PDA

View Full Version : NVIDIA SLI investigation


Pages : [1] 2 3 4

ChrisRay
02-06-05, 01:53 AM
Introduction


Introduction: When Nvidia first announced SLI I was impressed by claimed performance increases. Naturally I was also a bit dubious. Hearing alot of information around the internet I have been reading both negative and positive opinions about the SLI setup. Naturally I have wanted to test an SLI setup myself to better formulate my own opinion. After formulating that opinion I also wanted to share my results with others. So Welcome to my SLI Investigation Thread.


Rendering Methods:There are a few key terms which you will want to know regarding the rendering methods available to an Nvidia SLI setup. The first is SFR. SFR stands for Split Frame Rendering . The method is pretty well described by its name. The other mode available is AFR. Alternatively known as Alternate Frame Rendering which is also pretty well defined by its name. In my experience I have found AFR to be the better of the two. However if one does not seem to be functioning properly I do reccomend trying the other. Nvidia also has specifically optimised AFR and SFR methods within its drivers. I do not reccomend modifying these if they already have a custom built rendering profile built by Nvidia. These can be seen and noted by there bizarre label in the profiles which we will cover shortly.


Profiles: At the very heart of Nvidia SLI lies Nvidia Profile system. Originally introduced last year during the Geforce FX era. Profiles allow the drivers to manually configure settings such as Anti Aliasing, Anisotropic Filtering and now SLI rendering modes. The Nvidia drivers detect the program .exe file and load up the proper settings. To protect users from rendering errors Nvidia has made the SLI profiles non switchable within the control panel. Power Users however may find multiple ways to enable SLI in any application they desire.

#1: SLI Profile Editor released at *********.com (http://www.*********.com/utilities/nvidia/sli/index.php) . This is my preferred method for editing SLI profiles. Its quick. It's fast and it allows easy entry of SLI profiles within the drivers.


http://www.nvnews.net/articles/chrisray_iq/sli/profileeditor.jpg


#2: Manually Editing the nvapps.xml which can be found in the Windows\System folder. A Properly enabled SLI application will look something like this below. I dont reccomend employing this method if you are not at least familiar with markup language. I've illustrated a few examples to give users an idea of what to look for.

<PROFILE Label="Nvidia Vulcan Demo">
<APPLICATION Label="vulcan.exe"/>
<PROPERTY Label="multichip_rendering_mode" Value="1" Itemtype="predefined"/>
<PROPERTY Label="aa_default" Value="10000000"/>

#3: A new application (http://www.nvnews.net/vbulletin/showthread.php?t=45204) has been released by Grestorm. This is tool is not just useful as an SLI profile enabler but an all around package for tweaking graphic settings and performance. It is easiest of the available methods for enabling SLI. Currently the drawback is that it requires Microsoft Net 1.1 Framework (http://www.microsoft.com/downloads/details.aspx?FamilyId=262D25E3-F589-4842-8157-034D1E7CF3A3) to be installed on your computer. You will be able to switch and change between various SLI profiles as well as enable some specific compatibility bits. It will lock out Nvidia specific assigned profiles by default and I dont reccomend editing these unless your absolutely sure you know what your doing. The application itself can be downloaded here. (http://grestorn.webinit.de/yape/nHancer1.0.3.zip) One final note. This is the only application currently available that lets you globally force SLI to be enabled in either AFR, SFR or Single which is excellent for certain applications which have trouble with profiles.

http://www.nvnews.net/articles/chrisray_iq/sli/nhancer.jpg


Selecting the Right Rendering Mode: As described above there are 2 rendering methods for enabling SLI. SFR: 2 And AFR: 1 . Please denote the numbers next to the two rendering modes. You will be using them to enable and switch between the various rendering modes. The SLI Profile Editor released at *********.com (http://www.*********.com/utilities/nvidia/sli/index.php) will allow you to choose between AFR and SFR when setting the profile. Clearly labeled by the number denotion in the application. You may also edit the NvApps.xml file with a proper profile as shown above. The last important thing to note. Number 4 disables SLI and forces you to use a single GPU to render.


Watch Out for Nvidia Specific Profiles: In the NvApps.xml there are specifically optimised AFR and SFR modes for certain applications. These can be seen because they have neither a 1, a 2, or even a 4, in them. An example shall be listed below. Altering these with a non optimised SLI profile will likely prove to be detrimental to your SLI performance.

