Scalable computer graphics is one of the mantras we have in this industry, it’s our faith, and our goal. Due to the nature of seemingly unlimited parallelism in graphics, the scaling opportunities seem almost equally unlimited. But they are limited, and more than you might think.
In 1999 when Nvidia introduced the NV10, or the GeForce 256, it was the first integrated vertex shader (known as a transform and lighting engine at the time). We could see the road ahead pretty clearly: Moore’s law was kicking in, transistors were becoming free, and the lowly VGA controller grew to become the Graphics Processor Unit, well, almost.
In 2000 Nvidia proved our vision was right and brought out the first chip with pixel shaders, the NV15, and now the GPU had truly arrived.
Moore’s law enabled 3Dlabs, ATI, Nvidia, and Intel to add transistors and processors to the point that today you can get an ATI chip with 48 processors in it; and even an integrated chip, the lowliest of the low end, has two shader processors in it. Why not if transistors are free?
So the GPUs have been scaling nicely, faster actually than Moore’s law, due to the nature of parallelism in graphics. Clearly it was an opportunity waiting to be seized, and seized it has been.
But chip scaling was not the only opportunity, or so we were told. We were reminded of a trick 3dfx developed in 1999 with the Voodoo, they called it Scan Line Interleaving, or SLI, and they weren’t the first to do it. Metabyte-PGC (later part of AlienWare, now part of Dell) had a technique as did ATI with their Alternate Frame Rendering approach on the Rage Fury in 1999.
But in 2004, Nvidia reinvented the idea of scaling with multiple GPUs, and introduced the new SLI, which stands for Scalable Link Interface, which is a clever way to rename the acronym and leverage its branding.
But no sooner did ATI finally get their act together on Crossfire, one year after announcing it, then Nvidia and Dell trumped them with, QUAD SLI. Yes, four AIBs in one, very hot, PC.