Low fps, cpu & gpu usage

Those with more experience probably know why this is sometimes happening.

I bought new GPU, GTX1050Ti and as I was testing what it can do, I experienced slow fps (see attached image), CPU load and GPU load were not at maximum, temps were cool all around.

My guesses are that some thread wanted more computing power from single core or thermal throttling, but how thermal throttling would happen when temps are ok?
There might be also other possibilities that I'm not aware of yet.

Maybe there are some settings I should play with to avoid this issue, should I look nvidia control panel or in game?

 

It works fine 99% of the time, only some specific situations like that town in first post's screenshot get slower framerate without 100% utilization of CPU or GPU



I guess I could try to see those graphs when I get that odd slowdown.

Banana bench is not working, there is message what it gives to me.

Probably next week I will install new i7 cpu, that would be possibility to compare lot of things and might be that slowdown disappears with that too.
--- Post updated ---
This is how graphs look when I'm getting less than 60fps (limiter set to that) and lower than 100% CPU & GPU utilization.



Graphs started to lag really bad around half filled, really long delay with graph update near the end.

 

It happens mostly when driving at night with lights on and effect is strongest at East Coast town. CPU utilization with Sunburst is 60-70% with rally lights, gpu load is also 60-70%, none of CPU cores is above 80% and crazy part is that when entering town, fps drop, cpu and gpu load drops as well as temps.

Without Vsync there is not much change to framerate, but there is lot more jerkyness, it is not so smooth.

If I turn lights off, gpu load increases and cpu load decreases, when I turn lights on again gpu load decreases and cpu load increases. I think that somehow CPU is not being able to feed GPU with enough data or something weird. Odd thing is that none of the cores is hitting anywhere near 100%

Benchmark and tests show everything running as fast as they should, but maybe those tests don't use some bandwith what light calculation uses.

I try to get that i7 installed next week, we will see then if CPU is creating some bottleneck that does not show up in cpu usage.

 

It would be too easy if just lights would be too heavy on GPU, in picture GPU load 67%, CPU load 62%, wouldn't GPU load be 100% if it would be taxed to max? Also temperature is under 60 and when at 100% load temp goes to 60 or bit more, so that indicates also how GPU is not working very hard.



There is a bug in that disable shadows setting, if set anything else than none, lights are not following the car, they work only on small area where you changed setting to partial or all, so it is not currently possible to set shadows anything less than full for night driving.

Light quality has also big impact on fps, but again anything less than high and it becomes quite hard to see, but as GPU load is so far from 100%, that leads me to believe there being something more complicated happening.

 

I'm getting results that don't make much sense.

If shadows are at full, turning lights off will reduce GPU usage, I did full fan cycle first lights off, then lights on, you can see from this graph how lights on cycle is much longer, for some reason turning lights on lowers gpu load so much that gpu thinks there is not enough workload to keep going at full throttle and it decreases clock frequency, still there are no any kind of change in CPU load.

I put settings screenshots and .hml that MSI Afterburner makes in zip, it is easier to check data if opening that .hml in Afterburner, but part of that log is here:



Also two graphs you requested are attached. One thing to note though is that if shadows are disabled, turning lights on increases GPU usage, but if shadows are all enabled, turning lights on decreases GPU usage. Could there be another bug lurking in the shadows?
--- Post updated ---
I did test with a 3d mark fire strike, got gpu score bit over 7500 when 1050 Ti in guru3d test got 7400, then on Catzilla benchmark I got total score of 14349 and Catzilla certification.

There was one physics test in Catzilla benchmark where I noticed CPU load being below 90% and GPU load being quite low as test did test CPU, FPS was quite low, what it got me thinking is that as i3 has 2 cores and 4 threads, could it be that sometimes code can't utilize 4 threads and instead of speeding up threads slow down?

I think that next test I do with hyper threading off if I find that possible somehow, maybe it confuses more

New update:
I figured out that if I pause physics with 'J'-key, it should result increase in GPU load as CPU resourced are freed.
Despite CPU load not being even at 80% on any of the cores/threads with this testing attempt, when I pause physics I get increase of 5-10% in GPU load (lights on, shadows all enabled), still GPU load does not reach 100.
With lights off and shadows off GPU load gets to 99 and GPU load can get to 99 with shadows enabled, if lights are off, but it wanders a bit then, some locations in town I get around 80% GPU load daytime lights off and shadow all on, if I then pause simulation with J-key, I get 99% GPU load, which indicates CPU limit, despite CPU load on highest core/thread is 72-89% (CPU2 of 4 threads).