<PROFILE Label="Tomb Raider">
<APPLICATION Label="TRAOD.exe"/>
<APPLICATION Label="TRAOD_P3.exe"/>
<APPLICATION Label="TRAOD_P4.exe"/>
<PROPERTY Label="prevent_cpl_aa" Value="1" Itemtype="predefined"/>
<PROPERTY Label="multichip_rendering_mode" Value="2C10401" Itemtype="predefined"/>
</PROFILE>

ChrisRay
02-06-05, 01:53 AM
Synthetic and Demo Performance


Test Setup and benching methodology: For all synthetics and demos I used the default rendering method available to ensure valid and comparable results. The old Madonion in me wanting the best performance set the Quality tab to High Performance for benchmarks in Aquamark, and the 3dmark applications. Any other changes will be denoted by the graph illustration.



Athlon 64 3800+ @ 2.4 Ghz
Asus A8N SLI Deluxe
1 Gig DDR400 (Dual Channel) CAS 2.5
Dual 6600GT's

Drivers

Nforce 6.39 Platform Drivers
Nvidia 71.80 Display Driver




3dmark2001SE

http://www.nvnews.net/articles/chrisray_iq/sli/3dmark01.jpg


3dmark2001SE: 3dmark2001SE is remarkably CPU bound platform. It has actually been used for studying CPU performance in many cases. However certain tests within the application still show notable GPU bound tests. These tests are illustrated above. It should be noted that the shader power of an SLI setup is substancially increased to a single GPU. Paticularly last generation.

Nvidia Optimised Profile



3dmark2003

http://www.nvnews.net/articles/chrisray_iq/sli/3dmark03.jpg


3dmark03: 3dmark03 is almost a completely GPU bound software. Showing very little CPU reliance it is an excellent test for pure GPU bound benchmarks. As can be seen by the tests illustrated here. A Second GPU can significantly increase potential of this benchmark. Once again suggesting more shader limited scenerios will greatly benefit from an SLI setup.

Nvidia Optimised Profile


3dmark2005

http://www.nvnews.net/articles/chrisray_iq/sli/3dmark05.jpg


3dmark05: The latest introduction of 3dmark available. This benchmark minimum requirement for all game tests is Shader model 2.0. It is custom optimised to include SM 3.0 profiles for its tests. Once again we can see that Dual GPUS produces a pretty respectable gain for shader limited situations. The demo is probably using an optimised version of AFR if you note its vertex performance.


Nvidia Optimised Profile


Aquamark

http://www.nvnews.net/articles/chrisray_iq/sli/aquascore.jpg


Aquamark3: Aquamark3 shows what we have seen in previous synthetic tests. A second GPU will produce pretty significant gains for the the application. An interesting note is the CPU score. It would appear SLI creates a small CPU overhead CPU limited situation. This will probably be observed in gaming enviroments when you are limited by your CPU rather than your graphic card.

Alternative Frame Rendering Profile



Nvidia and ATI tech demos.

http://www.nvnews.net/articles/chrisray_iq/sli/demobench.jpg


Tech Demos: Having real benefits in synthetics I wanted to see if I could reproduce the performance in other shader tech demos released by ATI and Nvidia. The SLI profiles for these demos have been forced by me using the applications listed in the first post. The tech demos reproduce what we have seen in the synthetic benchmarks. SLI greatly benefits shader limited situations.

Profiles listed in graph

ChrisRay
02-06-05, 01:54 AM
Gaming Performance


Test Setup and benching methodology: My benchmarking methods are typical to prior benchmarks done by me. I use fraps and record 3 preset runs through a standard level in each game illustrated. I then divide the results by 3 to reduce margin of error. Margin of Error is usually with 1 FPS and I take great pride in the accuracy of my benches.



Athlon 64 3800+ @ 2.4 Ghz
Asus A8N SLI Deluxe
1 Gig DDR400 (Dual Channel) CAS 2.5
Dual 6600GT's

Drivers

Nforce 6.39 Platform Drivers
Nvidia 71.80 Display Driver





Far Cry


http://www.nvnews.net/articles/chrisray_iq/sli/farvolcano.jpg

http://www.nvnews.net/articles/chrisray_iq/sli/farchive.jpg


Far Cry: The performance benefits to SLI will just jump out and strike you. The performance increases took settings that were unplayable on a single GPU to levels that were smooth and offered an enjoyable playback experience.


