Promised performance increases observed - with Patch 16.1 - updated 05.02.2015

[RedFox]

The major bottleneck in this game is whatever script they are using to run the AI and pathfinding. I notice that in large battles,especially siege battles, that when you command a huge group of units to move from A to B that there is a 5 second 'pause' while it computes everything on the CPU. They really need to trim that portion of their code,or offload the computations somehow to the graphics card-which has plenty of computational power to spare, on high graphics it uses less than 50% of my gpu's potential.

Is it a delay of 5 seconds or a freeze of 5 seconds? If it's a delay, they've offloaded pathfinding to a separate thread. Pathfinding is arguably the second hardest thing to get right in game programming, even in AAA titles. I wouldn't be surprised if they're having some serious issues there. Still, I'd think they would keep the working pathfinding they had in Shogun II...

You can't really offload pathfinding onto the GPU that efficiently. Maybe if you had to calculate paths for 2000 soldiers in parallel, then you might make use of it, but Total War only deals with large clumps of units in formation, so pathfinding should take no time at all...


It's becoming increasingly obvious how many issues the game has and more points to the simple fact of incompetence on behalf of CA. I think we can stop the discussion regarding performance issues now - it's painfully clear that these issues can't/won't be fixed. Shamefur Dispray, CA.

[Legatus_Aemilius]

Is it a delay of 5 seconds or a freeze of 5 seconds? If it's a delay, they've offloaded pathfinding to a separate thread.

It's a literal freeze of a 5 seconds, and then the game continues lagging as expected.

but Total War only deals with large clumps of units in formation, so pathfinding should take no time at all...

True...but it also depends on what they are using to calculate their pathfinding. I hear that they use floating point numbers, while integers from my understanding actually perform far better. Guess accuracy trumps everytime though. All in all, the algorithm they use is pretty important.

And I agree, performance issues are a fact now(and have been for a while) so it's best to discuss other things!

[RedFox]

Is there anything I could test in Rome 2 to validate he VRAM allocation and usage of Rome 2 and to backup my original post?

I'm afraid there really aren't that many options to collect statistics on DirectX resource usage with an external tool that I know of. Apparently Visual Studio only supports debugging from source and I had little success with AMD's GPUPerfStudio, even though it should support DX10/11 external debugging.

There does seem to be GPUPerfAPI though, might be able to access GPU performance counters for DirectX 10/11 through that. Perhaps then we could collect some data regarding DirectX resource usage and driver delays.

It's a literal freeze of a 5 seconds, and then the game continues lagging as expected.

That's a bit surprising... Maybe they're calculating pathfinding per soldier.

True...but it also depends on what they are using to calculate their pathfinding. I hear that they use floating point numbers, while integers from my understanding actually perform far better. Guess accuracy trumps everytime though. All in all, the algorithm they use is pretty important.

And I agree, performance issues are a fact now(and have been for a while) so it's best to discuss other things!

They could be using floats or they could be using custom fixed point numbers. Either way, floats are reasonably fast for any such calculations, so I'm very sure it's because of non-optimal algorithms, or because they use some inner-platform scripts to implement game logic, which would greatly reduce performance.

[DarrenTotalWar]

PC's are so complex

[Totalheadache]

PC's are so complex

well not really...coding etc and game design perhaps. but we arent talking about anything to complex here, simply bad optimastion. now the code behind it might be complex but we can all grasp what's going on here.

[Leonidas II]

Well , it was shown multiple times (at least in previous titles that use the warscape engine) ,that the engine uses only 1 core of the CPU effectively.
It doesn't spread the load correctly.
Also it renders shadows and parts of the animations on the CPU .Combined it causes a massive bottleneck.
That has been since Rome Total War like this.

Edit : Multicore support was added to the game with empire.
Still it doesn't seem like the engine can use it correctly.
Otherwise we wouldn't have such a massive CPU bottleneck.

Edit 2 : alQamar has published such a good post in a german forum relating to the CPU bottleneck , I wish I could post it here

[Totalheadache]

Well , it was shown multiple times (at least in previous titles that use the warscape engine) ,that the engine uses only 1 core of the CPU effectively.
It doesn't spread the load correctly.
Also it renders shadows and parts of the animations on the CPU .Combined it causes a massive bottleneck.
That has been since Rome Total War like this.

