NVIDIA Linux SLI

With our previous article that we published moments ago, demonstrating the performance of the GeForce 7800GTX 256MB under Linux with the 1.0-8174 Rel80 drivers that were finally released today, there's no disputing that the Windows XP NVIDIA ForceWare users can generally see a significantly higher frame-rate with the same hardware components, in addition to other features that aren't yet supported by the proprietary NVIDIA Linux drivers. However, how do NVIDIA's initial Rel80 Linux drivers (1.0-8174) fair in the world of Scalable Link Interface? Today we will be investigating all of these areas of SLI as we measure the level of performance on this Athlon 64 system with Enemy Territory v2.60, Quake 4 v1.0.5, and Doom 3 v1.3.1302. To start with, below is the system setup used during the testing for this article. The basis for this system is Tyan's K8E-SLI, which we recently reviewed here, and it has proved to be an exceptional desktop and workstation motherboard and is based off of the nForce Professional 2200 Chipset rather than the nForce4 SLI.

As mentioned previously, all of the game testing today had occurred with Enemy Territory, Quake 4, and Doom 3, with the testing for this article being a single 6600GT against a 6600GT SLI comparison. We will also be delivering additional SLI results from other GeForce 6/7 cards soon. The various settings in which we benchmarked Enemy Territory are standard settings while Doom 3 focused upon High Quality, High Quality - 2x Quincunx AA/4x AF, High Quality - 4x AA, 9-tap Gaussian/8x AF, and High Quality - 8x AA/8x AF. Quake 4 was benchmarked at Low Quality, High Quality, High Quality - 2x Bilinear AA/2x AF, High Quality - 4x Bilinear AA/4x AF, and High Quality - 8x AA/8x AF. We originally intended on delivering performance results at multiple resolutions, however, one of the problematic areas we have found thus far in our initial testing has been no SLI performance gains unless the game resolution is that of the X.Org server. In the case of this article, the testing was performed at 1280 x 1024. (EDIT [2005-12-06]: NVIDIA has told us the lack of SLI scaling is caused by a problem in XVidMode, that they have yet to correct, but you should see improved performance if using the XRandR extension to change resolutions.) There are also a few additional bugs with the SLI portion of these drivers that have yet to be worked out. On the following pages are our 1.0-8174 results under these three games using both single and dual graphics cards. If you would like more information about Linux SLI in regards to the hardware and software setup, it can be viewed in our Linux SLI Primer. During our Linux SLI testing, we ran the 6600GT pair in Alternate Frame Rendering (AFR) as well as Split Frame Rendering (SFR) and SLI Antialiasing (SLIAA). In Alternate Frame Rendering, one GPU draws the next frame while the other matched GPU is processing the frame after that while in Split Frame Rendering the graphics load is simply split between the two contenders. In the final mode, SLIAA, the Antialiasing load is split between the two GPUs.