: News
: Reviews
: Editorials

: Archives
: About
: Contact
: Advertising
: Privacy
: Links
: Forums






Not Cheap... Not Slow


by Josh Walrath


            I am sure quite a few of you still remember the 3dfx Voodoo 2 days quite fondly.  I know that I do.  When the Voodoo 2 was announced, 3dfx made it known quite clearly that two of them could be put together to double the performance, as well as enable resolutions up to 1024x768 at a full 60 frames per second in the popular games of the time.  Quite a few people bought two of these monsters, and it was the highest end combination that could be had for over a year (the NVIDIA Riva TnT was not the Voodoo 2 SLI killer that was initially promised).  It wasnt until the Voodoo 3 and the TnT 2 were released that the Voodoo 2 SLI drifted into obscurity.

Behind the beige exterior lies one of the most powerful 3D rendering machines on the planet.

            The years passed, and NVIDIA bought up 3dfx.  Along with 3dfx came many patents, trademarks, and personnel.  While there has not been a wholesale implementation of 3dfx technology in NVIDIA products, some of the ideas that 3dfx pioneered found a place to live.  NVIDIA chips have had the ability to be linked together since the GeForce 4 Ti days, and quite a few NVIDIA based chips have been utilized in the simulator market in multi-GPU applications.  One of the things that held NVIDIA back from producing a consumer level multi-GPU product has been that of the underlying architecture of the PC.  First off AGP 2X and 4X did not allow for a second AGP connector, and the PCI bus was just too slow for modern GPUs.  While the AGP 3.0 specification would have allowed for a second AGP 8X slot, it was never implemented by any chipset manufacturer.  Not only that, but other bottlenecks such as limitations in AGP upload bandwidth, and overall memory system bandwidth made the concept of a modern multi-GPU setup unappealing from a performance perspective.

            Things started to look a lot brighter once DDR 400 and dual channel memory architectures started to pop up.  Main memory bandwidth was starting to become a non-issue, and the performance of the latest AMD and Intel processors made CPU overhead considerations of GPU arbitration a thing of the past.  The biggest technology jump that made a consumer level multi-GPU product viable was that of PCI-Express.  Here was a interconnect technology that promised a huge amount of upstream and downstream bandwidth and was not hindered by the architectural limitations of previous standards.  NVIDIA knew this technology was coming, and they took a very long sighted approach towards developing their consumer level multi-GPU technology.  So, deep inside NVIDIAs engineering labs, SLI was born again.

            Scan Line Interlace transformed into Scalable Link Interface.  3dfxs reliance on analog scanlines, and later on with the Voodoo 5s digital scanlines, proved to be more of a hassle than NVIDIA wanted to tackle.  So, instead of scanlines NVIDIA took two different approaches to rendering a scene.

Once the panel is removed we see the two 6800 GT's just itching to go.  The PC Power and Cooling 510 SLI has nicely wrapped cables that tend to not get in the way.  Noise levels can be significantly elevated by the extreme amount of fans a well cooled SLI case should boast.  This particular case has 9 fans in total.

            The first is Alternate Frame Rendering (AFR).  This simply sends each card a separate frame, and the master card then takes the output from the secondary card and alternates that frame from its own generated frame.  This is the rendering format we will see the majority of games utilize.  There are several upsides to this type of rendering.  The first is the simplicity of the setup and a large compatibility with most titles.  It is quite easy for the driver to send each card alternating frame data, and since each frame is rendered in full by each card, it is relatively easy for the master card to display the finished frame from the secondary card after it was transmitted.  The next big advantage is that of geometry scaling.  Each card is sent the full geometry for the scene it is working on, so at any one time each card is processing a different frames worth of geometry.  As compared to the other rendering scheme to be described, where each card is sent the exact same geometry data, we can theoretically see a doubling of overall geometry output.  AFR (and AFR2 for applications which might have issues with AFR) will probably be the dominant way to run applications in SLI.  Currently this is used for most of the benchmark applications such as 3D Mark 03/05, Codecreatures, and others that would benefit from the extra geometry pushing power that AFR gives.


Next: SFR


If you have found this article interesting or a great help, please donate to this site.


Copyright 1999-2005 PenStar Systems, LLC.