If I disabled shadows, I can max out everything else (except dynamic reflection) and get 100fps in previous situation and I can turn then lights on and get only 10fps hit on performance.

I'm quite certain that there are something with shadows that causes my CPU to not be able to feed GPU enough well, thus it will be interesting to see what CPU upgrade will do to this situation, I should get thermal paste at monday, so I can install CPU and I try to do fake scientific comparison of frame rates then.

 

Some reason bug with lights not moving with car, did not happen now on Jungle Rock Island, so I recorded performance.
Shadows and dynamic reflections off, everything else maxed out, poor i3 can still somewhat record a video. Car has all lights I could put to it, I that is 8 driving lights. I had also camper and oil barrel spawned on island, so when bug with lights is fixed, keeping shadows off will give quite good performance.


Oh, forum did not want to combine post with previous.

 

I noticed that if I select high light quality and partial shadows in game main menu, then load map and switch to full shadows (disable shadows = none), I get 100% GPU usage and lights and shadows work as they probably should.

Also I noticed that if I select disable shadows = all, and high light quality, I get working lights and great performance with usable lights, I can have tens of lights and several vehicles without dropping below 60fps on Utah Industrial starting location.

So there is something awfully goofy going around with menu that should select graphical options, I guess next version will have that bug squashed.

 

And this is what I get with i7 CPU, notice how one core/thread is showing lot higher usage than others, with i3 load was more even between cores/threads, but despite i3 did not have 100% CPU load on any of the cores, it is now quite clear that i3 could not feed GPU enough fast, now I get close to 100% GPU load.



Maybe it gives some ideas what to improve or maybe not needed to improve, idk, I can get constant 60fps quite easily with changing graphics settings, also with i7 I get now close to what GPU can do, with i3 fps was less and it seems shadows like from CPU power.

 

I have question for @Nadeox1 or someone on dev team who might know answer, why shadows are calculated on CPU with my computer, where as many others seem not to have shadows calculated by CPU or so many seem to claim?

Does shadow computing use CUDA cores and is there some issue for BeamNG not realizing it can use 1050 CUDA cores?

I spam video here too, that is two Off road roamers high beams on, both have extra spot lights added, you can see how I'm limited by single core performance, if I turn shadows off I don't get that kind of fps issue. Around 20fps with two cars is quite little, GPU load is not very high, with different settings I can get GPU load higher as well as FPS higher, while CPU load remaining similar, so it does indicate that CPU is the limiting factor, with those settings shadows off my GPU would probably not be able to reach 60fps, but I guess even Integrated Intel HD could do the same FPS on that situation.


In practice I can't use shadows at all on many maps, because I drop below 60fps at places and I'm not sure at all if upgrading GPU would help as it is clearly i7-6700 CPU that is hitting maximum here, unless it is so that for some reason GTX 1050 Ti can't calculated shadows used in BeamNG and higher level GPU's can?

In this clip I turn shadows on and off so it is easy to see how big effect they have and how shadows off CPU can almost keep up with GPU, 27fps shadows on, 90fps shadows off.


I get 161 in Cinebench single core and @VeyronEB mentioned not having this issue even he gets only around 100 in Cinebench, so I don't know if there is some bug that causes shadows to run on CPU with my computer or is it really so that for others there is less CPU load by graphics thread.

I know shadows are heavy and that 2x 6 lights make a lot of shadows, but if others don't get same kind of CPU load and slowdown, there must be something that needs closer investigation.

Some graphs that may or may not be useful, but as sometimes those are requested also. FPS is bit higher as not recording video. Is that blue in shadows on graph gpuUI or preRender? How do I make that to be less if that is reason of this slow performance?

This is shadows on:


This is shadows off:

 

I got same kind of experience with i3 + GTX 1050, there is not really much of difference between i3 and i7 in single core performance, but I get higher framerates overall as more cores. With integrated graphics GPU is limiting factor, so can't do comparison tests with that and GTS 8800 runs UI really really fast compared to GTX 1050 or especially iGPU, but 8800GTS is slower than iGPU in graphics rendering in BeamNG.

