How effective Arma 3 really is at AI + simulation + terrain, for better understanding

[St. Jimmy 272]

We should post here some results. This isn't just normal ran a mission and post fps and usage results where scripts, AI and terrain are smashed together. We should post results how well Arma 3 does terrain rendering alone, AI and some other logics alone. Then you could post some results together if wanted/needed.

I'll just make simple AI tests and terrain tests.

I'll edit/post here results when I get them. Takes some days (even weeks) to get all and think what more to test.

Feel free to post your results also with some different specs :)

AI TESTS

How to set-up video settings:

View distance: 3000m. AI seeing is depending on this
Everything else low as possible. 3D resolution @ 50% and my windowed reolution is 1500x850 (doesn't really matter much)
Object view distance can be set-up at 1m with:
setObjectViewDistance 1;

System specs:

i7 2600K @ 4,4GHz without hyper-threading
8GB 2133MHz CL9-11-10-28

Launch parameters:

-cpuCount=4 -maxVRAM=2048 -malloc=tbbmalloc -maxMem=6096 -exThreads=7 -noLogs -nosplash -world=Altis -nopause

Don't care too much if there's xx vs xx amount of AI when I'm not testing combat. It was just easier for me to mark down the number of AI in the test. So if there's 16vs16 and it's a test about AI moving only then they really don't fight against each other or there's just AI only from one side.

Player very far from the acion so no audio simulation even gets to kill the fps/bottleneck.

Watch the sky to free as much as possble CPU from ground rendering.

Basic AI groups fighting each other in VR map. 300m gap (hell of a shooting) simple move command.

16vs16
- Average Total CPU usage 37,6%
- Average Core 0 95,4%, Core 1 24,8%, Core 2 14,3% and Core 3 15,7%
- fps for first minute 745fps-840fps and after 3min 930fps


32vs32
- Average Total CPU usage 38,7%
- Average Core 0 95,8%, Core 1 24,6%, Core 2 16,8% and Core 3 17,6%
- fps for first minute 550-690fps and after 3min 800fps


48vs48
- Average Total CPU usage 40,8%
- Average Core 0 95,4%, Core 1 26,5%, Core 2 20,1% and Core 3 21,0%
- fps for first minute 400-500fps and after 3min 700fps


64vs64
- Average Total CPU usage 42,5%
- Average Core 0 95,7%, Core 1 28,4%, Core 2 22,9% and Core 3 23,1%
- fps for first minute 320-420fps and after 3min 650fps


80vs80
- Average Total CPU usage 56,6%
- Average Core 0 95,4%, Core 1 50,4%, Core 2 42,3% and Core 3 38,3%
- fps for first minute 250-410fps and after 3min 550fps


96vs96
- Average Total CPU usage 60,9%
- Average Core 0 95,6%, Core 1 46,2%, Core 2 52,2% and Core 3 49,7
- fps for first minute 200-340fps and after 3min 460fps


112vs112
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps


128vs128
- Average Total CPU usage 68,9%
- Average Core 0 94,6%, Core 1 65,1%, Core 2 62,2% and Core 3 53,3%
- fps for first minute 120-290fps and after 3min 390fps


144vs144
- Average Total CPU usage 73,1%
- Average Core 0 95,4%, Core 1 66,6%, Core 2 71,0% and Core 3 59,6%
- fps for first minute 85-215fps and after 3min 310fps


160vs160
- Average Total CPU usage 61,7%
- Average Core 0 94,5%, Core 1 58,8%, Core 2 51,0% and Core 3 42,7%
- fps for first minute 70-130fps and after 3min 270fps

CPU usage comparsion picture between 32vs32(red)-80vs80(green)-160vs160(blue). I could post more but not necessary IMO. This is just to see some reference.

http://i.imgur.com/vsjTJ30.png

THOUGHTS

This test alone doesn't tell too much. The first core high usage is Arma thing because your core 0 is used always at full when you don't have any vsync or something that brings fps down.

Core usage rises the more AI there is BUT there's also firefights in here and likely bullet physics take big part in this also.

