To optimize my scene file, I went through and gradually took everything out to see how big a performance hit each item was giving. I thought other folks might like to see the results--and I would love it if anyone could help explain some of the strange things I found. I also think it would be great if anybody else wants to post how much various things affect their own performance.

Also, maybe I just didn't find it, but there doesn't seem to be much commentary at all on frame rate and performance. The Wiki doese have a few very general frame rate hints. If anybody can think of other resources for optimizing games, that would be very cool.



I did all my testing on a slow machine. This is important because things that slow down my slow machine don't affect my fast one at all. For instance, reducing polys doesn't benefit my fast card any because it hasn't reached any kind of bottleneck.

Here are the highlights that I picked up:
1. I started off with about 600 WMB polys and 1800 entity polys (2767 portals). (The main character alone is about 900). One of the main things that helped was bringing the number of entity polys down to about 1000. That sped me up from 20 to 30 FPS. I was very surprised that it had a hard time with so few entity polys. A single character in our PS2 games had more than that.

2. The other big observation I had was that none of the things that I thought would make a big deal really mattered. Dynamic lights, stenciled shadows, a hundred dirt sprites, visibility calculations, and block texture resolution. None of it really made any difference. (Except--staring at two big dynamic lights, full-screen, with stenciled shadows turned on slowed it down over 30%.)

3. The other strange thing is that when EVERYTHING was gone from the level, jumping down from 800x600 to 640x480 DOUBLED the FPS. I could understand a 1/3 improvement, but not double.

4. This likely says a lot more about my own card than about general responsiveness of the engine, but it's a start.

5. Also, there seem to be little plateaus. For instance, it seemed to stick at almost exactly 19.4 for many different changes--and then again at about 29.2. Similarly, my fast machine (Dual 2GHz & Quadro4) usually sticks at 63.2 FPS, even in just an empty room. But in some complex rooms (like the dynamic lights/stenciled shadow test below) it'll hit 75 to 85 FPS. Just kinda strange.

Test machine info:
Dell Latitude C800
847 Mhz
Windows 2000 Professional
512 MB RAM
ATI M4 AGP4x
(I think it has either 32 or 64MB of RAM)



Variation #1:
Modification FPS
Full-Screen, 800x600, Hi-Rez 19
Re-compiled with Visibility Calculations "OFF" 19
Turned Stenciled Shadows OFF 19
Turned the steam particles OFF 19
Turned the Dynamic Torch lights OFF 19
Turned Animated Sprite fire OFF 19
Deleted about 100 dirt Sprites 19
Cut texture maps from 256 to 128 19
Deleted all entities except torches 22
Deleted torches 29
Replaced all level geometry with single hollow block 29
Deleted the character 30



Variation #2:
2 dynamic lights 29
Dynamic lights OFF 29
Add Stenciled Shadows 19



Variation #3 (same as #1 but at 640x480):
Re-compiled with Visibility Calculations "OFF" 23
Turned Stenciled Shadows OFF 28
Turned the steam particles OFF 28
Turned the Dynamic Torch lights OFF 28
Turned Animated Sprite fire OFF 28
Deleted about 100 dirt Sprites 29
Cut texture maps from 256 to 128 29
Deleted all entities except torches 29
Deleted torches 29
Replaced all level geometry with single hollow block 29
Deleted the character (or just go to first-person) 59

Another interesting thing happens when you press F-11 twice. It brings up the polygon tesselation highlighted in red. The strange thing is that in some cases, this will cut my FPS in half. But at other times, it will actually push it higher than I've ever been able to get it (by about 25%!). I hope somebody can help us understand things like this that will let us push our performance higher.

Your Screen Shots show frame rates which A6 at its current stage of development as a general guide to a standard i.e. 50 - 75 fps is now well capable of sustaining in extremely complicated levels.

I have numerous levels of differing sizes and all can sustain those types of frame rates - in most instances for many users 60 - 75 fps is now a consistant reality - given a recent machine spec.

Examples I have would be average of indoor level 5000 portals and up to 400 (yes 400) entities). Outdoor level 16000 portals with as yet small but growing number of entities = all showing 60 - 70 fps.

I am finding that .mdl entities have little bearing on framerate so you can use a lot if needed.

Main frame rate killer seems to be large textures tiled repeatedly increasing portals and wide camera views (panoramas) in outdoor areas. If you design levels well much of this can be avoided. Dont use such textures where they are not necessary and dont let the view show too much at any one given point in a level. (seeing down a long city street as far as the horizon - you cant do that and expect to keep frame rates high) - makes sense though should be stressed if your frame rates are low check these things first.

dynamic lights, entity shadows, materials, shaders, all entities, sounds, particals all are rendered in RUN time. so more you got visible at any given time, more time it takes to do all them calculations each frame.

level geometry, more portal surfaces and more varied shadow/light mapping visible the more you want to up the presision levels in the build.

F11 portal hightlighing is all run time rendering.

hardware and driver settings has a big effect on runtime.


compairing performance between a PC and a consoul machine is like comparing a Porsche to a Mack Truck. can't. each are designed with specifics in mind that the other doesn't.

Quote:

---

dynamic lights, entity shadows, materials, shaders, all entities, sounds, particals all are rendered in RUN time. so more you got visible at any given time, more time it takes to do all them calculations each frame.

level geometry, more portal surfaces and more varied shadow/light mapping visible the more you want to up the presision levels in the build.

