RAM and SLI

So, multiple VSAs per board...

One of the biggest questions I wanted answered about the VSA-100 was related to its scalability. If you can have two (or more) VSA-100 processors on a board, is it possible to decouple them from SLI mode for Windows 98 muilti-monitor support in 2D, 3D, or both? According to Scott Sellers, this is not only possible, but required only trivial changes to the hardware PCB and supporting circuitry. In fact, 3dfx had plans for a board with 2 DB15 connectors at launch, and held back only to keep from saturating the channel with too many products. They are very eager to see how much demand exists in the market for this kind of product, so let us know whether or not this would persuade you to purchase a 3D accelerator!

The biggest question

The biggest complaint we've heard about VSA-100 (especially from 3dfx's competitors) is the need for multiple chips to compete with others' single-chip offerings. The redundancy in silicon, as well as larger PCB design, electrical cost, and cooling requirements should make VSA-100 much more expensive to produce than an equivalent single-chip solution.

Scott acknowledged this, but maintained that 3dfx chose explicitly to go the route of multiple processors. He stated that it's actually much more difficult to engineer a chip that can scale massively if required. This allows 3dfx an incredibly broad spectrum of products, ranging from low-end consumer/OEM all the way to high performance graphics workstation products. The main benefit is that the company does not need to spend time/effort on any additional technology past the PCB and local bus/chipset level. Single-chip solutions however, must re-engineer their chips for different markets, but adding custom features for high-end solutions, and cost-reducing in order to hit OEMs.

More on redundancy

Another issue we've encountered with VSA-100 is the need for redundant memory. By now, everyone's seen the specs for the Voodoo5 6000, which boasts 128MB of RAM. Unfortunately, at first glance this may seem a bit misleading - while there actually is 128MB of physical RAM on the board, that RAM is segmented, or separated between all four VSA-100 chips on that board. This is done because each chip must render different parts of one scene, and so the textures much exist in the buffers for each chip - this was the same situation for 3dfx's Voodoo2 SLI. What this means is that in a 4-chip, 128MB solution, the card actually has access to only 32MB (128MB total / 4 VSA-100 processors) of RAM. Memory costs today are by no means trivial, but the redundancy alone seems wasteful.

This sounds dismal, but are there any upsides? Certainly. As a parallel processing system, the major benefit is bandwidth. The individual bus between each VSA-100 and memory bank allows for 2.93GB/sec of memory bandwidth at 183MHz. This is doubled for a 2-chip configuration, and quadrupled for a 4-chip setup. Also, textures aren't the only use for memory. The frame buffer, which can take up several MB depending on the color depth, resolution, and refresh rate, can be shared across each memory bank, leaving more space free for textures.

In addition, remember what we discussed before, specifically multi-monitor support? Well, if 3dfx creates such a product, the memory bank for each chip would be decoupled from SLI along with the processor, giving each processor full, independent access to its memory. In a hypothetical game spanning multiple monitors, this would equate to having literally two (or four) 32MB VSA-100 cards installed.