Many people come into this forum and ask questions like "what does Texture Sharpening do? What is the difference between Application/Balanced/Aggressive in the driver panel? The answers have been discussed, but they are usually buried in some old driver thread. You can find Detonator drivers at the following places.
*note* Application/Balanced/Aggressive is known as Application/Quality/Performance in driver 43.45.
*note* Application/Balanced/Aggressive is known as Quality/Balanced/Performance in driver 43.51.
*note* Application/Balanced/Aggressive is known as Quality/Performance/High Performance in driver 44.03.
The purpose of this thread is to create easy access for everyone, and a reference point that can be quickly linked to whenever someone asks about these settings.
It may be prudent to take a look at the summary of my findings, here. I provide concise conclusions in that post, as well as dicuss the driver settings as they pertain to GeForceFX cards.
Tests run with a GeForce3, WinME driver 42.30 and DX8.1
SBA and FastWrites disabled by default
Percentages are the dropoff from the baseline IQ(ie no FSAA or AF)
2x FSAA: 3946 -21.9%
4x FSAA: 2700 -46.5%
2x AF: 4478 -11.3%
4x AF: 3881 -23.1%
8x AF: 3472 -31.2%
2x FSAA: 5080 -22.8%
4x FSAA: 3370 -48.8%
2x AF: 6014 -8.6%
4x AF: 5429 -17.%
8x AF: 4913 -25.4%
2x FSAA: 5110 -22.8%
4x FSAA: 3377 -49.0%
2x AF: 6087 -8.0%
4x AF: 5479 -17.2%
8x AF: 4955 -29.3%
It is clear in drivers previous to 4x.xx series, before nvidia implemented this system of A/B/A, that drivers were most closely matched to Balanced. So let's see how much faster/slower Aggressive and Application are from Balanced. Notice that both Balanced and Aggressive suffer less of a performance drop when enabling AF than Application does.
Application: -23.3% -22.3% -19.9% -25.5% -28.5% -29.3%
Aggressive: +1.0% +1.0% +0.2% +1.0% +1.0% +1.0%
Application seems to be 20% slower than Balanced or Aggressive across the board. I have no idea why.
Now here's some FSAA/AF benchmarks on Balanced with and without Texture Sharpening.
2xFSAA/2xAF: 4679 -28.9%
2xFSAA/4xAF: 4134 -37.2%
2xFSAA/8xAF: 3694 -43.9%
4xFSAA/8xAF: 2781 -57.7%
2xFSAA/0xAF+TS: 4868 -26.0%
2xFSAA/2xAF+TS: 4131 -37.2%
2xFSAA/4xAF+TS: 3696 -43.8%
2xFSAA/8xAF+TS: 3697 -43.8%
4xFSAA/4xAF+TS: 2782 -57.7%
4xFSAA/8xAF+TS: 2786 -57.7%
From the numbers it seems like all Texture Sharpening does is enable an extra degree of AF up till the max of 8x. However, the numbers for 2xFSAA+TS do not match up with 2xFSAA/2xAF.
2x FSAA: 152.6 -47.1%
4x FSAA: 100.1 -65.3%
2x AF: 240.7 -16.5%
4x AF: 196.2 -31.9%
8x AF: 181.7 -37.0%
2x FSAA: 152.5 -47.6%
4x FSAA: 100.1 -65.6%
2x AF: 252.6 -13.2%
4x AF: 220.6 -24.2%
8x AF: 209.9 -27.9%
2x FSAA: 152.6 -47.1%
4x FSAA: 100.1 -65.3%
2x AF: 252.6 -12.5%
4x AF: 220.3 -23.7%
8x AF: 210.0 -27.2%
Same story here. AF has less of a hit when using Balanced and Aggressive.
Benchmarks with AF/FSAA on Balanced
2xFSAA/2xAF: 138.7 -52.3%
2xFSAA/4xAF: 124.6 -57.2%
2xFSAA/8xAF: 118.6 -59.2%
4xFSAA/2xAF: 97.6 -66.5%
4xFSAA/4xAF: 89.5 -69.2%
4xFSAA/8xAF: 89.4 -69.3%
2xFSAA/0xAF+TS: 152.5 -47.6%
2xFSAA/2xAF+TS: 124.6 -57.2%
2xFSAA/4xAF+TS: 118.6 -57.2%
2xFSAA/8xAF+TS: 118.6 -59.2%
4xFSAA/0xAF+TS: 97.5 -66.5%
4xFSAA/2xAF+TS: 92.8 -68.1%
4xFSAA/4xAF+TS: 89.5 -69.3%
4xFSAA/8xAF+TS: 89.4 -69.3%
Again, there are some weird results regarding Texture Sharpening.
2x FSAA: 141.8 -48.9%
4x FSAA: 97.8 -64.7%
2x AF: 230.3 -17.0%
4x AF: 190.1 -31.5%
8x AF: 177.4 -36.0%
2x FSAA: 141.8 -49.3%
4x FSAA: 97.8 -65.0%
2x AF: 241.9 -13.5%
4x AF: 213.2 -23.7%
8x AF: 204.5 -26.9%
2x FSAA: 141.8 -49.3%
4x FSAA: 97.7 -65.0%
2x AF: 241.1 -13.7%
4x AF: 213.2 -23.7%
8x AF: 204.6 -26.8%
I added Quake2 Trilinear results just for kicks. You can see that the percentage drops match up with Bilinear, but Bilinear obviously has better performance.