I need to test more the other things to get better picture. 320 AI fighting against each other when you're not drawing any terrain and not close to need to hear all the gun shooting gives nice fps. But very far even from the over 1000 AI battle what richiespeed recorded so intersting to see when I go to more AI if Arma 3 can even handle that amount of firefight. At least that was somehow possible in Arma 2

Fps naturally rises the more the scenario was going because some AI likely died from that close battle.

I'll open this when I've results

Basic AI groups fighting each other in VR map. 1000m gap simple move command and danger mode.

16vs16
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

32vs32
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

48vs48
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

64vs64
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

80vs80
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

96vs96
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

112vs112
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

128vs128
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

144vs144
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

160vs160
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

THOUGHTS

Coming...

Basic AI groups fighting each other in Altis map. Heavy towns.

16vs16
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

32vs32
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

48vs48
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

64vs64
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

80vs80
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

96vs96
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

112vs112
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

128vs128
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

144vs144
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

160vs160
- Average Total CPU usage
- Average Core 0, Core 1, Core 2 and Core 3
- fps

THOUGHTS

Coming...

Basic AI groups moving far away in VR map. Simple move command. Only movement order happening so AI is just running forward on the flat map.

16vs16
- Average Total CPU usage 37,9%
- Average Core 0 95,3%, Core 1 27,5%, Core 2 12,7% and Core 3 16,3%
- 820-870fps

32vs32
- Average Total CPU usage 35,8%
- Average Core 0 95,5%, Core 1 21,8%, Core 2 11,9% and Core 3 13,8%
- 760-810fps

48vs48
- Average Total CPU usage 36,1%
- Average Core 0 95,0%, Core 1 20,9%, Core 2 13,0% and Core 3 15,4%
- 650-700fps

64vs64
- Average Total CPU usage 35,6%
- Average Core 0 95,0%, Core 1 20,7%, Core 2 11,8% and Core 3 15,0%
- 550-580fps

80vs80
- Average Total CPU usage 35,6%
- Average Core 0 94,6%, Core 1 18,7%, Core 2 12,9% and Core 3 16,2%
- 490-530fps

96vs96
- Average Total CPU usage 35,0%
- Average Core 0 94,5%, Core 1 17,3%, Core 2 13,0% and Core 3 15,4%
- 440-480fps

112vs112
- Average Total CPU usage 35,4%
- Average Core 0 93,9%, Core 1 17,3%, Core 2 13,6% and Core 3 16,6%
- 380-420fps

128vs128
- Average Total CPU usage 35,4%
- Average Core 0 94,2%, Core 1 17,9%, Core 2 13,3% and Core 3 16,3%
- 320-360fps

144vs144
- Average Total CPU usage 35,8%
- Average Core 0 93,8%, Core 1 18,3%, Core 2 14,1% and Core 3 16,9%
- 320-360fps

160vs160
- Average Total CPU usage 36,3%
- Average Core 0 93,1%, Core 1 18,3%, Core 2 15,4% and Core 3 18,4%
- 270-300fps

320vs320
- Average Total CPU usage 38,4%
- Average Core 0 95,2%, Core 1 20,6%, Core 2 18,0% and Core 3 19,9%
- 100-130fps

640vs640
- Average Total CPU usage 42,3%
- Average Core 0 83,4%, Core 1 32,6%, Core 2 21,6% and Core 3 31,4%
- 30-40fps for 4 minutes and after that drops to 4-14fps.

CPU usage comparsion picture between 64(red)-160(green)-320(blue) AIs.

http://i.imgur.com/k4MRPtB.png

Not much difference so I'll show some comparion with higher AI counts.

CPU usage comparsion picture between 320(red)-640(green)-1280(blue) AIs

http://i.imgur.com/qyFt4U0.png

In this 1280 AI movement goes pretty well about 4min the fps bogs down from 30-40fps to 4-13fps

THOUGHTS

As you can see the CPU usage doesn't differ no matter what amount of troops are moving on the battlefield and the CPU usage paragraphs are also pretty identical. Everything is pushed on just core 0. Fps just falls down by the amount of moving AI.

