Why does adding AI kill the GPU usage?

I'm not talking about overall performance (framerate), I'm talking about my GPU's actual load (see graph).

My stock Nvidia GTX 460 starts out at 80% load with one AI and drops to 40% with 15 AI. The framerate follows the same trend (roughly halving) and I think that's solely because of the GPU drop, not because of the added geometry/textures.

And my CPU (stock Intel 2600K) doesn't budge at all during the entire process. So are the AI routines running on the GPU instead of the CPU?

What gives?

 

Good spot. I will try to replicate this.

 

See attached.

There was a dip in GPU usage as individual AI were added to a private session, but this stabilised back to it's original usage a few seconds after each AI joined. Likewise, after adding 29 AI in total, GPU usage went back to where it was when the private session started. So, I can't replicate what you are experiencing for some reason.

Oddly, when pulling back to the main menu, there was a spike and stabilising in GPU usage up to ~78%? I am not sure why this is.

My specs can be found on my profile.

 

Your session started with your GPU only experiencing 50% load.

That's a problem that ISI really needs to focus on.

 

I just noticed this.

I took the last place on a grid with 15 AI and GPU usage was at 87% with 26 FPS
I remained stationary and let the AI come back around.

While the cars where in rear view and passing
GPU usage dropped to 40%, FPS dropped to 10
You could see the stutter which wasn't there at the beginning of the race.

On the next lap I tried with the rear view mirror off.
FPS were better until the cars passed and then dropped to 8FPS with GPU usage at 28%
You would have expected an improvement not a reduction in performance.


This only seems to happen for the Meganes the Open wheelers show no GPU drop off.

 

That's not my experience. Here's the historic cars at Spa...

 

I get some drop off at Spa but not as much as the Meganes at other tracks.

My max GPU load at Spa is only 70% max.

 

You should also have a look at the CPU usage. It will probably be nearly an inverse of the framerate and GPU usage. GPU usage is tracking the framerate very closely and this is not a surprise. It means your system is CPU-bound which is fairly typical for a simulation. I have seen this with my system when I race with lots of AI cars.

 

Vince, he did look at CPU usage and stated it doesn't go up when adding AI. I have seen the same thing. My CPU usage is always 30% regardless of how many AIs I add. But the frame rate drops like a rock even though my GPU load never goes over 70%. Something is really not right in this build. I hope they fix it soon.

 

My CPU is barely even working with 15 AI cars.

 

And before anyone says "Turn off Hyper-Threading": It doesn't help.

 

Barely even working, with one core at 80%? I'd say that's quite a load, actually...

 

In a game that should already be optimized for two cores, yes, that is barely even working.

 

Yes, it is a big load and, as I said, it is to be expected with a simulation. Remember that ISI has said that the game will utilize two cores. It will not distribute itself across four or more processors so you can't just look at the aggregate CPU usage. You have to look at the usage of individual processors.

Also - the thread that is heavily loading the CPU might not always run on the same processor unless they are setting its affinity to a processor. If they do not do that it is possible that it can 'bounce' around and use other processors.

 

I wasn't aware that the software was dualcore. Then I agree, it's wasting alot of cycles.

 

My GPU usage is not good either. 45 to 75 depending on settings The more details I set, I get slightly more load but the target performance is way too low for any setting. If I'm not getting good performance, the GPU should be at least trying to give it to me. I notice this in a couple other racing titles that are not optimized well. The ones that are optimized well, run good and put my GPU to work. The results are good though.

 

I would not say this game is wasting cycles. It is a simulation with enhanced physics modelling so it is just using a lot of the CPU. If you want to optimize things for yourself you can turn off things like special effects, tire wear, fuel usage, and damage but that is up to you.

I did an experiment a while ago that was interesting. I used Fraps to measure the average frame rate over a replay of one lap with 50 cars on the grid and I got the same result at 1024x768 as I did at 1920x1200 which demonstrated to me that the game is CPU-bound with that many AI cars. This was done with RF1.

 

Vince, what I meant by wasting cycles is that at MOST I get 80% usage of a single core, and only for short bursts. The software is written to utilize two cores, and as the framerate is rather poor I would expect to see at least one core maxed and thus indicating that I am CPU limited. As it is, I average around 50% usage on one core and the others are idling or doing up toward 25% in bursts. Same goes for my GPUs. They are fairly evenly loaded, but I never see anywhere near full use, usually in the 40-60% region.

This indicate to me that the bottleneck is elsewhere. My solid state drive is not working very hard once all the textures are loaded. My memory is paging nicely and below capacity, so that's not the bottleneck. I don't see where the bottleneck is, besides software.

Other games that run smooth I see a nice even load on either CPU, memory or GPU, which indicate I am hardware limited to either of those units, and can load the other units more if I need to. If I see I'm CPU bound I can add eyecandy without penalty. If I'm GPU bound I can remove eyecandy and get better framerates, or add CPU stuff like AI without penalty.

With rF2 I can't do any of these things. Adding AI reduce GPU load. Adding eyecandy reduce CPU load without increasing framerate or GPU load. Counterintuitive.

 

I just wish that one of the ISI coding experts (eg, not Tim or Scott) would comment on this issue. :/

 

This is understandable to me. It makes the system more CPU-bound which lowers the framerate because of increased computations and that reduces the load on the GPU because it renders fewer frames.

Yes, this is counterintuitive and makes me wonder how the GPU load is measured. I am surprised it would reduce CPU load unless you are becoming GPU-limited which will result in a drop in framerate but if GPU load is not changing that is not the case. It always does for me with RF1 so I tend to drop shadows and special effects down a notch.

In the end, we should remember that this is a beta release so it is unlikely to have received much if any optimization which usually does not happen until the feature set is frozen which has not happened yet either.

BTW - Scott and Tim are not coding experts. Terrence, Joe, and Jeremy are but they rarely post here.