Quake3 Demo w/43.45
Since I apparently don't have/can't run demos with retail Quake3, I used the demo. The results speak for themselves.
Tenebrae Quake w/43.45
High Performance: 6595
Whatever was slowing down "Application" mode in previous drivers seems to have been changed/fixed in 44.03.
Quake2 Trilinear w/44.03
Quality 0xAF: 275.0
Quality 8xAF: 204.4
High Performance 0xAF: 275.6
High Performance 8xAF: 208.3
It appears that nvidia changed their AF algorithms across the board with 44.03, so that Quality is just as fast as High Performance(or Performance, since HP should be as fast as or faster than P).
I tried using RivaTuner's OpenGL AF patch script. In Quake2 Trilinear I achieved 177.3 fps vs 177.4fps with 4xFSAA/8xAF Application and 205.1fps vs. 204.5fps with 4xFSAA/8xAF Balanced. All in all I have to conclude that either the patch script doesn't work anymore with newer Detonators or nvidia already incorporated the optimizations into the drivers.
Apparently the patch script does work in driver 43.51. Discussion.
I also wanted to try the FSAA optimizations that supposedly exist in RivaTuner as documented in this thread:
I should note that RivaTuner users should be familiar with the OGL AF settings "performance optimization" and "quality optimization." Application uses "quality" and Balanced/Aggressive use "performance."
I think some screenshots are in order to find out what, if any differences there are to IQ between quality/performance AF as well as what Texture Sharpening does, but I'll get to that later. Maybe later today, maybe next week.
edit: added new information and benchmarks with the 43.45 reference drivers.
edit: added new information about the 43.51 drivers.
edit" added new information and benchmarks with the 44.03 reference drivers.
Excellent work ... that was very well written ... thanks for the analysis.
The RT setting you cant find is available under the poweruser tab, it is not a basic setting.
anyway, i have taken screenshots in Quake3 with colored mipmaps, now i'm just need to upload them somewhere.
Ok, here are some Quake3 screenshots with r_colormiplevels 1. All screenshots have 2x FSAA enabled. I'm not sure why the shots turned out so dark, maybe it had something to do with print screen.
edit: As is clearly seen, 0x+TS is not the same as normal 2xAF. 2xAF is better.
Screenshots to come later at 4x FSAA, and maybe some shots from either U2 or UT2003, although since I don't know if you can view colored mipmaps in those games it will be harder to discern IQ.
it is the 4th key listed under D3D on my system in RT.
Now, if you are not using the Alias display, but are just using the actual Name display for the keys, you will not see it, you will see a string of numbers. You need to click on the word "Name" next to the little green arrow at the top of the column so that it says "Alias", and you should now see the names of several more keys that previously just looked like numbers.
i will definitely try that. i will also try to do some benchmarks with Texture Sharpening on a gf4 to see if i can mirror the some results.
i'm also interested in any change in AF between gf3/4. we'll see if i can't get that stuff included next week too :)
That RT AF script can be also found in the newer versons of Atuner.
And it's a lot easier to use.
It's under the "D3D Extra Options Panel", then click on "Optimise Anisotropic Filtering"
I gained about 20% across the board, with the setting optimised and couldn't see any drop in IQ.
This is a great thread, StealthHawk. Perhaps it can become a sticky ?
Well, I thank you for your compliment, but I don't want to sticky my own thread ;)
Anyway here are the Quake3 screenshots at 4x FSAA and Balanced
Balanced 4x 0xAF
Balanced 4x 2xAF
Balanced 4x 4xAF
Balanced 4x 8xAF
Balanced 4x 0xAF+TS
Balanced 4x 2xAF+TS
Balanced 4x 4xAF+TS
Balanced 4x 8xAF+TS
Now for my commentary. I noticed that sometimes a higher degree of AF was being used than what I specified in the drivers. To ensure accurate results I used RivaTuner to set the AF level and TS options. I also retested Quake2 at the setting of 4x/4x and 4x/2x+TS. The results were the same as before. Although there is a discrepancy in the benchmark results, the screenshots appear to be identical, although 4x/2x+TS is faster than plain 4x/4x.
I think I may test these settings in 3dmark2001 to see if the same strange results occur in D3D.
Yeah, well, since you're a mod and all that
STICKY THE DAMN THREAD SON
And I'm back with some more benchmarks. This time with data from a GeForce4Ti4200 on an AthlonXP 1800+ and nvidia drivers 42.01 winXP DX8.1 installed.
Application: 7273 -13.9%
2x FSAA Balanced: 7017
2x FSAA+ Balanced: 7044
4x FSAA Balanced: 4902
4x FSAA+ Balanced: 4890
4xS Balanced: 4572
4xS+ Balanced: 4565
As I witnessed on my system in the first post, setting the drivers to Application degrades performance by over 1000 3dmarks whether using a gf3 or a gf4 card.
I also tested the FSAA boost in RivaTuner, when the boost is enabled it is denoted by the "+." I was going to use Uttar's AA Analyzer to take screenshots to compare quality, but my friend didn't have DX9 installed, and the program requires it to run. Looking at the benchmark results though, there is no measurable gain from the rumored FSAA optimization, at least in 3dmark2001.
|All times are GMT -5. The time now is 04:22 AM.|
Powered by vBulletin® Version 3.7.1
Copyright ©2000 - 2013, Jelsoft Enterprises Ltd.
Copyright ©1998 - 2013, nV News.