AMD 5900X + RTX 3080, limiting factor? Monitoring data included.

Hello. This topic is going to sound like "payed a lot of money to get performance, and not there yet...", but I will back it up with monitoring data.
So recently I did big hardware upgrade and I made a lot of experiments in order to find my the sweet spot between best visual quality and framerate. Here are my observations.

First the harware configuration
cpu: Ryzen 5900x watercooled and some OC (30% higher limits, some undervolt)
mem: 2x16GB ddr 3800 cl 16
gpu: RTX 3080 MSI Suprim X, aircooled, overclocked (110% power, +500mhz mem, +100mhz core)
monitor: Samsung chg9 ultrawide 3840x1080 144hz
Windows 10 20H2, nvidia driver v460.97, ACC v.1.6.6

For the purpose of my experiments I used following scenario:
Quick race at Monza 2020, 16:00, 28% cloud, BMW GT3 player car, AI GT3 random cars.
Start the race from last position (in most cases) or in the middle and maintain that position 0.5-0.9 seconds close to the AI. Do only first (opening) lap and when cross start/finish line quit.

As a variables I tested:
AI cars: 4; 19; 29 with my starting position last or 10th
Graphic presets: epic; high
Screen config option: triple (screen dived on 3, distance 495, side angle 13); single (vfov 37)
SMT (cpu hyperthreading): yes; no

Below in the screenshot you can find most important data from the results.


I have not analyzed all the data, but have some short conclusions.
- SMT (hyperthreading) doesn't have impact
- triple screen costs me noticeable fps loss (15+ fps in mid scenarios)
- in most scenarios gpu is fully utilized (that is good)
- in some scenarios neither gpu or cpu are fully utilized but getting low framerate - and that is my issue.

Additionally I tested performance impact with only changing individual setting vs the epic preset. Using triple screen option, 19 ai cars, same scenario as above. Results are in the below screenshot.


I keep all the log files and can share if someone wants to lose time in analyzing

So my custom preset is (20-epic-custom3) with triple screen setting and epic preset + following changes
View Distance: high, Shadows: high, Effects: high, Post-Processing: low, Volumetric Fog: off, Mirror Quality: mid, Mirror Framerate Limit: 45, Mirror View Distance: 100

Below is comparation of some data between epic (in red) and my custom preset (in green).
Some notes on the graphs.
"Max CPU/Thread Usage [%]" - is load of the single most loaded cpu core at the given moment.
"CPU usage [%]" - is the total cpu load (all cores)
"Average Effctive Clock [MHz]" - hwinfo calculates this and it's kind of another indicator for the load on the cpu. If it close to the cpu frequency means the load is high. Here is not shown the my boost clocks - but they are about 4750-4850mhz
"Core clock [MHz]" - is the gpu core


Now to the issue. As you can see on epic preset gpu is fully utilized most of the time. But with my custom settings it is not. And the other thing is that "Max CPU/Thread Usage" is also not high enough to say that I am cpu limited.

I mostly play online on servers with 15-30 cars, usually on monza and with sunny/low cloud conditions. My ideal target is to get 140fps flat without compromising too much quality. But I doubt that this is possible with hardware...
I also wanted to buy better monitor (samsung g9 ulrawide 5120x1440, but i gave up)

P.S. All log files here https://drive.google.com/drive/folders/1SLAr2ik4yyntFleu-1pSy2-uU2AS6X-q
The logs can be viewed with HwInfo Generic Log Viewer or with MSI Afterburner
Cut 10 seconds from the beginning and 10 seconds before frames drop to 30 (fps limit in the menu) near the end.

 

Georgi,
Thanks for the Info - is there a possibility you report CPU Load per CPU Core (or Thread if SMT=on)? Perhaps I did not read your graphics correctly, but I frequently see that ACC loads one Core (or Thread) to 100% for a longer period of time while other CPU cores are bored. As a result you see low general CPU load (averaged over all Cores) while in reality you are already reaching the limit of your CPU.

 

Hi @Dezember6th.
I have uploaded all log files here https://drive.google.com/drive/folders/1SLAr2ik4yyntFleu-1pSy2-uU2AS6X-q
The logs can be viewed with HwInfo Generic Log Viewer or with MSI Afterburner

In the log files there is info for each core, but I did not include it in the tables because there are too many cores.
In the graph there are only few brief moments where the single core load is above 90%, that is what I am trying to show - this graph Max CPU/Thread Usage [%] - is showing exactly this the usage on the most loaded core.
For example for moment it may be core 1 80%, next moment it would be core 6 90%, the graph will show 80%,90%...

 

This eerily reminds me of another thread: https://www.assettocorsa.net/forum/...ed-but-framerate-decrease.67725/#post-1177908

I do think you are CPU-limited, in a fashion. ACC's graphic engine uses just 4 cores - since you have more physical cores than that, SMT has no effect. With the epic preset, you're limited by the GPU, by lowering the options you shifted the balance away from the graphics card. Don't forget, you are bound by the latency of your whole system: SOMETHING is limiting the calculation, and the GPU can't start rendering until it knows what to render. Could be RAM, could be the CPU changing cores too much, who knows...

 

Yes sure it is possible, that is what I am trying to find what is the limiting factor.
In epic average most used core is 74.1% on custom preset is 76.9% .. not much difference.
RAM in my case is well optimized (56ns latency in aida), and even if it is ram I think that it would clasify into cpu load during mem transfers. If it is the cpu it would be 100% some core usage during the flare drops. For sure it is not core switching at least if you open the log you see that the load is not jumping so much between cores.
And that gpu is not utilized and cpu also also could be in programing too.

But even if I agree that I may be cpu limited lets say and want to spend money on faster cpu .. what should I get, I dont think that there are any consumer cpus much faster )

 

Dx12 is your answer.
More usage of cores then dx11 with this unreal engine.

 

Just as calculation example, lets say you have 100fps, so 10ms frametime. If you have 9ms which the CPU needs for a frame and the GPU does it's thing in parallel, then waits for the rest and in the end there is an additional 1ms for GPU only, which could not be run in parallel, you'll have 10ms/100fps.
But your frametimes are still are mostly CPU limited and because the percentage is an average over a longer time period, it will show 90%.


VR users might have some more details, I've seen some of the tools show cpu/gpu frametimes in other games.

 

Correct. I think at this point in time you have reached the limit.

 

fpsVR is such a tool, shows CPU & GPU framtimes, fps, temps., CPU usage live in Game + you can record and analyse with the history viewer.

https://store.steampowered.com/app/908520/fpsVR/








@georgi iovchev

Thanks for sharing your data and findings.

 

I also had to find out I´m still GPU limited with Epic settings when installing my 3080. I´m afraid the 3090 would be needed to get the FPS boost. But even with this monster 140 FPS won´t be hit all the time. A nice FPS boost can be achieved by lowering the foliage settings to mid and volumetric fog to low. To be honest I don´t watch at the grass that much and the fog setting is really taxing.

 

I doubt that 3090 will boost fps that much, having in mind its price, its just not worth.
Yes I know lowering foilage is giving performance. But on tracks where there are trees they start to look ugly.

 

I know maybe my FOV 17 saves me

 

Aah that variable I have not measured. Probably it will have impact also, cause depending on it more or less objects wil be visible.

I dont know if Kunos have plans, but I think that there is potential locked in the number of cpu threads that the sim uses. I dont have high hopes, this probably is related to engine or schedule changes.

 

The higher the FOV the lower the FPS. I guess there won´t be any more major optimizations. Going to DX12 seems to be too much work cause Kunos heavily adjusted the unreal engine to get it to work on a racing game.