Voodoo 5 - Architecture and Performance

Architecture and Performance

The VSA-100 graphics chip is a direct descendant of "Avenger", more commonly known as Voodoo3. It was built on a 250 nm semiconductor manufacturing process, as with Voodoo3. However, the process was tweaked with a 6th metal layer to allow for better density and speed, and the transistors have a slightly shorter gate length and thinner gate oxide. VSA-100 has a transistor count of roughly 14 million, compared to Voodoo3's ~8 million. The chip has a larger texture cache than its predecessors and the data paths are 32-bits wide rather than 16-bit. Rendering calculations are 40-bits wide in VSA-100 but the operands and results are stored as 32-bit.

One of the design goals for the VSA-100 was scalability. The name of the chip is an abbreviation for "Voodoo Scalable Architecture." By using one or more VSA-100 chips on a board, the various market segments for graphics cards are satisfied with just a single graphics chip design. Theoretically, anywhere from 1 to 32 VSA-100 GPUs could be run in parallel on a single graphics card, and the fillrate of the card would increase proportionally. On cards with more than one VSA-100, the chips are linked using 3dfx's Scan-Line Interleave (SLI) technology. A major drawback to this method of performance scaling is that various parts of hardware are needlessly duplicated on the cards and board complexity increases with each additional processor.

3dfx changed the rendering pipeline from one pixel pipeline with twin texture mapping units (Voodoo2/3) to a dual pixel pipeline design with one texture mapping unit on each. This design, commonly referred to as a 2x1 configuration, has an advantage over the prior 1x2 design with the ability to always output 2 pixels and 2 texels per clock instead of 1 pixel and 2 texels per clock.

This is the first 3dfx graphics chip to support full 32-bit color depth in 3D, compared to 16-bit color depth with all previous designs. The limitation of 256px x 256px maximum texture dimensions was also addressed and VSA-100 can use up to 2048px x 2048px textures. Additionally, 3dfx implemented the FXT1 and DXTC texture compression techniques.

The VSA-100 supports a hardware accumulation buffer, known as the "T-buffer". When rendering to the T-buffer, VSA-100 can store the combined outputs of several frames. This mechanism allows for creation of effects such as motion blur (if used temporally) and anti-aliasing (if used spatially). VSA-100 supports rotated-grid super-sampling anti-aliasing (RGSS AA) modes, with a maximum anti-aliasing level determined by the number of VSA-100 chips in the SLI configuration. One chip allows 2X AA, two chips allows 4X AA, four chips provides for 8X AA and so on. The RGSS method of anti-aliasing combines multiple samples of each frame, resulting in higher quality than the brute force ordered-grid over-sampling of ImgTech PowerVR, ATI Radeon DDR and NVIDIA GeForce 2.

The chip implements a 128-bit SDRAM interface, again similar to the Voodoo3. Memory capacity and bandwidth is separately dedicated to each VSA-100 processor. While capacity is not cumulative across the entire card, bandwidth is effectively cumulative and thus a card with 2x VSA-100 processors has similar bandwidth to a single-chip graphics card using 128-bit DDR memory. Memory is clocked synchronously with the VSA-100 chip. Later, unreleased boards were planned to offer a 64-bit DDR memory design instead, in order to reduce board costs through lower complexity, while offering similar RAM performance.

While VSA-100 is an AGP 4x capable graphics processor, 3dfx did not implement AGP texturing.