It is not always an advantage to try to be a man of your word, yet a promise is a promise. Days before the G80 launch while chatting in the B3D IRC channel I told Rys that if I manage to buy a 8800GTX before the start of December Id owe him a public apology. G80 was officially announced on November 8 and I was holding a Gainward Bliss 8800GTX in my hands on the 14th. Thus, I have no other choice than to send him my apologies. Dont worry though this one is amongst the cases where I simply love being wrong. Enough with the nonsensical drivel and on topic: Chris wrote an excellent analysis over the image quality on a G80 and promised to deliver higher resolution results in due time. Sometimes things dont pan out as we expect and here I am trying to fill in the gap. When it comes to as fundamental changes as with D3D10 it is always hard at first to find adequate material (yes even synthetic applications) to test hardware like the 8800GTX. Unless of course someone has advanced coding skills, which I completely lack in. If you havent read yet Beyond3Ds outstanding architecture analysis on the G80, Id urge you to do so. Furthermore I tried to find testing scenarios that have not been covered this far, which was not that easy either. My system setup hasnt changed much apart from the GPU, so those that are still using a single core CPU as me and arent bound to low resolutions might find a bit more interest in the pages to follow. Here the system setup in detail again:
Tagan TG580-U22 (580W) Thermaltake Xaser III V1000A ASUS A8N-SLi Deluxe Athlon FX-53 Thermaltake Sonic Tower 1536MB PC3200 DDR SDRAM Gainward Bliss GeForce 8800GTX 768MB Creative Soundblaster Audigy2 Samsung SyncMaster 1100DF WinXP Pro SP2 DirectX 9.0c nForce 6.70 ForceWare 97.02
First, I wanted something that represents in relative terms a worst-case scenario when it comes to anisotropic filtering performance. One of my worst-case scenarios is Humus selective Super sampling demo, which of course serves a very different purpose. Here is a screenshot of the demo:
Its purpose of course is to illustrate what selective Super sampling could do to shader aliasing both in terms of performance as in image quality. The trick being that w/o any selective SSAA enabled and some or all available filtering optimizations present, shader aliasing gets more prevalent. The usual performance drop with past generation GPUs went up to 50%, so lets see what the G80 is capable of.
No shader AA quality mode: ~1300fps No shader AA high quality mode: ~1070fps 4x selective SSAA quality mode: ~765fps 4x selective SSAA high quality mode: ~655fps
In the first case, we see a drop of 18% without selective AA - which is quite impressive considering the fore mentioned 50%- and roughly 14% with 4x sample selective AA enabled. Chris measured in a real time gaming scenario like Serious Sam2 a drop of roughly 3%, yet you should remember that I set out for a worst-case scenario. The bottom line of the story is that its hardly worth bothering with the quality mode at all, considering the rather miniscule performance drop with high quality AF enabled. We had seen some fill rate test in Archmark after the 8800GTX ended up in some users hands and some were quick to suggest, free trilinear. I am not entirely sure if Archmarks results are representative, but lets have a second look on them:

High quality 1xAF:

Texturing Mode R5G6B5A0 Z0 S0 --Textured fillrate--------------------------------Bilinear filter----------------------------1 10.386 GPix/s 2 7.139 GPix/s 3 5.603 GPix/s 4 4.546 GPix/s ----Trilinear filter---------------------------1 9.567 GPix/s 2 7.179 GPix/s 3 5.633 GPix/s 4 4.570 GPix/s

Quality 1xAF:

Texturing Mode R5G6B5A0 Z0 S0 --Textured fillrate--------------------------------Bilinear filter----------------------------1 10.378 GPix/s 2 7.137 GPix/s 3 5.604 GPix/s 4 4.545 GPix/s ----Trilinear filter---------------------------1 9.618 GPix/s 2 7.181 GPix/s 3 5.631 GPix/s 4 4.571 GPix/s
The results here are nearly identical between the modes, but lets try the same this time with 16xAF enabled:

High Quality 16xAF:

Texturing Mode R5G6B5A0 Z0 S0 --Textured fillrate--------------------------------Bilinear filter----------------------------1 10.376 GPix/s 2 7.122 GPix/s 3 5.603 GPix/s 4 4.545 GPix/s ----Trilinear filter---------------------------1 8.521 GPix/s 2 4.579 GPix/s 3 3.054 GPix/s 4 2.292 GPix/s

Quality 16xAF:

Texturing Mode R5G6B5A0 Z0 S0 --Textured fillrate--------------------------------Bilinear filter----------------------------1 10.333 GPix/s 2 7.126 GPix/s 3 5.603 GPix/s 4 4.545 GPix/s ----Trilinear filter---------------------------1 9.401 GPix/s 2 5.590 GPix/s 3 3.872 GPix/s 4 2.753 GPix/s
If I assume that the application is not measuring something wrong, once 16xAF is enabled, there is a cost with trilinear and obviously a smaller one in quality mode, since the latter uses a trilinear optimization often referred to also as brilinear. The gain due to the trilinear optimization accounts in that case 16%.

