The raw fill rate of a graphics system based on the VSA-100 does scale linearly with the number of processors. How that impacts actual content performance depend on the complexity of the content, the resolution, and the color depth. At 1024x768 resolution and 32-bit color most titles are fill-rate limited, the fill rate of the graphics system is the bottleneck to performance. In such a case the performance will scale with the number of VSA-100 chips used up to the point that fill rate is no longer the bottleneck. The good news is that the most complex titles, the ones using the most textures per pixel like Unreal Tournament, will benefit the most from multiple VSA-100 chips.

Technically we could do single-board multi-head products, and we haven't ruled it out, but we don't have any plans to talk about, either.

Our goal has always been to be the best 3D accelerator for OpenGL and D3D, in addition to supporting Glide. Now, our efforts are completely focused on providing the best DirectX and OpenGL drivers. The Open Source community will continue to evolve Glide as required by the market.
Software developers have expressed their wishes to focus their efforts on D3D and OpenGL. OpenGL and D3D are very mature and flexible APIs with advanced 3D features and we have been very successful in getting our new features added into both APIs. We have the ability to add revolutionary and unique features to OpenGL through its extension mechanism, and Microsoft has been consistently aggressive in evolving Direct3D every year with compelling new 3D capabilities. These factors have caused OpenGL and Direct3D to be the preferred APIs by the development community, thereby making a proprietary API such as Glide less important for the industry.
We plan on supporting current Voodoo3 customers and will provide compatibility support for products announced through the end of 2000.

It's the features that the TNT2 Ultra doesn't have, like DXTC and FXT1 texture compression, SLI, full-scene AA, and T-Buffer effects, that took the extra time. That's why VSA-100 products are positioned against GeForce and NV15, not old stuff like the TNT2.

The T-Buffer and accumulation buffer or A-buffer produce similar results but differ in operation and implementation. Actually I don't know of anyone who is including an A-buffer in a PC-based graphics system today. Of course the key to enabling developers to use cool features like motion blur, soft shadows, and depth of field blur, is exposing the features through a standard API. Once that's done the hardware implementation details should be a don't-care for the developer if the hardware is well-behaved. We're working hard to insure that all T-Buffer effects are exposed through industry-standard APIs.

The GeForce's fill rate is obviously too low for a high-end gaming card. I think that so many key applications are fill rate limited that the benefit of GeForce's geometry acceleration is questionable. At a published 480Mpix/sec. it trails our slowest SLI board by 40%. I think we've all read the unauthorized specs for NV15 that were published on the Web this week and have seen that even that "next generation" product still trails our entry-level SLI's fill rate.

I wish it were as simple as saying "T&L doesn't help today but it's important at the end of this year." The real issue is how you balance fill rate, multi-texturing capability, digital effects, and geometry processing, within a specific system to achieve the best results. A good question to ask is "Where are the performance bottlenecks in my system and how do I best relieve them?" As I said earlier, when playing the top titles at the resolutions and color depths that serious gamers prefer the bottleneck is fill rate. It doesn't matter how much geometry acceleration you have if you can't render the image fast enough. In that regard the GeForce is an unbalanced part today and as games become ever more complex it will be an even more unbalanced solution in a year. When will the CPU's floating point capability become the bottleneck for the best titles? I think that people might start looking at getting a new card with geometry acceleration, and by that I mean a card that's leading-edge at the time that you buy it, no sooner than the end of 2000.

Graphics memory bandwidth often limits fill rate, hence performance. Part of the reason for returning to SLI was to optimize our memory bandwidth to cost ratio. A Voodoo5 5000, with two VSA-100's operating in parallel, has an effective 256-bit memory interface that provides bandwidth equivalent to a 128-bit DDR interface. DDR memory is going to be expensive and scarce in the first half of this year. Embedded memory forces severe compromises in logic cost and complexity since embedded DRAM processes are always a generation behind the mainstream dedicated logic and memory processes. I can't tell you specifics of future plans but our goal is always to provide the highest bandwidth with memory that's readily available and reasonably priced.

Well, Savage2000 has a 128-bit SDR DRAM interface that they're running at 166MHz. That's the same bandwidth as a single VSA-100 so it's great for a mainstream product but it's only half the bandwidth of a Voodoo5.

Our goal is to have a strong ICD in the box. I know that we've disappointed end users with our ICD situation for Voodoo3. We will keep improving all of our drivers so there will of course be regular maintenance releases.

What a great question. While it's easy to get frustrated when you buy a PC product, really almost any PC product, only to find that there's something better in six months. I think we all understand what an incredible benefit it is to have technology improve so quickly. Why is it a benefit? Even though what you bought today could be obsolete in six months it is still superior to what you would have bought six months ago. In other words, you can afford better things today than you could have six months ago. Voodoo Graphics was far ahead of the industry because we dared to build a product that everyone else, all our competitors that is, thought was totally absurd. 3D only? Multiple chips? A pass-through cable? A proprietary API? Are you guys nuts? Of course after we had figured it out it the competition came on hard but it still took over two years, an eternity in this business. Competition in PC graphics is unrelenting, the rate of improvement in silicon technology shows no signs of slowing, therefore the short product cycles will continue. Our goal at 3dfx is to come up with those wacky ideas that lead the market by years, not months.

Interview by Chick's Hardware UK