3dfx VSA-100

The 3dfx VSA-100 (Voodoo Scalable Architecture) was a GPU designed for the Voodoo 5 series of expansion cards. It was announced at Comdex in late 1999.

As implied by the acronym, the primary advantage of the VSA-100 was its extreme scalability. Theoretically, anywhere from one to 32 VSA-100 GPUs could be run in parallel on a single graphics card, and the fill rate of the card would increase proportionally. However, 3dfx had only managed to release cards containing 1 or 2 VSA-100 chips, although the never-released prototype Voodoo 5 6000 featured four such GPUs.

The VSA-100 had hardware support for FSAA, but lacked hardware T&L support. As the latter feature grew in importance, the VSA-100 architecture soon became technologically obsolete.

Architecture

The architecture was a direct descendant of "Avenger", Voodoo3, but rather than one pixel pipeline with two texture mapping units the VSA-100 featured two pixel pipelines with one texture mapping unit each, which was enabled by the process shrink from 350nm to 250nm. This doubled its fillrate in single texturing applications while maintaining fillrate par with "Avenger" per clock while multitexturing.

The VSA-100 supported a hardware accumulation buffer, known as the T-buffer. This solved a problem which still plagues modern GPUs, that the GPU has no way of knowing what it drew in the last frame(s). When rendering to the T-buffer, a VSA-100 could store the combined outputs of several frames and enable effects such as motion blur (if used temporally) or antialiasing (if used spatially). Modern GPUs use pixel shaders to emulate these effects, the VSA-100 could do them directly.

In a near-catastrophic piece of bad design, the VSA-100 only supported the same 128-bit SDRAM interface as the Avenger (it seems that they share the exact same memory controllers) and in an era where RAM bandwidth was taking over from pixel fillrate as the king of performance, DDR equipped Radeons and GeForces were much higher performing than VSA-100. In order for a VSA-100 to perform on par with its contemporaries, it absolutely had to be used in a dual-GPU configuration, which raised costs but also doubled the memory bandwidth to more competitive levels. The planned VSA-101 (on 180nm) was to feature a 64-bit DDR interface.

VSA-100's other major limitation stems from the re-use of Avenger's memory controller. The memory had to be run at the same clock as the VSA-100 core proper when Nvidia's GeForce series could use slower or faster memory to target different markets. VSA-101 was to share this limitation. The low-end Voodoo4 4500 (single VSA-100) was not what the VSA-100's designers had in mind and it could barely out-perform a Voodoo3, let alone a GeForce2 MX.

3dfx clearly intended VSA-100 to be massively parallel but such a design required multiple VSA-100s per implementation, which raised costs. Two VSA-100s in parallel (e.g. Voodoo5) gave a 4x1 pipeline with the equivalent of 128-bit DDR, yet the GeForce2 which was available at the same time had both 128-bit DDR and a 4x2 pipeline setup. In order to remain price/performance competitive, the VSA-100 needed to be much cheaper to produce and it was, having merely 14 million transistors as opposed to GeForce2's 27 million. Sadly this cost saving simply transferred over to the added expense of implementing the pair on the same PCB, double packaging costs, double cooling costs and added PCB complexity.

External links

* [http://www.cnn.com/TECH/computing/9911/16/comdex.3dfx.idg/ 3dfx announces Voodoo 4 and 5]
* [http://firingsquad.com/features/3dfxv4v3/ 3dfx VSA-100 Interview]
* [http://guru3d.com/voodoo4/voodoo4.htm Voodoo 4,5 and the VSA-100 Graphics Processor]