I have done a lot of testing, setup changing, BIOS fiddling etc. I like to tweak last bit of performance out from the system (after all, I have been overclocking and making water cooling setups before they sold water cooling to public), but I can't find anything that would be the reason, that is not to say there would not be anything that I have yet to find, but so far not much luck and still benchmarks show perfectly good scores, only issue is within BeamNG shadows or something.

I have looked anything from PCI-E lanes, to thermal throttling, but none of such, all I know is that CPU load on thread which runs graphics likes to spike up and down really fast and goes to 100% with shadows enabled quite fast, then GPU load goes down as not enough data goes to GPU from that graphics thread. I have played with every setting in NVIDIA control panel too, threaded optimization is not doing anything was only odd thing I found out.

Not sure if I have listed full specs, but:
Win 8.1 Pro
i7-6700 @ 3.7Ghz
16GB 2666Mhz DDR4 Kingston Hyper-X using XMP profile 1
Gigabyte Z170X-UD3 with latest F21 Bios
Gigabyte GTX 1050 Ti OC 4GB
Samsung SM951 M2 SSD
Seasonic S-12 460W Fanless Psu, which I added a fan to so it stays nice and cool as everything else.
CPU cooler is Thermaltake Macho Rev.B, CPU temps stay below 45-50 (depending room temp) even at highest stress.

My Cinebench single core result is 161 and like every other benchmark I get exactly result which I should get.

I have done clean reinstall of graphics drivers, steam, BeamNG, actually BeamNG and steam are in different folder these now, but no change.

As you can see I get perfectly normal results, problem is only happening with BeamNG and shadows are main reason for that high CPU load.

I did download LFS and it seems I get over 900fps without maxing out CPU or GPU (Actually if I go to options menu, I get over 1000fps and CPU single core is then maxed out with 90% GPU load), I even had power profile balanced and GPU in ECO mode which gives bit less performance, but that is very light game, is there other Torque game or demo that I could test to see if same happens with other Torque titles?


In Catzilla benchmark I get perfectly correct scores, in 720p my result is 10th and with 576p 2nd best GTX 1050, there is CPU in test too, but that uses all cores, no issues with that result either.

Here are other benchmarks, for me it looks just fine and if there would be something wrong with my configuration, I would think it should show up in these tests?





I think what needs to be done is to figure out if other 1050 users are getting their CPU killed by shadows, there might be some bug, maybe bug is in Nvidia driver, card or something else, or then it is completely weird thing that affects only my machine.


Yet another thing is that I have to take PC to pieces and have some old HDD running Win 10 (that should work few days without activation) next for testing BeamNG in that if it does the same or not.

But it really feels like being a bug of some sort in some place, finding cause is just not easy it seems, but thanks from confirmation that graphs are not what they should be, now we need to find out why.

 

@Nadeox1, it seems that there is not much difference in Windows 10, I might not had quite same settings, so FPS is bit different, but same high CPU load limiting performance still is there, also in Win 10 I don't get 3.7Ghz.

Sorry that I don't edit shots, I don't have tools installed and I don't plan to spend any longer than necessary in Win10 as I don't like it at all. Anyways, I guess we can start to think it is something else than Windows related?



--- Post updated ---
I think I got close to same settings with Win8 for comparison, but in Win8 I get that 3.7Ghz boost clock, however with shadows on it really seem not to use that extra speed because CPU load is lower, however CPU load spikes really quickly to 100% when shadows are on, but it is just very quick spike that does not register in task manager, but with 50ms cpu load reading interval it shows up.




I'm not hitting power limits, or thermal limits either, so annoying and frustrating.
--- Post updated ---
Two more interesting shots, I messed up shadow settings in world editor, did not gained much fps in there, but what I found out is that with lights off I can get CPU core that runs graphics thread to max out in task manager.

Whatever those lights are doing is causing odd spiking CPU load which is not picked up by task manager. 2 x 6 lights, I wonder if they play ping pong between CPU and GPU there, cast light, trace light, paint shadow, something on that area? However that really does not explain why you others with same settings and vehicles in same place don't get that, I really want to think that it would be something that happens with only this graphics card model, but I would need to test other GPU in order to eliminate that possibility and it is quite impossible at the moment.


--- Post updated ---
More interesting things, it seems that MSI Afterburner causes CPU load to be less than 100% at task manager in that one CPU core, because turning it off, CPU load reaches 100%, however it has Zero effect on FPS, why it affects such way to task manager and why it has no effect on FPS is bit illogical.