Nvidia optimised Profile


Doom 3


http://www.nvnews.net/articles/chrisray_iq/sli/doom3hell.jpg

http://www.nvnews.net/articles/chrisray_iq/sli/doom3under.jpg


Doom 3: Unless you have been living under a rock without an internet connection. You probably know that Doom 3 has been one of the most demanding games released this last year. Bringing last years graphic cards to there knees. It is also optimised to take advantage of Nvidia hardware. One thing that should be noted. Disabling AA brings about substancial performance gains. This is probably due to the way anti aliasing effects two sided stencil acceleration and the 128 framebuffer limitation of 2 6600GTs.


Alternate Frame Rendering Profile



Everquest 2


http://www.nvnews.net/articles/chrisray_iq/sli/EQ2bench.jpg

Everquest 2: Quite possibly the most demanding MMORPG ever released. This game is known for bringing the most high end computers to their knees. Adding a second GPU does seem to benefit the gameplay experience. But the platform still remains remarkably CPU bound and minimum framerates dont seem to be affected.

Note: Everquest 2 is currently only SLI compatible with 71.80 drivers. It has a custom profile for them. Forcing SLI with earlier drivers will cause rendering errors and crashes with EQ 2. T



Nvidia optimised Profile

Call of Duty


http://www.nvnews.net/articles/chrisray_iq/sli/dutybench.jpg


Call of Duty: Call of Duty starts out CPU bound even on a single GPU platform. As expected from the earlier aquamark results. When you are CPU bound SLI creates a slight CPU overhead which is slightly detrimental to performance. Once the resolution and anti aliasing has been increased the benefits of SLI become readily obvious however. Call of Duty is not by default SLI enabled. I had to force AFR mode with application listed in the first post.



Alternate Frame Rendering Forced Profile



Half Life 2


http://www.nvnews.net/articles/chrisray_iq/sli/hl2bench.jpg


Half Life 2: Half Life 2 also started out CPU bound. As with other CPU bound situations it proved to create a slight CPU overhead. As with Call of Duty enabling higher levels of resolution and anti aliasing started to show the benefits of SLI however.


Nvidia optimised Profile



Conclusion

Conclusion: SLI has a lot of potential. At the moment it can be seen in just about every GPU limited enviroment. CPU limited enviroments will probably not benefit much from it however. Considering all the talk about Nvidia profiles I expected limited application support with SLI. However thanks to enthusiast users and the ability to edit the profiles. I havent found any real problem enabling SLI in any of my favorite games that do not already support it. I do hope Nvidia listens and adds an "Advanced" tab that allows users to switch SLI on and off via the control panel. The current method isnt what I'd call the perfect solution. Fortunately issues such as this can be fixed with driver updates. Given the the need for the profiles I would hope Nvidia steps up and increases there regular driver updates to meet the demands of the new SLI profiles they have developed.

ChrisRay
02-06-05, 01:54 AM
Shew, That was alot of work ;)

Clay
02-06-05, 01:57 AM
Wow Chris! You've been busy! Looks like a lot to pour over here. :thumbsup:

MUYA
02-06-05, 01:59 AM
SLI rawks :)

rohit
02-06-05, 03:49 AM
This was more like 66GT SLi review.
Great work !:thumbsup:
SLi Truly Rocks!:thumbsup:

bkswaney
02-06-05, 04:12 AM
Super nice!
Thanx for all that work. :D

zoomy942
02-06-05, 04:25 AM
extremely informative. nice job

Rollo
02-06-05, 12:09 PM
I would think this would be extremely useful to anyone considering a rig with 6600GT SLI in it. Thanks for all the hard work!
:clap:

superklye
02-06-05, 01:25 PM
This is a fantastic archive of information. Great job, Chris. It's very interesting, very informative and very well written.

jAkUp
02-06-05, 01:37 PM
Awesome! Nice Chris! See, I told you that you have alot of quality posts :)

Sazar
02-06-05, 01:49 PM
nice stuff chris :thumbsup:

did you notice the max fps on the games shoot up as well by a lot?

OWA
02-06-05, 02:08 PM
Great post Chris. Thanks for the effort.

