Optimizing FPS by Tweaking Game Settings

Hey guys,

Ive been trying to get the most out of my graphics card and after extensive Ocing test Im settling for 5.4k scores with 640/890 OCed 8600mgt clocks. Basically I dont want to go higher up for the fear of over extending my card, and I really havent seen too many over 650 to want to go over that bar.

But theres still alot of tweaking to be done in game to improve performance.
Some of the general graphics options ive noticed are:

Anti-aliasing
Shaders low-ultra
Shadows off-ultra
Terrain/World/Model texture quality low-ultra
Terrain/World/Model detail low -ultra
Dynamic lights
VFX effects
Decal Level
Softened smoke
DX10 effects: bloom, haze, grain, ect,
Fill me in if im missing major ones (this is just off the top of my head)

What do you guys usually do to strike out a balance between quality and performance when your system is under performing.
It'll be great to accumulate a magic formula of game settings for the optimal performance increase/quality decrease ratio.

If anyone have examples of a specific setting tweak that worked really work please note.

For example
Cod4 tweak: turn specular map off
Performance increase: 5~8fps increase on high settings w/o shadows, w/o antialiasing,

 

Detailed shadows are always the first to go for me. They're rarely ever worth the fps hit.

 

Textures will rarely have to drop to the lowest settings, provided you have enough vram\bandwidth, as there's little to no performance hit.

Particles and smoke kill fps in Shooting games so always put them low for online play.

 

Turn antialiasing off and max out anisotropic filtering, that is what I always do. It is a lot more noticable than antialiasing and I don't think it is as resource demanding.

 

Yeh I usally set anisotropic filtering 2-3x without noticeable performance hit. Theres also some direct X effects that looks great but really isnt performance intensive at all, bloom for instance is great. Anyone know anything about Decal details?

 

Decal details usually involve visable damage to the environment. Bullet holes, vehicle damage etc. Not a huge impact on performance unless you have a very bad system

 

Well if its multiplayer FPS eye candy is more of a hindrance than an advantage..

Most serious FPS players will turn every graphical setting down to the lowest and even tweak ini files etc to reduce graphical details further! why? Less distractions, easier to spot enemies, more fps etc etc

I don't have my m1530 yet but my friend gets well over 100fps in COD4 on his at 1440*900 with the right config

There are lots of good configs online for COD4, try http://www.team-dignitas.org/ for some pro player's configs

 

Depends on the game and on your system.
The only "magic formula" is to understand what's going on.

You might find this post relevant.

Grossly simplified, there are two basic types of bottlenecks. Either you're pixel-shader limited, or you're not.

If you're pixel-shader (or technically fragment-shader) limited, it means that there are too many pixels that require too much work, for the GPU to keep up.
There are two obvious solutions to this:
1: Fewer pixels (lower resolution)
2: Less work per pixel (simpler lighting or shadowing, disable bloom and postprocessing.

On the other hand, everything else is basically free. You can max out the polycount, texture resolution, animation quality, AI and everything else that isn't computed per-pixel, without losing a single frame per second.

On the other hand, if fragment shaders are not your bottleneck, if something else is taking even longer, lowering your resolution or disabling dynamic shadows (or any other fragment shader effect) won't gain you anything. Then the GPU will just run idle for a part of the time.

Technically, I should make a lot more distinctions, into fillrate, vertex shaders, CPU, RAM, bus bandwidth and a lot of other factors that can become bottlenecks. But usually, it's a fairly good approximation to think of "fragment shaders vs everything else".

Basically, divide the options into "those that have to be done for every pixel", and "those that have to be done per unit of geometry (vertices, polygons, meshes)
if something is done per model (animation, for example), then it doesn't really matter how many pixels it consists of. So feel free to yank up the resolution.

On the other hand, if you're computing dynamic shadows, you have to do that for every pixel, no matter the polycount of the model.

 

Cool post ^

Explains why in games like TF2 lowering texture detail makes very little difference to performance but an increase in shader detail can make a big difference

 

Yep. One thing I don't think I mentioned is that of texture detail (or texture resolution).

GPU's are pretty clever. And they're really good at texture lookups. So much that they don't actually care what resolution the texture is.
Finding the memory address of pixel (x,y) takes the same time no matter the resolution of the texture. Actually reading from that address also takes the same time no matter the resolution.
So generally, high texture resolutions are *exactly* as fast as low ones.

Which means that you don't really save anything much by lowering texture resolution. It saves you a bit of VRAM, which might help in some rare cases, and it makes the texture a bit faster to transfer between system RAM and VRAM. But those are both fairly uncommon operations. Usually, once a texture is loaded into VRAM, it stays there for a while.

 

So how clever is texture loading? Is every texture in the entire map loaded into vram during the loading screen? Or do the textures stream in as you move to different parts of the map? Also if the same texture is used in several locations is it only in memory once?

I'm pretty interested in this stuff, Gotta any good links / sites for reading up on the technical aspects of Games Development? I'm starting 1st Year Computer Science in October and I'm interested in a Games Development Masters too

Thanks

 

Depends on the amount of textures you have, versus the amount of loading screens.
Some games load everything on the loading screens.
Others don't.

Yeah, only once. Anything else would be silly.

Umm... Gamedev.net has a lot of articles you might find interesting, and there's generally a lot of technical content on their forums. Other than that, I suppose just reading up on rendering in general will at least give you a clearer picture of how hard it is to do X, and so you should be able to predict pretty easily whether people actually do it in real games.

 

Good posts Jalf

So like you said we can distinguish performance settings into the following categories:

Influence Impact on Performance
per Pixel High
per Unit of Geometry Medium
per Model Low


Lets get some examples up for each category, hopefully all of them.
Can someone explain the difference between vertices/Polygons/Meshes vs Pixels .
Also whats the difference between 256bit GPU vs 128bit GPU performance wise. And Memory clocks vs Core clocks which is more important for each individual category.

I would but I dont think I can put it as clearly as some other people. And hopefully afterwards we can compose a better guide for gaining GPU to Settings tweaks insights

 

The bus width and memory clocks are two key factors in determining memory bandwidth.
So here's a rough guide:

Bus width (in bytes) X memory clock X 2 (for ddr, double data rate) = Memory bandwidth

This is important as the more memory bandwidth you have the more video memory you can take advantage of. This area is more important for loading high res textures, higher resolutions and is also plays an important part in anti-aliasing.

Everything else really comes down to stream processors, shader clocks and then core clock.

 

The latter two are the same. And my point was that you *can't* prioritize them according to performance. Sometimes the vertex shader *will* have a greater performance impact than all the bloom lighting in the world.

It depends on how much the game uses each. If the game is really heavy on vertex shader effects, then that will be the limit, and then you can crank up the per-pixel stuff as much as you like, with no performance hit.

Not much more than I've done already. A vertex is a single point or node. Three vertices make up one triangle (so what the GPU really does, is position three vertices, and then draw the area in between them)
And a mesh is just a big bunch of interconnected triangles.

The point being that when things are being positioned, moved and deformed, it's done by manipulating the vertices that make up the models in question.
Some simple lighting effects are done the same way. (Compute a color for each vertex, and interpolate over the triangle. This obviously limits how much detail you can get, but it's cheap, because instead of computing a shade for every pixel, you only do it for every vertex.
The pixel shader is responsible for "filling in" the triangles afterwards, computing a color for every pixel covered by the triangle.

Memory clock gives you faster memory access. So you can do things like texture lookups faster.
Core clocks give you faster computations, so complex shaders will execute faster.