Edit : Multicore support was added to the game with empire.
Still it doesn't seem like the engine can use it correctly.
Otherwise we wouldn't have such a massive CPU bottleneck.

Edit 2 : alQamar has published such a good post in a german forum relating to the CPU bottleneck , I wish I could post it here

<this>

[IGdood]

Looks like CA better step their game up.

http://www.worldsfactory.net/2014/01...10k-characters

[alQamar]

It's probably best to turn off vegetation Alpha. If you implement 'proper Textbook' blending + transparency, it would consist of depth-sorting all visible transparent objects and possibly rendering the object twice in some cases, which gives horrible performance, but great looking transparency. Of course, if one has experience, you can circumvent this by configuring the graphics card for specific alpha-blended rendering. I bet CA went with the Textbook approach, hence bad performance.

If you turn off vegetation alpha, it might disable the expensive depth-sorting.

As for Video RAM usage, 1920x1080 32-bit backbuffer would take 7.9MB, if you add 4x MSAA, it would multiply backbuffer usage by 4: 31.6MB. If you add backbuffer (31.6MB), frontbuffer (31.6MB) and also the depth buffer (7.9MB), you get 71.1MB of Video RAM usage.

Now since they have water reflections, it would require using a framebuffer (custom back/front buffer, not the built-in type). If they do this correctly, they can save 71.1MB and only create the framebuffer and avoid any unnecessary operations. Could probably check that by running RII through VS DirectX debugger. Sadly I don't have RII installed, so I can't check.

So they might end up using up over 150MB of Video RAM just for the rendering device itself. I have no idea what they're filling the rest of video ram with. A simple DXT5 texture takes 1 byte per pixel and +33% for mipmaps, so a 1024x1024 texture would take ~1.33MB or 0.34MB for a 512x512 texture. You can fit a few thousand textures inside say 600MB, leaving the rest for model data and animations. I don't know what they're doing, but 1GB should be enough for high-fidelity graphics.


Was R2 really written by inexperienced programmers? I noticed that their new Aliens title looked pretty good - perhaps they allocated all the experienced devs on that?

P.S. - Sorry for all the technical details, I'm just trying to make a point here. An AAA title should not be this bad.

Hello Redfox thanks for joining this thread and sharing your knowledge that is beyond my horizon.
Is there anything I could test in Rome 2 to validate he VRAM allocation and usage of Rome 2 and to backup my original post?

[Leonidas II]

Rome 2 uses more than 2 GB of Vram . That's my information

@RedFox

May I ask you which subject you have studied ?

[alQamar]

And again Patch 9 is decreasing performance for my rig. Some other agree to have a better performance outside the benchmark and even in the benchmark.

I am really sick of this. The average performance even degraded over all presets I tested, compared to patch 8.1 final! What the heck is happening over at CA?



Rome 2 benchmark - Patch 9 beta, refer 8.1 performance in the OP.

maxed out settings:
on release: 33 fps average
today: 29.5 fps average (-10.6%) -7% compared to Patch 8.1 final

extreme settings:
on release: 41.6 fps average
today: 38.8 fps average (-6.7%)

and now the other part of the story - the praised people with non / not so capable hardware:
ultra settings:
on release: 63 fps average
today: 64,9 fps average (+3 %)

very high settings:
on release: 84 fps average
today: 87,8 fps average (+4.5%)

high settings:
on release: 141,4 fps average
today: 155,8 fps average (+10.1%)

[Tr0tskY]

Rome 2 benchmark - Patch 9 beta, refer 8.1 performance in the OP.

maxed out settings:
on release: 33 fps average
today: 29.5 fps average (-10.6%) -7% compared to Patch 8.1 final

extreme settings:
on release: 41.6 fps average
today: 38.8 fps average (-6.7%)

and now the other part of the story - the praised people with non / not so capable hardware:
ultra settings:
on release: 63 fps average
today: 64,9 fps average (+3 %)

very high settings:
on release: 84 fps average
today: 87,8 fps average (+4.5%)

high settings:
on release: 141,4 fps average
today: 155,8 fps average (+10.1%)

This is quite simple really when you think about it.