PSYCHODAD
02-06-05, 02:33 PM
Wow, nice job. Bet that took awhile to put together.

SH64
02-06-05, 03:56 PM
Very good work Chris ! .. its good to see someone finally managed to scoop on the Min FPS results in SLI mode .

ChrisRay
02-06-05, 08:37 PM
nice stuff chris :thumbsup:

did you notice the max fps on the games shoot up as well by a lot?

Yes Max FPS also went up. But I dont pay much heed to max FPS. Just AVG/Min usually because I feel its most important.

ChrisRay
02-08-05, 05:14 AM
Attention UT2004 users. I managed to get around the performance barrier of UT2004 SLI. By default Nvidia has it set to an optimised AFR profile. When you'd enable SLI you'd see zero performance or worse. Lose it.. Obviously I was completely unsatisfied with this result. So I toyed around with it. And Viola. The most basic Split Frame Rendering mode produces a nice performance enhancement for UT2004.


*Edit* Removed as I have found problems with this game.

Veger
02-08-05, 08:39 AM
wow, excellent information, good job chris...

blueworm
02-08-05, 09:05 AM
Great work cris.
1 question, I take it that it is possible to hand write the profiles? And that the cutesy app is just for ease?
I am a linux user and we are still waiting for SLI drivers, and the whole profiles thing has had me worried.

Rollo
02-08-05, 09:09 AM
Attention UT2004 users. I managed to get around the performance barrier of UT2004 SLI. By default Nvidia has it set to an optimised AFR profile. When you'd enable SLI you'd see zero performance or worse. Lose it.. Obviously I was completely unsatisfied with this result. So I toyed around with it. And Viola. The most basic Split Frame Rendering mode produces a nice performance enhancement for UT2004.


http://members.cox.net/omega1979/SLI/ut2004sli.png


Excellent work Chris. Now if you can get it working for Doom1, I'll be set! ;)

jbirney
02-08-05, 10:43 AM
Good info in there Chris!!!

eva2000
02-08-05, 12:31 PM
may interest some folks

Just reinstalled FarCry and ran some benchmarks at stock cpu, card speeds in SLI mode with 71.25 drivers

System:
3500+ 90nm @ 11x201 = 2211mhz
A8N-SLI Deluxe 1004.01 bios
2x 512MB OCZ PC3700EB singes @201mhz 3-2-2-8 1T
2x eVGA 6800U PCI-E @ stock 425/1100
1x 120GB WD 1200JB
580W Topower 6P686 PSU

71.25 drivers
4xAA/16xAF


HOC Volcano (http://fileshosts.com/AMD/Asus/A8NE_Deluxe/results/3500_90nm/OCZ/PC3700EB/2x512/LDT4x/201-201-3228-9-14-2223-1T-1.4-3.55vpsu-1.67/farcry/HOC_Volcano/4xAA_16xAF_425-1100_7125/result.html)

Resolution: 1024×768
Maximum quality option, Direct3D renderer
Level: Volcano, demo: hocvolcano.tmd
Pixel shader: default model
Antialising: 4×
Anisotrophic filtering: 16×
HDR: disabled
Geometry Instancing: disabled
Normal-maps compression: disabled

Score = 107.78 FPS (Run 1)
Score = 106.02 FPS (Run 2)
Score = 108.39 FPS (Run 3)
Score = 107.23 FPS (Run 4)
Average score = 107.35 FPS


Resolution: 1280×1024
Maximum quality option, Direct3D renderer
Level: Volcano, demo: hocvolcano.tmd
Pixel shader: default model
Antialising: 4×
Anisotrophic filtering: 16×
HDR: disabled
Geometry Instancing: disabled
Normal-maps compression: disabled

Score = 94.67 FPS (Run 1)
Score = 91.52 FPS (Run 2)
Score = 95.16 FPS (Run 3)
Score = 94.08 FPS (Run 4)
Average score = 93.85 FPS


Resolution: 1600×1200
Maximum quality option, Direct3D renderer
Level: Volcano, demo: hocvolcano.tmd
Pixel shader: default model
Antialising: 4×
Anisotrophic filtering: 16×
HDR: disabled
Geometry Instancing: disabled
Normal-maps compression: disabled

Score = 72.14 FPS (Run 1)
Score = 70.60 FPS (Run 2)
Score = 73.46 FPS (Run 3)
Score = 72.95 FPS (Run 4)
Average score = 72.28 FPS