ForceWare 78.03 beta

A fix for the texture shimmering bug?
Shortly after the launch of the 7800GTX, various complaints from the user base appeared about texture shimmering with anisotropic filtering enabled. First covered briefly by hardware.frs Damien Triolet in this review, followed later on by the German 3DCenters relevant article and Hexus.net with their own detailed description of the issue. Apparently it gets more noticeable if applications request a high negative LOD bias and if any under filtering related theories should have any merit, then it truly can be a quite annoying side-effect. Ive more a condescending approach to things and since this byproduct affects merely one specific zone of each frame, it also can be seen as a bug. NVIDIA response was immediate and not too long after the fore mentioned articles were published released the ForceWare 78.03 beta driver, which supposedly improves texture shimmering on the G70 class of accelerators. Ive tested the driver in more than a handful of games and theres truly a significant improvement in high quality mode (all optimizations disabled) and while theres still a very small amount of shimmering visible, I had to almost press my nose on the screen to actually notice anything in most occasions. Quality (optimizations enabled) on the other hand seems only very slightly improved and theres still a considerable amount of artifacts. Note that shimmering doesnt occur under all occasions or all games. Mostly games that request for a high negative LOD bias are affected and then it still comes down to the type of textures that are being used in an application for the side-effect to be even noticeable. System Setup: Tagan TG580-U22 (580W) ASUS A8N SLI Deluxe AMD Athlon FX-53 1024MB PC3200 DDR SDRAM GeForce 7800GTX Soundblaster Audigy2 Samsung SyncMaster 1100DF Games tested: Half Life2 (3DCenters 3dc_canals) Far Cry (Ubisofts research, regulator, training and techreports tr1volcano) UT2004 (UMark and 3DCenters Primeval) Chronicles of Riddick (Panoptical02, Sodini02) Battlefield2 (Guru3Ds guru3d) All tests conducted with 4xAA/16xAF enabled.

HALF LIFE 2

For HL2 I had also Transparency supersampling enabled, because the game is overloaded with alpha test textures. Thus for this game transparency AA is an absolute blessing. Please note that quality stands actually for slightly more aggressive optimizations since in that mode the anisotropic MIP filter optimization is actually disabled. I kept it enabled to detect an as high possible ballpark between enabled and disabled optimizations.

105.97

1280*1024

102.02

1600*1200
90.78 88.32 Quality HighQuality

2048*1536

The difference accounts across resolutions between high quality and quality as follows: 1280*1024 = -4% 1600*1200 = -3% 2048*1536 = -8% The difference in performance between the two modes and the framerate so high, that its hardly worth it enabling any optimization at all. Default quality (without the anisotropic MIP optimization) modes output is fairly similar to other products Ive seen in HL2. Slight shimmering here and there on floor textures, some moir patterns when you look at buildings for instance from specific angles etc. On the other hand with high quality most if not all fore mentioned side-effects are gone, and I wouldnt call 57 fps in such a ludicrous resolution like 2048 with 4xMSAA + TSAA and 16xAF unplayable at all; au contraire.

UNREAL TOURNAMENT 2004

UT2004 is one of shimmering poster childs and that because the application requests for some textures a negative LOD bias value up to -4.0 (when set to ultra high for textures in game). With quality (whether with the anisotropic MIP optimization enabled or disabled) there are a couple of spots in the game (like some floor textures in DM-Asbestos) where the shimmering is still there and immediately noticeable. With high quality theres only a slight degree of shimmering still present, but still to a higher degree than other applications (that dont request for an as high negative LOD value). To test the differences I used UMark2.0; first comes a pure flyby to test fill-rate differences from the CTF-Face3 map:

228.19

207.98 Quality High Quality

207.52

174.63
Differences in percentages: 1280*1024 = -9% 1600*1200 = -16% Of course the most demanding performance-wise maps are the ONS and Assault maps in UT and the case above could be only an indication for the DeathMatch or Capture The Flag maps and only if you play online and your PC isnt acting as a server.
Next I tried a CPU bound botmatch case scenario and 3DCenters Primeval demo seemed ideal for that task, with 12 bots and heavy firefights. In comparison I also tried a botmatch scenario in UMark with 4 bots only (again in ONS-Primeval). Resolution for both cases was 1600*1200.

