Interesting "Low" Performance (3770K/1080Ti)

So just got a 1080Ti (Asus ROG) its paired with a 3770K at stock clocks. (I'm waiting for my mac pro 5,1 dual socket.)

16GB ram

Videos attached and settings attached.

Grid map 210-240 FPS



WCUSA (57-59 FPS)



ECUSA

 

There is bit of settings missing from your settings screenshot.

 

In WCUSA with the AMD ASUS 380X 4GB with the same settings (mid 1080p card from 2015)

I used to get 30-40 FPS on WCUSA. So it can't possibly be a only 10-20 FPS bump? I didn't see it get over 63 FPS yet.

 

I just tested gtx1080, at first glance not much improvement over 1050Ti, also reason is what I have been telling long time, single core computing power of i7 6700 is not enough, not even with the 1050Ti with single car.

However I need to analyze data properly and do some charts, which will take maybe few hours, but I would say anything above 1060 6GB is not going to get much higher FPS, but might be able to keep 60fps more steadily with certain settings.

Anything with less than good 164 at cinebench single core test is probably having hard time feeding these faster GPUs.

Pausing physics is giving me huge fps boost at eca, from around 70 to around 120fps, but that still is not 100% utilization of 1080 and for certain with 1080Ti you would need overclocked Coffee lake i7 to keep it fed, if even that is enough.

I guess this tells already enough, my standard test loop at ECA with my standard settings, that one thread of CPU is working so very hard, but GPU is resting. This CPU thread has been working harder in 0.13 than earlier versions. At beginning and at the end physics are paused.

Near the end drop in GPU usage happens at ECA town.


Part of the graphics are processed by CPU and same thread runs graphics too, UI was hidden during the test.

Test methods and tools in performance testing link at the signature.

 

As @fufsgfen said, single core cpu is very important. At a certain point, gpu upgrades produce diminishing returns. I’m not gonna lie by saying switching from an ATI r5 to a gtx 1060 didn’t quadruple my FPS, because it did. But, once you are settled in the GPU department, single core perfomance prevails as becoming more important because CPU’s have the share a core+thread with graphics, and if the thread is not very good, bottleneck occurs, limiting performance.

The Ivy Bridge processors were good back in 2012, but in 2018 things are better. If you want to upgrade to a new CPU to not bottleneck the system, I would reccomend moving to AMD’s excellent Ryzen architecture. A brand new 1800x will set you back 240€ now that the 2000 series ryzens are out (don’t be tempted by them, they are definitely a “tock” product). That will eliminate bottleneck entirely, and, if you’re willing to splurge on a 1080, you shouldmsave some money up and buy the new ryzen.

TL;DR: you’re bottlenecking.

 

There is no CPU in existence that would not be a bottleneck for 1080Ti in this game.

My CPU is something like 164-167 in cinebench single core test. Desktop Ryzen is no much faster, fastest is Coffee lake i7 special edition that runs higher clocks than normal i7 and even that is only about 35% more than my i7 6700, that 35% I doubt is not enough for 1080Ti in 0.13.3 BeamNG, but what I'm sure of is that devs will try to find a way to get lower single thread load, as currently it is way too high, so at Beta if not earlier they probably will work on that.

That single core limit is problem even when running several cars, if they could somehow move 50% of that core load to other cores, we could run a lot more cars than we can now.

I wrote more of my tests here, also some thoughts on how to manage that single core load with graphics settings so that you can run BeamNG well and pretty:
https://www.beamng.com/threads/cpu-single-core-load-gets-bit-high-at-eca.56637/#post-901996

 

---

Whoops I misread. You’re totally right, and downgrading to 1070 and ponying up to a better CPU might be a good idea.

 

That's a single vehicle, although I do agree on CPU being a bottleneck. Wouldn't this be partial and combination to do with the GFX engine also? Especially since there is such a huge drop in performance from one map to another with the same/similar cpu load across grid map vs WCUSA?




--- Post updated ---
Just for testing purposes

Going from stock clocks to 4.3 GHz improved FPS from 54-55 (Stock) to 57-60 FPS (4.3 GHz)

Sadly I really can't test any further since this 3770K doesn't like 4.5 GHz on air.

 

With 0.13, not sure if it was in 0.12 pausing physics improves lot of my fps, with single vehicle, but in 0.8 there was no change.

My suspicion is that something physics related has been added that runs now on same thread as graphical stuff, which is GFX engine side, that does cause considerable load on that single core.

Now adjusting graphics settings helps that single core, which helps again more physics to run on that single core.