HOC Archive (http://fileshosts.com/AMD/Asus/A8NE_Deluxe/results/3500_90nm/OCZ/PC3700EB/2x512/LDT4x/201-201-3228-9-14-2223-1T-1.4-3.55vpsu-1.67/farcry/HOC_Archive/4xAA_16xAF_425-1100_7125/result1.html)

Resolution: 1024×768
Maximum quality option, Direct3D renderer
Level: Archive, demo: Archive1.tmd
Pixel shader: default model
Antialising: 4×
Anisotrophic filtering: 16×
HDR: disabled
Geometry Instancing: disabled
Normal-maps compression: disabled

Score = 98.28 FPS (Run 1)
Score = 110.53 FPS (Run 2)
Score = 113.70 FPS (Run 3)
Score = 113.84 FPS (Run 4)
Average score = 109.08 FPS


Resolution: 1280×1024
Maximum quality option, Direct3D renderer
Level: Archive, demo: Archive1.tmd
Pixel shader: default model
Antialising: 4×
Anisotrophic filtering: 16×
HDR: disabled
Geometry Instancing: disabled
Normal-maps compression: disabled

Score = 95.73 FPS (Run 1)
Score = 109.25 FPS (Run 2)
Score = 108.92 FPS (Run 3)
Score = 108.23 FPS (Run 4)
Average score = 105.53 FPS


Resolution: 1600×1200
Maximum quality option, Direct3D renderer
Level: Archive, demo: Archive1.tmd
Pixel shader: default model
Antialising: 4×
Anisotrophic filtering: 16×
HDR: disabled
Geometry Instancing: disabled
Normal-maps compression: disabled

Score = 78.09 FPS (Run 1)
Score = 82.36 FPS (Run 2)
Score = 83.01 FPS (Run 3)
Score = 82.20 FPS (Run 4)
Average score = 81.41 FPS


TechReport Regulator (http://fileshosts.com/AMD/Asus/A8NE_Deluxe/results/3500_90nm/OCZ/PC3700EB/2x512/LDT4x/201-201-3228-9-14-2223-1T-1.4-3.55vpsu-1.67/farcry/TechReport_Regulator/4xAA_16xAF_425-1100_7125/result1.html)

Resolution: 1024×768
Maximum quality option, Direct3D renderer
Level: Regulator, demo: tr1-regulator.tmd
Pixel shader: default model
Antialising: 4×
Anisotrophic filtering: 16×
HDR: disabled
Geometry Instancing: disabled
Normal-maps compression: disabled

Score = 54.05 FPS (Run 1)
Score = 62.03 FPS (Run 2)
Score = 62.10 FPS (Run 3)
Score = 61.91 FPS (Run 4)
Average score = 60.02 FPS


Resolution: 1280×1024
Maximum quality option, Direct3D renderer
Level: Regulator, demo: tr1-regulator.tmd
Pixel shader: default model
Antialising: 4×
Anisotrophic filtering: 16×
HDR: disabled
Geometry Instancing: disabled
Normal-maps compression: disabled

Score = 52.85 FPS (Run 1)
Score = 60.45 FPS (Run 2)
Score = 60.29 FPS (Run 3)
Score = 60.05 FPS (Run 4)
Average score = 58.40 FPS


Resolution: 1600×1200
Maximum quality option, Direct3D renderer
Level: Regulator, demo: tr1-regulator.tmd
Pixel shader: default model
Antialising: 4×
Anisotrophic filtering: 16×
HDR: disabled
Geometry Instancing: disabled
Normal-maps compression: disabled

Score = 47.95 FPS (Run 1)
Score = 54.38 FPS (Run 2)
Score = 54.39 FPS (Run 3)
Score = 54.42 FPS (Run 4)
Average score = 52.78 FPS

Blacklash
02-08-05, 01:03 PM
Great job Chris, as usual. You are often, IMO, the board MVP.

eva2000's frames at 1600x1200 with 4xAA/16xAF illustrate precisely why I want SLI. The SLI min/average frames with GTs or Ultras are just too impressive.

Soon as I break down and get that new monitor it's all over for me :D I'll have zero resistance left.

ViN86
02-08-05, 01:20 PM
wow, great post and info :D

(alf) (alf) (alf)