That very high AI count shows that 1280 AI in groups moving isn't an option in this game. Something is maybe bugged or the AI in Arma 3 seems to be very demanding. It goes "well" (30-40fps) for 4 minutes but after that it drops badly.

Basic AI groups moving far away in Altis map. Simple move command. Heavy town in the middle.

16vs16
- Average Total CPU usage 35,7%
- Average Core 0 96,1%, Core 1 18,9%, Core 2 12,1% and Core 3 15,8%
- 390-420fps

32vs32
- Average Total CPU usage 38,1%
- Average Core 0 95,1%, Core 1 21,2%, Core 2 17,2% and Core 3 19,1%
- 310-350fps

48vs48
- Average Total CPU usage 39,1%
- Average Core 0 94,6%, Core 1 23,0%, Core 2 17,9% and Core 3 20,9%
- 280-320fps

64vs64
- Average Total CPU usage 37,9%
- Average Core 0 95,6%, Core 1 20,4%, Core 2 15,6% and Core 3 20,0%
- 240-280fps

80vs80
- Average Total CPU usage 40,2%
- Average Core 0 95,1%, Core 1 24,1%, Core 2 19,1% and Core 3 22,3%
- 200-250fps

96vs96
- Average Total CPU usage 41,8%
- Average Core 0 95,3%, Core 1 26,5%, Core 2 21,0% and Core 3 24,5%
- 180-220fps

112vs112
- Average Total CPU usage 44,1%
- Average Core 0 94,9%, Core 1 29,4%, Core 2 24,4% and Core 3 27,5%
- 150-190fps

128vs128
- Average Total CPU usage  46,6%
- Average Core 0 94,3%, Core 1 31,9%, Core 2 26,4% and Core 3 29,6%
- 120-160fps

144vs144
- Average Total CPU usage 44,4%
- Average Core 0 95,5%, Core 1 29,5%, Core 2 25,1% and Core 3 27,8%
- 110-150fps

160vs160
- Average Total CPU usage 48,5%
- Average Core 0 96,0%, Core 1 34,5%, Core 2 30,6% and Core 3 32,6%
- 80-130fps

And for reference to the last one 320AI running on a salt field in Altis to see real difference because the map naturally alone eats some CPU. As you can see it's heavier for AI going through towns when they've to maintain their formation and choose paths.

160vs160
- Average Total CPU usage 35,3%
- Average Core 0 89,8%, Core 1 17,7%, Core 2 13,7% and Core 3 19,9%
- 80-130fps

Diagram of CPU usages with 320AI. Red RV, Green Altis on salt field, Blue Altis on heavy towns.

http://i.imgur.com/WhzFIpn.png

64(red)-160(green)-320(blue) AIs CPU usage comparsion.

http://i.imgur.com/xLSucgC.png

THOUGHTS

As you would expect running through the towns is heavier for CPU when AI needs to think more. Then we can also compare already how a real map added (with lowest quality possible except view distance at 3km) alone effects some performance. It seems to affect only to the first core because there's fps drop but overall usage is very much the same compared to RV map as you can see from the diagram. Other cores usage rises bit the more AI and pathfinding there is but it likely starts to fall again if tested with overkill amount of AI.

The more AI there was the more core usages were rising so those aren't pushed through just one core alone.

I'm not sure is forest and other heavy clutter test really needed after this. Maybe later but I'll start to test some other things now.

AI very close to each other on RV map with never fire command and player far away.

So the AI knows and is in combat but it just doesn't fire. This is where we see how much basic AI fighting eats resources.

16vs16
- Average Total CPU usage 40,3%
- Average Core 0 96,2%, Core 1 27,5%, Core 2 17,3% and Core 3 20,3%
- 570-700fps

32vs32
- Average Total CPU usage 44,9%
- Average Core 0 95,5%, Core 1 30,6%, Core 2 25,0% and Core 3 28,3%
- 460-490fps

48vs48
- Average Total CPU usage 58,7%
- Average Core 0 94,8%, Core 1 47,6%, Core 2 42,4% and Core 3 49,7%
- 420-460fps

