AMD Ryzen CPUs

ahh, i thought u meant that the 1060 was a bottleneck for the 1600 in ac

its not weird that you get more fps with 8 ai cars than you get with 24.

 

Oh well, 1060 <> 1600 ...
This might be a confusing configuration

I guess for online racing it will be fine, but for AI with a big number of cars it seems that the CPU is the more relevant favtor and (for now) it seems higher clocks are the dominant factor as multicore performance isn't that great yet.
You can always see how 1-2 cores are really on high load and the rest is at medium to low load.

 

Dang. I get the same thing with my computer, just a lot less FPS. Seems that the AI racing always does that no matter the system, and worse if you enable reflections to a lot of cube-map faces or raise the resolution. I have yet to hear of a system where the cpu isn't the bottleneck for AI racing.

 

Your video of 48 cars shows both the CPU and GPU being used less than 50%. Looks like only one CPU core is at about 84% while all 11 others are around 25%. The overall CPU utilization is below 50%.

How should we interpret these results? If both CPU and GPU are loaded less than 50%, how could there be a bottleneck? Could be it the wonky hyper-threading CPU reporting? For instance I do know that when I run prime 95 using two threads, my i3 being a 2c/4t CPU, shows two cores loaded 100% and 2 cores loaded 0%, for a "total" of 50% CPU utilization which is totally inaccurate. While only 50% of the virtual cores are being loaded, reality is that the silicon is loaded at ~100% since both physical cores are loaded 100%. This is my understanding, please correct me if I'm wrong here.

Are we looking at the same wonky CPU usage scenario here and in reality the CPU is much more loaded?

On the 48-car video. One virtual core is at 80% and all other virtual cores are at least at 20%. Would this mean that one physical core is at 100%, thus creating a bottleneck?

I'm writing this because I am interested in getting a Ryzen CPU but would like to understand results better.

 

You should know how the physics work in AC and what the CPU Usage in the 'Render Stats' app represents.

You really need to watch this if you don't know it yet, from 40:00 on:


So I guess that for the physics calculations AC only uses one or two threads, hence the very high cpu load on 1-2 cores/threads.
My take on that is simply that a higher clocking CPU is at the advantage over 6/8-core cpu's since the physics calculations aren't multitasked (or not multitasked to all available cores).

If you look at the 'Render Stats' cpu load you clearly see that it is almost all the time at ~90%*, so the physics calculations are bottlenecked by what 1 or 2 cores (on which the physics are running on) can handle.

So in the end 6-/8-core cpu's with lower clocks are at the limit of what 1/2 cores can do, but much CPU performance is wasted since the other cores aren't utilized, so 4-core with high clocks seem to perform better.



*meaning that it takes 90% of the frame time to calculate all the physics, which means:
more cars -> more physic load -> time to calculate physic per rendered frame increases -> gpu/framerate gets bottlenecked by the cpu

 

Assetto Corsa uses 4 cores. But, another physical or logical core is very useful to other things on your system and keep the minimum fps as high as possible.

 

Of course it uses 4 cores or more, you can see the load spread on multiple cores, but it's always just 1 or 2 that are (almost) on full load and the rest just does 'some sidework'.
If the physics could be spread on multiple cores the cpu wouldn't bottleneck or it would be at a much higher framerate.

 

Something strange that you guys have, if I put enough cars all my 4 cores are loaded 100%, my GPU stays at 35-50%...

 

Interesting thing I just noticed, I ran more tests on Road America, regardless of what I tried (forcing number of threads in the assettocorsa.ini file, thanks for the suggestion mms) there wasn't a noticeable difference.

But in one run the AI crashed me on the 2nd straight, while I was spinning I got around my usual frame rate while I was facing the other cars, though while I was facing a direction that didn't have opponents in my FOV the frame rate jumped up by 30-50%.
So the physics load isn't the only thing creating the bottleneck and the only explanation I can think of is that the animation of the other cars needs those 30-50% of cpu load while the GPU is on low load.

 

OK, since sleep is overrated and I had an idea...

I wanted to see how the frame rate and cpu usage is affected by 'doing it live' vs. replay, so this is a recording of the loaded replay file.