Afterburner running:

Afterburner not running:


Another thing is that when I adjust graphics details so that I'm GPU limited, CPU load still is very close to 100% on that one core and if I then turn shadows on, GPU load drops, CPU loads goes to 100% and my FPS gets killed because CPU is not enough fast. In pic same situation and settings as those above, but shadows off, so system is GPU limited then.


It is actually quite hard to get settings adjusted so that I would be GPU limited in that setup.

With Dynamic reflections at max, I get lowest FPS, but it is still CPU that is limiting:


Graph from above situation has more pink in it, which might be also because Afterburner is not running:


All the testing and adjusting still does not give me solution, it just looks like to be that I'm running out of CPU power. For me it is hard to believe that this would not happen with other machines, after all according to tests my CPU power is in correct range, single core performance is quite high level, no adjustments in BIOS or in software is changing this.

 

@Nadeox1, not only 4 lights, but I add 2 more, those extra spots that I have mentioned (I think I wrote to some video description too), so it is 2 x 6 lights.
I see it kills your FPS too, so it starts to look more normal and anyone claiming otherwise has just misunderstood this test.

Yeah, I hide UI when taking graphs as it seriously affects accuracy of data collection.

Many addon maps have shadows that kill FPS also, related to number of objects and I guess lights on map, I have tested with roamer as that makes same kind of FPS drop and CPU load as those addon maps.

I have read many commenting that this does not happen to them, but perhaps they just haven't tested quite same way.

So, thanks from those graphs, now I know that to get more performance it is indeed CPU single core performance where to focus.

Update: Ah, you added edit after I quoted.

One crazy thing about this GFX limitation is that it is GFX limitation on CPU as old shadows are calculated by CPU. Indeed I think that is single greatest limit in terms of performance that BeamNG currently has and not so easy to go around, splitting each light/shadow to own thread would be quite a nightmare, changing engine so it can do newer shadows is not simple thing either.

For maps limiting drawing distance improves things, but terrain can't be limited very much.

I'm sure at some day you will find nice way to improve performance though as you are quite clever people

 

@Nadeox1 You know, there probably is something wrong in game or in Torque 3D, because I can run 12 lights just fine on maps which don't have much of mesh objects like buildings or roads, but more there is mesh objects like in ECA or some mod maps, higher single core load I get and things slows down.

Now these are just my thoughts, which I throw up in the air, they might be on something or completely bollocks, but maybe some food for thought, maybe useful, maybe not.


So buildings and mesh road objects seem to introduce some extra CPU load, at ECA using many lights this issue can be shown out, on some mod maps lights are not needed, just number of objects in view cause this high CPU usage.

Improving that would make it possible to have more complex maps and I'm not entirely sure that it is not a bug that is bit hard to flush out, but maybe it is in nature of such objects, some calc that needs to be done, however after ton of testing it is very apparent there is something that might be worth to look.

Limiting visual distance helps as there is less objects so less CPU load comes to play, but is it really just masking the issue or is it really just something that needs to be calculated regarding the objects? Then why would adding more lights make issue happen more easily, trees get shadows too, why they would not cause same issue? Why buildings at Hirochi would not cause same issue?

Maybe new version already has changed this behavior, but to my understanding there is something happening with mesh objects that might not be quite way it is wanted.



These screens just show how huge difference there is in CPU load and behavior.

This is graphics settings high, no dynamic reflections or SSAO though, hardly any CPU load even there is really lot of stuff:




At Hirochi my GPU is not fast enough to run 60fps, but same way no huge CPU load, GPU usage at 99-100%, JRI is the same:


At ECA, things go differently, only 80% GPU usage in this spot and single core load at 100%, town is even worse:

 

... that's what Nadeox1 has been telling you. The thing that is "wrong" is that BeamNG is using old/bad shadow techniques. There are actually plenty of old techniques which either never worked, didn't scale well, etc. I read big chunks of this thread and skimmed the last half. As you've already guessed, fixing old ideas is probably less advisable than just gutting chunks of the graphics code and replacing them with better or more modern things.

I mean absolutely no offense in saying this, but it's no mystery that the graphics engine is not powerful and is a bottleneck. Your time may be better spent on other topics. ;-)