Dual Graphics Showdown - SLI vs Crossfire

The two giants of the graphics market are set to go toe to toe with their dual-GPU implementations. James Wang pulls them both apart and explains how each technology works, and which one will deserve your hard earned cash.

CrossFire in all its glory. Classicly styled cable linked Voodoo style. Only digital.

Computer graphics are what is more commonly known as an embarrassingly parallel problem. This is a label given to problems which can be simply split into parts and solved independently without much fuss. Such problems are easily solved by throwing more resources at them. Indeed, adding more pipelines to the graphics processors has been the basis of performance growth in 3D graphics for some time. But what happens when you can’t squeeze more pipelines into the chip, when the limits of physics dictate your limits have been reached? Why, throw more chips at it of course!

The Golden Era
 
The desire to have more than one GPU per system goes back into ancient times. During the antiquity of PC 3D acceleration (circa 1997), gamers were mainly playing at a resolution of 640 x 480. Higher resolutions were out of the question not simply because they would be too slow, but because there wasn’t enough memory on the graphics cards to store the framebuffer. When 3dfx released the 8MB Voodoo 2, 800 x 600 became the new rave. But for some, this still wasn’t enough. Luckily, the Voodoo 2 had just the right feature for them. And with it, an age was born.

Scan-line Interleave (SLI) was a technology invented by 3dfx and Quantum3D to allow multiple Voodoo cards to work in parallel. The fi rst Voodoo used SLI to make inroads into the arcade and simulation market. The Voodoo 2 saw the fi rst use of multi-board graphics for PCs.

Two Voodoo 2 boards could be linked together via a cable. Both cards worked on the identical frame but each only did half the regular work; the first card would render the odd scanlines while the second card renders the even ones. The fi nal output is produced by merging the two frames.

Since each card only needed to store half the scanlines, it saved half of the framebuffer memory. Two Voodoo 2 boards in SLI could therefore render frames in excess of 1024 x 768. Gamers rejoiced. Although each board cost US$300 at the time and a separate 2D graphics card was needed to provide 2D output, many gamers forked over the cash to buy three graphics cards. They were blessed with 3D gaming at 1024 x 768 – not just a true luxury for the time, but also a sign of the future. Never again would we settle for anything less.

The Dark Ages
 
The arrival of AGP pushed multi-board graphics into obscurity. With only one AGP slot per motherboard and PCI graphics cards a dying breed, multi-board systems disappeared.

Two ill-fated products did try to break out of the mould. The first was ATI’s Rage Fury MAXX which used two Rage Pro chips on a single board. This was launched in late 1999 and was handsomely beaten by the GeForce 256. 3dfx made one last attempt with SLI on what turned out to be their last graphics card, the Voodoo 5 5500. Released in mid 2000 and powered by two VSA-100 chips, it faired no better against the competition. NVIDIA’s new GeForce 2 series consistently outpaced the Voodoo 5 and went on to become one of the most successful GPUs ever. After all, who hasn’t fond memories of their GTS or MX? From these two examples one couldn’t be blamed for dismissing multi-GPU graphics as a one hit wonder of the Voodoo era, and relegated to 3D history.

In retrospect, this turned out to be wrong.

The main reason why the Rage Fury MAXX and Voodoo 5 failed is because the chips that they employed were underpowered. Both the Rage Pro and the VSA-100 were produced on relatively old fabrication processes. This gave them a clock disadvantage against NVIDIA’s offering. The MAXX and Voodoo 5 also lacked transform and lighting, which hindered their geometry performance. The last nail in the coffi n was their choice of memory, modestly clocked SDRAM compared to fast DDR memory on the GeForce line-up. In essence, these two cards failed not because the dual-chip concept wasn’t good but rather their chips and memory were too slow. After all, if a single VSA-100 chip isn’t as powerful as half a GeForce 2, how can two of them expect to beat the whole thing? The lesson was this: build the fastest single GPU, then think about linking them together. Which is exactly what NVIDIA did.