---

Actually, the interesting thing about this test is that those things had practically zero (negligible) impact on performance. More precise level building--zero impact. # of entities--zero impact. Entity shadows-zero. Shaders and materials--zero (since most cards still revert to A5 mode). Sounds--zero. Removing all of those things didn't do anything to speed up the frame rate.

Now, if you were worried about really fast systems with no geometry bottle-neck, then those concerns would kick in. But for the average consumer PC, it seems that those concerns simply don't matter much compared to geometry count. I think that transparency is also a big bottleneck, but I'm not sure. I'm trying to figure out what else really does matter, but so far all those high-end features are all standing in line behind other, far more significant bottle-necks. Again--on the AVERAGE consumer PC.

are you compairing these performances based on standing still in your tests or moving? I've built levels where while moving ( even just changing camera pan angle) the frame rate can drop significatly, soon as you stop movement it jumps back up. This all falls back on what appears into view that needs to be rendered.

vid card can make a differnace too as you noted. the engine will revert back to A5 ( all software rendering mode) either by a not direct x supportive car or by scripting it to change. Some of the newier vid cards just out perform others on hardware rendindering speeds (A6 mode) Like i use a gforce 5200 FX card, always have stuck with NVida since the TNT2 GLFire cards but now they just can't seem to perform like the ATIs lately. I've yet to get my games to get past the 60FPS problem with this card.

the video resolution you mentioned is easy to understand. Monitors have a Hrz rating. this is the moitors refreash rating based on screen resolution (i.e. 800x600 ,1024x768 ...) the higher the resolusion is upped on the monitor the lower the hrz rate drops ( slowing how fast a screen is redrawn).

In newier vidcards and monitors a vidcard can detectif its output Hrz is over what the monitor can handle at a resolution setting and will adjust itself.

unfortunaly you will still find many monitors on the market today that arn't ment very good at performace at higher resolution settings

example my monitor is rated at 85 hz at 1280x1024 resolution my GForce card is rated at 120 at the same resolution, so the vid card has to scale back to the 85 (and my monitor is a bit above the average monitor for hrz rating)
but a typical monitor hits 60hrz at

this all reflects back to your frame rate. basicly its 1 frame per hz at peak performance. so when shopping for monitors, its better to watch the hrz rating range then the viewable size. some good bucks can buy some very nice monitors rated at 120hz over 2048x rsolution ones typicaly used in CAD engineering and professional graphic work

Just turn off vertical sync in your display properties to eliminate the factor Grimbar mentions. My monitor incidently does 800x600 and 640x480 at the same video bandwidth which is also the monitor's maximum bandwidth. Limits don't start to kick in until 1024. Your monitor may be the same here so sync rates may not be a factor at all for you.

Interesting figures.. wish I could add more but I need to get a hold of a low end pc myself before I can do much performance testing. I am also observing that my levels are not stressing my dev machine noticably. My main annoyance is build times.

Quote:

---

always have stuck with NVida since the TNT2 GLFire cards but now they just can't seem to perform like the ATIs lately.

---

@Gimber
Ill say that this is not true. I still use NVidia and they still rock in 3D world. I have some games that i even cannot run at all on ATI. NVidia has much more stable drivers then ATI.

To make a fair comparison, I had to do the test standing in the same spot. If I turn just a little bit to the left or right, I pick up or lose dynamic lights, transparency, geometry, etc. so my frame rate fluctuates quite a bit if I test it looking at different things, and I wanted to rule that out.

As an example, dynamic lights normally don't affect my frame rate, but if I move so that a dynamic light is filling my frame, it slows performance significantly. But generally, geometry is the bottleneck. That's why I was quite surprised that compiling with higher precision didn't help more.

r00tsh3ll:
The thing I need to keep in mind about video cards is fairly different from the concerns I generally see posted on this forum because I am selling to the average consumer, not the hardcore gamer. NVidia and ATI are both outsold by Intel--especially in the average PC market. They're not nearly as sexy, but they're more prevalent.

Also, I have to worry about cards made 3 to 5 years ago, rather than what's going on in the market lately. I mean, if a consumer has an ATI or NVidia from the last few years, I have no worries that they'll be able to run the game at a perfectly acceptable rate.

I WOULD like to offer some nice eye-candy for folks whose video cards can handle materials and shaders. However, my estimate is that very few video cards in the general populace can handle these effects yet. JCL recently quoted a survey that reports the average PC these days is a 2 GHz with a GeForce or Radeon--but that's a very misleading survey--at least for the average buyer of a 3DGS-made game. This survey was taken by Valve for Half-Life 2, and basically was directed at hard-core gamers. I certainly believe that the average hard-core gamer has blazing PC specs, but the average PC is going to generally have about 1/4 to 1/2 that power, so unless I decide to compete head-to-head with Half-Life 2, I need to make sure my game will run well on the minimum specs for 3DGS: about a 500 MHz with a decent video card. And as far as I can tell, "decent" means that it was a cheap card from 3 years ago or a very nice card from 5 years ago. Anything other than that can't seem to run anything decent and they're lost to me.

However, I'm just estimating on a lot of this stuff, so I love the insights that you folks are sharing about what the real limitations are of older cards.