I can't tell if it is sound or physics actually, pausing physics pauses sound too? Could it be that sound engine is what makes that extra load now?

How do we run BeamNG without running sounds, that would be interesting test.

However it is not only one thing that loads that core, which makes it bit harder to pinpoint what there is going on, but my guess is some collision related, sounds (engine, car related at least), some GFX stuff, for me SSAO has effect on that CPU core load as despite having under 60% GPU load, disabling SSAO increases FPS, so it has to be coming from lower CPU load.

Now lower single core load here could mean that you are maxing out single core, with lower load you would still be maxing out single core, but it manages to do bit more loops of code, so fps increases.

HwInfo is absolutely fantastic program for diagnosing these kind of things as it logs so much data, which you can then open in Excel and make graphs, that will reveal quite clearly what affects which and if it is CPU or GPU responsible for less than excepted performance.
Just don't keep it running always as it sometimes tends to shut down computer without warning, like pulling the plug, sounds wonderful, does it? Like a russian car, does the job, don't except everything to work though

Anyway, I think it is more complex that just GFX engine vs process, there probably is sounds, GFX and at least part of collisions on that core, sounds run each frame, so does turbo model etc. all that needs to be synced I guess and running them async might be quite difficult, but I know only very little of multi threading stuff, I did only few in the house utilities with C# where threading was useful. all I remember was that some things had to be kept on same thread so that they did run in sync or hell broke loose, some really difficult problems to diagnose happened when attempted to thread too much, still having nightmares.

Collisions, those are things that multiply load I guess, at WCUSA you have loads more collisions at gridmap you might get higher FPS, but that also means your CPU load is same as WCUSA. Each frame adds some percentage, so more frames or more collision models and CPU load goes up.

Your total CPU load might be something miniscule, but one core might be working really hard.

You can see from my thread I linked earlier, that there is not huge change in single core load, but it is obvious CPU is bottleneck. Measured CPU usage percentage is really terrible value, most programs log it so slowly that fast spiking nature of the load is not showing up at all, so you get some kind of average, that can be 20% lower than highest spikes.

Kinda like during one second you might get 10 100% spikes, but task manager shows 80% as that is average of that 1 second.

So you kinda can't trust taskmanager or Afterburner percentages, they are something roughly there, you need much faster sampling rate, 50ms starts to reveal something, which then leads to load of other problems and challenges.

My own tool is enough fast, but there is no logging with it, HWinfo is nice as it logs so much of different sensors, even it does not show that quickly spiking CPU load either.

In game performance graph thing is nice, if you can read it, I really don't get too much out of it, I don't know what is lot and what is little in there, also I don't know how well it shows single core limiting performance.

Sorry about book again, it just is not very simple topic.

Update: About MSI Afterburner, it somehow can affect GPU load, it might lie or something, but I don't generally consider it reliable. Long time ago I did some testing and it could prevent GPU reaching 100% load, maybe it caused high CPU usage or something.
Also more FPS you get, more CPU power any FPS counter will rob, fraps I think uses less than others, but even ingame fps counter has CPU load increasing effect, I think it started to show up above 100fps though.

Then that ingame performance graphing thing, it works only when you hide the UI right after starting it, then wait, unhide UI, press immediately print screen button and paste image to your image viewer/editor.
It twists results a LOT if graph is kept visible and after unhiding UI it auto adjusts graphs in next update, so you really need to capture what is there quick to see real results in best resolution.

Testing is kinda hard, there are so many variables that can mess up with results, making you think something, when in reality is something else.
It took quite bit of testing to find methods that I use, which I like to stick to keep results comparable and such that I know they are reliable with least possible variables messing the data.
--- Post updated ---
On maps, also number of objects, especially meshroad objects is creating some CPU load on that single thread, it is like if each would have some kind of overhead and at some point it becomes really big.

Now shadows then seems to be affected by that too, because when lot of meshroad objects are used, disabling shadows has a huge effect on FPS and that CPU single core load.

It is not a polycount, also at WCUSA they made large areas as single object, which probably has helped with this issue, but then they will have less possibilities using LODs to control GPU load.

JRI has always been map that runs at very low CPU single core load, despite it has lot of trees, but there probably is not so much of mesh road or similar to those objects.

 

Interestingly, at the morning I found out that hiding UI can have about same FPS boost as you had with overclocking.

Using a bus in chase view at WCUSA, details set to high with all checkboxes checked, no dynamic reflections and shadows partial.

UI overhaul is something to look forward.

 

Hiding UI boosted it from 60 to 70 FPS at 4.3 GHz.

