SLI and PGC
How does everyone else do it?
Using two graphics engines at the same time isn't new to the world of 3D acceleration. Many of you are already familiar with 3dfx's SLI (Scan Line Interleaving) technique first introduced with the original Voodoo accelerator. Although the first Monster 3D Voodoo1 cards couldn't run in a dual SLI configuration, Quantum3D had an Obsidian board that used multiple Voodoo1 accelerators.
SLI didn't become a huge feature until the introduction of the Voodoo2. Anyone remember 3dfx's "Power of Two" advertising campaign? Users could hook up two Voodoo2 cards in SLI configuration to almost double 3D performance. With two separate 3D accelerators, one would render all the even horizontal scan lines, and the other would render the odd lines.
Overhead and synchronization issues kept the 2x performance increase claims in check, but SLI still performed very well. The Voodoo2 SLI setup can still hold its own against the Voodoo3 as we saw in our Voodoo3 3000 PCI Review benchmarks.
AGP problems
3dfx had a good thing going with SLI, but then the AGP slot appeared. Splitting up the SLI work between two PCI cards was pretty even, but the new AGP slot made SLI implementation almost impossible according to 3dfx. Finding a way to split up the work between an AGP card and a PCI card would be difficult. The AGP card would end up waiting for the PCI card most of the time because of the evenly shared workload between the cards.
Back in February, Metabyte announced their multiple accelerator solution, PGC (Parallel Graphics Configuration). We took a trip to their offices, and took a look-see for ourselves. You can read the write-up about our visit to Metabyte right here.
Metabyte's PGC solution involved splitting up the rendering workload a different way. Instead of having each card rendering every other line, the PGC technology would have one card do the top half of the screen and the other card render the bottom. The division between top and bottom could be adjusted to give the AGP card a larger portion of the screen to render, making sure that the AGP card never has to wait for the PCI card. Therefore the AGP card could do 60 or 70% of the screen while the PCI card could do 40 or 30%, and both would finish at the same time.