This game is CPU bottlenecked. By attempting to shift some of the workload to the GPU, the CPU is now able to work more efficiently, equalling extra frames. Of course, this will only work if your GPU is not bottlenecking either (note that GPU bottlenecking can be from any aspect of the GPU processes, not just reaching max clock speeds or max VRAM usage, it can be much more technical than this)

By moving workloads around, they have shifted where the bottleneck may occur. In other words, the see-saw is the same size but the fulcrum has shifted. It's still not a big enough see-saw though.

Personally, I know that playing on anything other than high unit detail will still bottleneck ANY cpu in the world like absolute crazy when using ultra unit sizes. It's just not doable. Everybody needs to drop this illusion of making it work. It's just not doable on this engine. Instead, the LODs need to be worked on heavily so that they look far more realistic, and the less detailed 3d models also need to be worked on. This kind of graphical fidelity does a lot more to improve the game than the heavy handed approach CA have taken with Rome 2, which ultimately, does not actually work on modern machines due to the limitations of CPU cycles on silicon chips....


I digress. I have personally found that shifting some of the workload onto the GPU has helped me a lot on this patch, even though I have a 4670K@4.2GHz. I know that I can never use the highest unit detail settings and I know that this game will probably never be able to optimise much beyond shifting workloads around and cutting actual graphical "quality" in favour of frames. Once you get this game running at a consistent 45FPS it's difficult to go back to 30 or so. It feels so gross and cheap compared.

[johnny_b_good]

Are you used that in game benchmark tool or benchmarked with fraps or some program in game?

[alQamar]

both I use fraps to have a visual control (just overlay, no logging), and use the benchmark output, which is reliable now since patch 3.

[mcantu]

wow...those results show none of the improvements that others are seeing.

could there be a system specific bottleneck thats holding fps back?

[alQamar]

wow...those results show none of the improvements that others are seeing.

could there be a system specific bottleneck thats holding fps back?

I could not explain myself a bottleneck, I reverted to 8.1 final again to be 100% sure that these decreases are due to the update. Everything is 100% reproducible.
currently I do extensive testings for bottlenecks but the only bottleneck my system has is the GTX670 FTW, still patch 9 uses less VRAM than 8.1 did. The behaviour is completely unresponsible.

If you, or anyone is interested into that specifically you might have a reads here.
http://www.twcenter.net/forums/showt...ormance-issues

[alQamar]

Well some bits of truth there Trotsky but this all does not explain why the game gets slower with every patch especially slower when CA is stating the opposite in their patch logs. It is a farce.

Just for my intention the models they made are just far too complex and that's the reason why Shogun 2 worked so great. If they managed to downsize the game again, Rome 2 would still look good without killing our hardware. Just from my records the game cannot be handled by the nowadays hardware on extreme settings or beyond. I know this sentence might lure some apologists to say "its running fine on extreme" but they will not share their benchmark screenies nor telling us how a 20v20 or 40v40 battle runs objectively when zoomed in. As you said the 3D models are far to complex and CAs high stakes are the reason for our biggest disappointment. The level Shogun 2 had graphically was fair enough and still supersedes Rome 2 technically in certain conditions (Antialiasing, tesselation, HDR, SLI / CF) etc.

Rome 2 is a shame, not because CA is incapable but because they plotted a game and tested it on workstations or 3D servers or something like that, but not ordinary PC hardware. Otherwise they would have noticed that it looks like a blurry mess on very high and high or even below, or it is unplayable on extreme in the benchmark scene. If they missed to recognize this fact, they at least know it since release and I won't stop to point their noses on that, mates.



just my 2 cents.

[Tr0tskY]

No I agree. This game, performance wise, was doomed the moment they opted for these super high polygon models. There is just nothing that can support it and they probably knew it during development but just wanted to go for the graphical wow factor over actual playability.

[alQamar]

it seems so be invited to discuss it in the other thread linked in #91

[Crazyeyesreaper]

Alot of it game engine but Direct X is fairly deficient as well. No matter how well threaded they make the game there are still limitations via Direct X.

After all there is a reason AMD's Mantle API can speed up Battlefield 4 by up to 45% and keep in mind that engine will scale to 8 cores sadly Direct X means not all those cores are used properly regardless Frostbite engine is one of the better optimized engines and it can still see a 45% increase from a better base API.

Between API and Engine limitations CA has hit the wall and hard. They need to start from scratch but having pissed away their money its not gonna happen. CA will be ridden into the ground by SEGA for quick profits.