 

So that is more than what you can get with overclocking, almost 15% fps boost ?

I'm not sure how in games UI usually works, but I would think that is a tad lot CPU load for an UI.

Also weight distribution UI app alone was some percents.

 

WHY ARE YOU GETTING A MAC!? ITS probably gonna be worse than ur current setup and probby would not run beamng. Also WHY ON EARTH are you using an i3 with a 1080TI!?

 

Who is using i3 with 1080Ti?

Then again, performance would be exactly the same with i3 vs i7 in BeamNG with single car as single core performance is what is limiting and at least Skylake i3 and i7 have same single core performance, upgrading from i3 to i7 does nothing to this low performance issue, despite impressions of i7 being so much faster, it is not, only it can run more cars, but that is about it.

Even upgrading GPU from 1050Ti is not improving lowest fps much at all on these problematic maps, because it is still single core performance of the CPU that is limiting.

Fast Coffee lake i3 probably will murder many i7 CPUs as single core performance is quite bit higher.

 

wait i think i got the processor wrong but anyway why use a mac for gaming......

 

I don't know much about modern macs, but I think they are same hardware as a PC and you can do dual boot and have benefits of Mac OS and Windows gaming both in on package.

 

I play BeamNG on a Mac Pro 4,1/5,1 with dual hex-core X5670s and an RX 580 8GB. I would dare say that the performance you'll see on your Mac Pro could be even lesser: these Xeons are not the CPUs intended for this game. If I stick the RX 580 in a modern i7 machine, it would do way better.

Still, the Mac Pro does fairly well: on every map, medium settings @ 1440p will typically grant me an even 60FPS. However, when I had my GTX 1060, the game stuttered regardless of the FPS count.
--- Post updated ---

Why not? I use a 2009 Mac Pro for BeamNG. I upgraded the Mac with dual 6-core Xeons, a PCIe-based SSD, an RX 580 8GB, and 32GB of RAM. This computer rocks.
--- Post updated ---

The 5,1 Mac Pro was released in 2010--he's not getting a new Mac. However, it was such a powerful and expandable computer, it can be used today without any issues.

In case you are unaware, many creative professionals still use the previous-generation Mac Pro because it's a fantastic and very expandable workstation. People run Titans in these machines. The highest CPU config for a 5,1 Mac Pro would include two hex-core X5690 Xeons @ 3.46GHz; the highest RAM config would include 128GB of DDR3 1333MHz ECC memory. Pair that with TBs of storage (within the 4 default sleds + two extra optical bays), PCIe-SSDs, and one or two heavy-duty GPUs, and you have yourself possibly the best 'old' computer there is.

 

To update this, I just go the 4,1 --> 5,1 cMP up and running, it does pretty well for having slow and old processors for 2009 Its 2 E5520's 4 core each. @ 2.26 GHz. (47-48 FPS) @ the same settings as my 3770K system.

I need to purchase another set of, x5680s. (The first set I bought was delidded wrong and was DOA.)
--- Post updated ---
For any one who is curious..

This is Gridmap with one D-Series.

Processors are E5520's @ base of 2.26GHz and boost of 2.5 GHz.

 

So as some as you know I have a system (Classic Mac Pro)

Swapped in my single CPU tray

W3690 (6C/12T) overclocked to 4.1 GHz (similar single core performance as a 2600K I7)

It cracks 3-4 FPS fps better in the same situation, than the 3770K @ 4.3 GHz. (Same 1080TI)

 

It is bit surprising that you can get faster fps than 3770k with similar single core performance to 2600k.

Single core of 3770k should be some 5-6% faster than 2500k at stock clocks, but then again your CPU has probably more cache, which might be helping or then memory performance is helping or both.

Userbenchmark lists 2500k as 98 and 3770k as 105 single core mixed points, best CPU at is 140 points in their listing, less than 10 points per year improvement.

On one thread tdev mentioned testing Vulkan, I hope it is possible to implement for BeamNG as current engine has high single core bind from textures/drawcalls, shadows rob same single core performance, so maps that use more textures on view are getting higher hit in fps than others.

With Vulkan more threads CPU(s) can do is more important again.

With Gfxbench you can benchmark your system with DirectX and with Vulkan, test scene is bit simple in terms of CPU load though, but when I did check CPU load with Afterburner there was considerable difference between Vulkan and DirectX, even it did not show up at end score because I have fastest single core performance possible at the moment.

I would think that with your 1080Ti and W3690 you might get some difference as DirectX11 test might be enough heavy for your CPU while Vulkan might be able to keep your GPU fed better thanks to threading.