Primeval (12)

64.51 56.25

Quality Primeval (4)

84.49 58.100

High Quality

While the difference in 3DCenters timedemo is at 13%, the test in UMark showed a 30% difference. Quality with the anisotropic MIP optimization disabled is 82.48 and thus still a 29% difference to high quality. I should note though that Primeval as a map represents the absolute worst case scenario for the entire game. While it clearly shows a worst case scenario it doesnt reflect the performance of the entire game in my opinion. A couple of further tests with UMark, again with only 4 bots:

ONS-Dawn

105.49 80.24

CTFMoondragon

163.71 134.84 103.99 86.05

Quality High Quality

ONS-Torlan
Differences in percentages: ONS-Dawn = - 24% CTF-Moondragon = -18% ONS-Torlan = -17% The picture now is a bit clearer; the performance difference is in between 15-25% on average and is by far not as small as in HL2 for instance. But UT2004 is still so screamingly fast on a G70 with a high end CPU that the game is also nowhere near unplayable, especially when your absolute worst case scenario is still over the 55 fps mark.

BATTLEFIELD 2

Battlefield2 is another poster child when it comes to texture shimmering but nowhere near as bad as Battlefield 1942. Frankly Ive played the game on a couple of different systems and I wouldnt touch that one with any optimizations enabled on any GPU right now available. Its obvious that BF2 badly needs the long awaited patch in order to reduce the rather huge memory resources this game currently needs. Else: I expect to see further performance improvements after the patch and hopefully for all accelerators. I couldnt see any shimmering with high quality enabled, so lets see what performance looks like:

59.08 1600*1200 52.52

Quality

34.78 2048*1536 30.64

Differences in percentages: 1600*1200 = -11% 2048*1536 = -12% In 2048 the game isnt playable for my taste, yet it definitely is in 1600*1200.

CHRONICLES OF RIDDICK

Finally a game that captured me in my office chair for hours. If theres going to be a sequel of that one, Im definitely looking forward for it. The 1.1 patch contains four inbuilt timedemos, from which I used two to see what the exact performance penalty is for high quality texture filtering mode. Colour me blind but I couldnt find any difference between quality and high quality in this one. Panoptical02:
60.46 1280*1024 58.94 Quality 43.78 1600*1200 42.69 High Quality
More or less a 3% difference in both resolutions; which is in the same ballpark in the other demo too. Sodini02:

73.88 1280*1024 71.54

53.32 1600*1200 51.70 80

FAR CRY

Far Cry used to make quite a difference between quality and high quality on NV40. I measured in my 6800 write-up here an average difference of ~27% and that with only 4x sample AF. I used HardwareOCs benchmarking utility which contains the timedemos used in that write-up above, yet used instead of the original, techreports custom tr1volcano demo. I didnt see any differences here either, in terms of image quality between high quality and quality and I checked thoroughly the spots I recalled exposing some minor sideeffects in the past. However should I have missed a spot, feel free to point me to it and Ill have a look at it, since its virtually impossible to replay entire games from start to end for a venture like this.

Training 1280*1024

74.83 69.2 60.61 54.04 68.65 62.78

Training 1600*1200

Regulator 1280*1024

Regulator 1600*1200

52.53 46.93 100.06 85.78 73.79 63.8

Research 1280*1024

Research 1600*1200
Difference in percentages: Training: 1280*1024 = -8% 1600*1200 = -11% Regulator: 1280*1024 = -9% 1600*1200 = -11% Research: 1280*1024 = -14% 1600*1200 = -14% Albeit I wouldnt really compare 4xAF with 16xAF results (NV40 vs. G70), if Id ignore the difference in the amount of AF samples, the performance drop is half as low on the G70. Techreports tr1volcano demo:

102.13 94.98 72.58 68.61 44.2 41.8
Percentages: 1280*1024 = -7% 1600*1200 = -5% 2048*1536 = -5% I dont understand myself why the drop is bigger in 1280, but the difference between the two modes here is smaller than in the other demos. I was merely curious to see how it fairs in 2048 and HardwareOCs benchmark utility doesnt support resolutions beyond 1600.

TEMPORARY CONCLUSION

Temporary because this is not only a beta driver, but I expect the community to keep an eye on issues like this in the future, probably to a higher degree than before; now that as many have become aware of it. Whatever the source or purpose of the indicated problem is, for the time being I can say that this driver delivers a definite improvement (mostly in high quality mode) to AF quality over former drivers. For a temporary fix this driver does a fine job and heres to hope that future official driver releases will increase image quality even more without any noticeable differences in performance. I will of course revisit the topic myself in the future as of course do I plan to measure performance and IQ with upcoming long expected triple-A games. Id like to thank all the individuals that helped me compiling this little write-up. Stay tuned -AiL