64vs64
- Average Total CPU usage 65,3%
- Average Core 0 87,2%, Core 1 61,0%, Core 2 56,3% and Core 3 56,7%
- 210-260fps

80vs80
- Average Total CPU usage 64,5%
- Average Core 0 86,7%, Core 1 66,8%, Core 2 51,1% and Core 3 53,3%
- 130-170fps

96vs96
- Average Total CPU usage 70,5%
- Average Core 0 79,2%, Core 1 78,1%, Core 2 62,5% and Core 3 62,3%
- 85-95fps

112vs112
- Average Total CPU usage 69,0%
- Average Core 0 80,4%, Core 1 78,0%, Core 2 57,9% and Core 3 59,7%
- 75-85fps

128vs128
- Average Total CPU usage 73,8%
- Average Core 0 79,2%, Core 1 81,1%, Core 2 66,4% and Core 3 68,3%
- 70-79fps

144vs144
- Average Total CPU usage 76,5%
- Average Core 0 80,1%, Core 1 83,0%, Core 2 70,8% and Core 3 72,1%
- 65-75fps

160vs160
- Average Total CPU usage 76,0%
- Average Core 0 80,5%, Core 1 83,3%, Core 2 68,8% and Core 3 71,4%
- 60-70fps

CPU usage comparsion between 32vs32(red)-80vs80(green)-160vs160(blue)

http://imgur.com/FTAMk5i

THOUGHTS

Performance starts to drop quickly in these tests the more AI there is. But CPU usage is pretty high so high AI numbers are really multi-core. As you can see core 0 usage drops bit the more AI is added but other cores are then used nicely.

What's bit interesting is that the fps is worse than in 300m gap AI fights. I would've easily guessed it's other way around but this seems to really show then that AI using tactics constantly and when it's ready to fire but not allowed it's pretty heavy.

Maybe some vehicle/aircraft performance testing how much they alone have effect

...

Player in the acion. In big battles this really bottlenecks likely because of audio simulation.

...

continue maybe in the next post which is about terrain and objects (including models) and maybe AI stuff if here's not enough space.

Edited January 27, 2015 by St. Jimmy

[St. Jimmy 272]

Only two simple VR map test results there but maybe someone is interested. Also maybe this now reserves me a new post?

Also thoughts so far:

That always >90% Core 0 usage was interesting thing that I haven't really before noticed. It's that core that kind of says how much fps you have. As you can see at 800fps the core 0 usage is at max

Edited January 18, 2015 by St. Jimmy

[343guiltyspark 10]

The AI in this game is not much different from operation flashpoint.

the game is designed to run with lots and lots of triggers and waypoints. the AI just react to being shot at.

[warlord554 2065]

You posted your basic system specs. I'd like to see your ingame video settings during these tests

[brightcandle 114]

The game is basically single threaded by its design (this is the time of a single frame captured, 12 grey lines at the top represent all my cores 6c12t 3930k@4.4Ghz):

https://dl.dropboxusercontent.com/u/3638175/Capture%20Palagia.png

This is why you see core 0 so high on load and everything is low, the game is basically single threaded except for the bits of rendering that are a bit parallel and I suspect that is the graphics driver or DX doing that and not BIS.

[St. Jimmy 272]

You posted your basic system specs. I'd like to see your ingame video settings during these tests

It's in the post but maybe you missed it. AI tests will be everything as low as possible (object view distance at 1m) except view distance at 3km (because view distance has some effect how well they can spot everything and maybe even pathfinding) to test as much AI only as possible. Later can come some more about how graphics also effect more.

Some Altis heavy town AI crossing and comparsion that for AI moving on flat added. Slowly but steady testing :P Really takes some time to record and compare things etc. Bit big project and I'm constantly making adjustments in my mind what kind of tests are needed. Not much of results so far more like confirmations of some facts how things go.

[Greenfist 1863]

Here's a small video I made for another thread about AI's performance at a distance:

There's 200 opposing units fighting 3.6 km away. As the camera moves closer you can see the fps drop gradually, but then jump back up as the camera teleports back.

The game was started without sound, because that had a huge impact on performance ( <-200 invincible soldiers firing constantly).