As you can see the physics load is 0%, but somehow the framerate only increased by maybe 10%, so there must be another reason for the cpu bottleneck. Maybe in the end it really is some rendering thing about the opponents cars.

But you can see how one thread is near or at 100% the whole time.

So now I'm even a bit more baffled than before.

 

Well, I wasn't talking about how AC uses and loads threads. My whole point is how pretty much every software (AC included) I've seen that displays CPU power has a wonky way of reporting CPU usage with hyper-threading which can be very misleading.

For example, I have two computers. One has an i5-4670K with 4c/4t and has an i3-4160 with 2c/4t. When I run Prime95 loading 2 cores, both CPUs are reported being used 50%. In the case of the i5 this is accurate, but in the case of the i3 this is not real. The i3 is really loaded pretty much 100%.

I run another experiment, this time using 7-zip to compress a file using 2 threads. I clock the i5 and to match the i3 @ 3.6 GHz. Running the 7-zip test with 2 threads, both CPUs perform a task virtually equally, on both the CPU usage is reported to be loaded 50%. This doesn't make sense as the i5 and i3 don't have the same resources. Now I do another test, this time compressing 2 files simultaneously with 2 threads each. The i5 finishes each file almost at the same speed as when it was compressing only one file, CPU usage is reported at 100%. On the i3, running the 2 files concurrently takes almost as double as doing a single file, CPU is reported used 100%. In short, the 50% reporting of the i3 on the first test is not real, the CPU was being loaded ~100%. It is just that the CPU monitoring software just averages the loading across cores without realizing whether they are physical or virtual cores.

My point is that when using an hyper-threaded CPU, the "total" CPU loading result is misleading. You need to look at individual core usage. In the case of the 48-car video, a single core running at 80% plus all others at 20% means there was one physical core which was loaded 100%.

I reach this conclusion based on the experiments I described.

 

Really interesting results.
It's really strange seeing the cpu AND the gpu not used at 100% either of them.
Thread from another user with his strong i7 6700 and 1070 and all components at 50% with the formula hybrid mod comes to my mind.
Something is the bottleneck but we cannot see it either in cpu in any core or the gpu.
I believe that Stefano can only solve this mystery for us @Lord Kunos

Στάλθηκε από το m2 note μου χρησιμοποιώντας Tapatalk

 

The total cpu usaveis below 100%, but that isnot the relevant value here as it looks.
Render Stats is always around 90% despite that there are still cpu ressources available but not used.
Though in the replay video you can see that one single thread is at 100% load almost all the time.

 

Exactly what I mentioned some pages back.
There is the possibility that multicore beasts like ryzen 1700 or 1800 get wasted because a single core gets limited.
And this can happen in an i7 too.


Στάλθηκε από το m2 note μου χρησιμοποιώντας Tapatalk

 

Yes, but usually the i7's are higher clocked than Ryzen and therefore probably perform better.

 

As a flightsim user, the clock speed of the CPU is particularly important. That being the case, AMD have some work to do in the overclocking department, because the i7 7700k is hitting 5Ghz with decent cooling.

 

Even with LC the Ryzen wall seems to be around 4GHz, some struggle to get to 3,9GHz (with acceptable voltages), some run 4,1GHz.

In another forum someone linked a tool to change clock speeds live, so I'll compare my 3,6GHz against lower clock speeds and see if the difference scale with clock speed.

 

Oh well, the mystery continues to grow.

2,6GHz:


3,6GHz:


Higher clock is slightly better, but it looks like less than 10% that 3,6GHz is faster over 2,6GHz, but 2,6 -> 3,6 = +38% clock speed increase.

 

Have you tried to change
[PHYSICS_THREADING]
THREADS=-1 ; -1 = automatic , 0 = disabled , 1...N number of threads dedicated to physics
C:\Program Files (x86)\Steam\steamapps\common\assettocorsa\system\cfg
assetto_corsa ini

if AI use the same physics maybe dedicate more threads might be better?

 

mms suggested it already via PM, but it didn't make a noticeable difference to me.