biggest load on the GPU or CPU, the only real results for an individual's system come from taking the time to try out different settings. Years of tweaking flight sims has shown me that one person's wonder settings can often have little, or no effect for someone else
 
ACC:
upload_2019-8-22_15-30-51.png


Did some testing in 1024x768 and 70% resolution scale.
Gpu never went above 75% load.

I used the same settings I'm using to always achieve 60 fps with slight drops at race starts. Both my 1070 at 3440x1440 and my old i7 2600k @ 4.4 GHz are reaching their limit with these settings.
I only have 12 cars visible though and used Dan's Replay with the same procedure!

Conclusions:
  1. Hyperthreading does a good job when having less than 4 real cores but another real core always gives better results!

  2. At 2 cores without HT it becomes really really bad...

  3. 2 cores with HT are still pretty good.

  4. No idea why HT on gives less fps with 4 cores... Gonna have to invest this further! In witcher 3 HT works wonders and it basically only uses one thread more than ACC.

  5. I will test a live AI scenario with HT on vs off soon since we got told that live AI uses more threads.
 
Last edited:
When the game is given enough cores to run optimally, HT becomes detrimental because Windows by default is always juggling processes between all threads. Among those threads, you now have HT threads instead of always having real cores. If you use process lasso, you can mandate ACC to only use specific threads (chose the real cores), and the HT impact should be reduced.
 
Btw... Has anybody any knowledge about what the min fps in the afterburner/rtss benchmark are?
Numbers are indicating something like "10% low"...

I really like that value because it instantly tells me where I should set my fps limiter at.
With gsync I found that a gpu limit is doable with fluctuating fps but if it's your cpu you get massive hiccups.
 
When the game is given enough cores to run optimally, HT becomes detrimental because Windows by default is always juggling processes between all threads. Among those threads, you now have HT threads instead of always having real cores. If you use process lasso, you can mandate ACC to only use specific threads (chose the real cores), and the HT impact should be reduced.
upload_2019-8-23_11-12-21.png


I checked temperatures first to know that 0&1 are on the same physical core. No idea though which one the virtual and physical core is...
Probably it's like this: HT takes one physical core and simply splits it into two virtual cores that are identical so there's no difference in theory?

Anyway, this is pretty interesting!

Not that it matters for me at all since running on my "looks okay but not great" settings for 3440x1440 on my GTX 1070 a normal benchmark results in this:

upload_2019-8-23_11-17-48.png
 
Probably it's like this: HT takes one physical core and simply splits it into two virtual cores that are identical so there's no difference in theory?
Yes that somewhat seems correct.


The Intel Core datasheet sates each core has two physical execution units sharing the same core resources. As such the OS sees two physical cpus. This information is passed down to the OS (topology awareness) and is its job to know which cores share resources. A good job would mean avoiding tasking busy cores with additional jobs. But, windows has 100's of processes competing for resources. You only have 4 real cores.


The fact we see performance drops with HT ON means that busy cores (running for example the ACC 3~4 main threads) are still receiving jobs on execution unit 2 of each core (that will trigger when execution unit 1 stalls every so often). When execution unit 1 (acc stalled thread) is ready to work again it must wait for execution unit 2 to complete. At this point windows juggles everything into another more empty execution unit from another core, which incidentally means it must move that specific core cache contents into the cache of another cores. (This is where the Ryzen higher inter-core latency causes penalties in cache sensitive software, games being the worst offenders).


With HT off, windows is tricked into keeping ACC threads in the same cores, because now it doesn't see "empty" threads elsewhere. Empty is in quotes "" because in reality they are not empty. Moving jobs to them will affect the jobs they already have. As an added benefit, ACC busy cores are no longer receiving random jobs from the other 100 windows processes, because those are marked as 100% busy. Win-win.


The takeaway is that core counts (threads) always trumps latency penalties, up to the point where there is enough threads to satisfy the game, after which the HT latency starts to be noticed.


https://dev.xdevs.com/attachments/download/302/324641.pdf
 
Last edited:
If you're on a 6 physical core or above system and mainly want to optimize performance in sims, this would be my optimization flow:

- Disable HT
- Find your max stable frequency at a fixed voltage that keeps your temps in check without AVX offset. Use cinebench r20, realbench 2.5.6 and blender benchmark to test (all use AVX to a reasonable degree without being power viruses like p95 avx)
- Then lower your vcore until you're unstable and bump it up slightly like .005 or .010v
- Then start your memory tuning beyond XMP. This will help a in minimum frame rates and really pay off in VR. https://github.com/integralfx/MemTestHelper/blob/master/DDR4 OC Guide.md

Don't get hung up on comparing your CPU or RAM to someone else. There's too much variance in silicon lottery, mother boards and ram bin lottery that makes this is a pointless exercise.
 
PSA: The 9600k had a price drop so quite cheap now. If your main goal is sims, pick one of these up and bump it above 5ghz all core without much effort. You'll get 6 physical very fast cores. You don't need any high end cooling or expensive motherboards with beefy VRM's either.

Should last for a few years. Get 3200mhz ram atleast.
 
Last edited:
PSA: There 9600k had a price drop so quite cheap now. If your main goal is sims, pick on of these up and bump it above 5ghz all core without much effort. You'll get 6 physical very fast cores. You don't need any high end cooling or expensive motherboards with beefy VRM's either.

Should last for a few years. Get 3200mhz ram atleast.
Friend just did that a few weeks ago. Seems like he got the lowest prices in a few years: (Germany)
9600k = 219€
16gb gskill 3200 cl16 = 72€
Gb z390 elite = 168€
Be quiet straight power 11, 650W = 99€

He bought a cm master box (full mesh case) last year and threw in a thermal right macho Rev b and 2x silent wings 3.

His system runs 5.0 GHz all core with 1.29v vcore (1.27 vrvout) and it doesn't go above 70°c.

Crazy system!
 

Latest News

Online or Offline racing?

  • 100% online racing

    Votes: 96 7.8%
  • 75% online 25% offline

    Votes: 130 10.5%
  • 50% online 50% offline

    Votes: 175 14.2%
  • 25% online 75% offline

    Votes: 348 28.2%
  • 100% offline racing

    Votes: 480 38.9%
  • Something else, explain in comment

    Votes: 5 0.4%
Back
Top