By the way, is your fps really at 700-800 in some of those test with 50-100 AI? :eek: (I can't get past 150 no matter what I do.)

[IndeedPete 1038]

Ah, interesting topic! I'll keep an eye on this thread, keep up the testing. ;)

[St. Jimmy 272]

Here's a small video I made for another thread about AI's performance at a distance:

There's 200 opposing units fighting 3.6 km away. As the camera moves closer you can see the fps drop gradually, but then jump back up as the camera teleports back.

The game was started without sound, because that had a huge impact on performance ( <-200 invincible soldiers firing constantly).

That's a nice test. Yeah the audio in bigger battles really drops performance. I didn't even remember that there was a no sound parameter. Good that you mentioned it!

By the way, is your fps really at 700-800 in some of those test with 50-100 AI? :eek: (I can't get past 150 no matter what I do.)

I'm using MSI afterburner + RTSS to see my fps. I believe it should be right because it's gradually dropping the more things are introduced and those tests have been made with the very minimum video settings in pretty empty places except view distance which is 3000m. I've put 3D resolution at 50% and object view distance is at 1m :D

Yes the fps is really that high when there's pretty much nothing to draw and only simulation going on.

I just added Never Fire AI tests. What's interesting that the fps is worse when the AI is in Never Fire mode and using constatly some tactics than when they're allowed to fire and battle. And in the never fire but constantly aware of each other mode CPU usage is higher so it's pretty bottlenecking situation.

Edited January 27, 2015 by St. Jimmy

[chrisb 196]

Terrain VR: Empty other than AI & me. Total on the terrain was 2289.

There seems to be a max of 1152 units per side! I removed 2 groups of 8 units from each side giving the above total (inc me).

Res: 1920 x 1080 @ 16:10

FSAA: 4x AA

ATOC: all grass & trees

PPAA: disabled

AF: Ultra

Sampling: 100%

Texture: Ultra

Objects: Ultra

Terrain: Ultra

Shadow: Ultra

Particles: High

Cloud: Ultra

PIP: Disabled

HDR: Standard

Dynamic Lights: Low

Overall VD: 3000

Object: 1736

Shadow: 100

2288 AI one simple move waypoint@

These are on the terrain 6.7km from me.

Just as it is (Preview): 7.7 fps average over 5mins

With Simulation Manager Module set @ 2000 radius: 193.9 fps average over 5mins (obviously they are there, the units, but they're not there, until they get within the radius, just sort of cache thing..;)).

Edit: I should point out, the settings are as they are, because they are my standard settings for A3. Other than the VD which is usually 5000. I do remove grass/clutter every now and then depending on mission and/or testing i.e. terrain @ low. But not here it was set @ Ultra as said.

Always use your sim/man if you want to use large numbers, set the radius to whatever suits your team/squad i.e. more if you have a sniper, less if not.. But beware they will lose any AI mod behaviour you might be running, well the ones I've tested they do.

Edited January 27, 2015 by ChrisB

[ceeeb 147]

There seems to be a max of 1152 units per side! I removed 2 groups of 8 units from each side giving the above total (inc me).

The limit is 144 groups per side, which is well known since OFP. I'm not aware of any limit to the number of units per side/group, I do remember testing a group of 1000 units in Arma.

[chrisb 196]

The limit is 144 groups per side, which is well known since OFP. I'm not aware of any limit to the number of units per side/group, I do remember testing a group of 1000 units in Arma.

That would be right, based on groups of eight. Of course you could have 1152 of all factions, which would be quite a few if you include civilians (doubt they count those), but certainly Independents. But really, I never go that far with AI on the terrain. It is possible to run quite a number with the sim/man or a cache system, provided you don't meet hundreds at a time. Not that you couldn't, many systems would take it. But it would be pretty useless because if your playing using AI (we do), then they would be simple target practise, with little in the way of any sense, because you would be running too many at once.

I would have thought they may have stopped a limit, as its really not needed because systems would decide. However, the series has always been o.k. regards amounts on pc. The only really low restriction was for OFPElite (orig xbox), which was a limit of 70 if I recall correctly.

Edited January 27, 2015 by ChrisB