Interview with Scott Sellers

One on one and no holds barred

After the press conference, we had the opportunity to sit down with Scott Sellers and grill him about the aspects of VSA and Voodoo4/5 that intrigued or baffled us, and trust me, we had a number of questions! The conversation is compiled here in no particular order.

3dfx CTO, Scott Sellers

With the recent release of the Voodoo3 3000 PCI and Voodoo4/Voodoo5 products, we had to ask why 3dfx was suddenly focusing so much on the PCI market. Well, Scott said that 3dfx had always heavily developed for PCI, and that it was the highest-volume graphics product market out there. Just like the throngs of people with Nintendo Game Boys and NES systems (far more than Dreamcast/Playstation may ever sell), the legacy market of PCI remains healthy, and is probably 3dfx's alone to dominate.

We then asked why 3dfx is staying on a supposedly dated 0.25 micron manufacturing process for a complex 14-million transistor part, while competitors Nvidia and S3 have moved to 0.22 micron and 0.18 micron, respectively. Scott matter-of-factly answered that Nvidia uses a 5-layer metal 0.22 process, which results primarily in smaller die size and lower power consumption. By using a six layer 0.25e (enhanced) micron process, 3dfx is able to shrink the die accordingly by cramming more onto the same relative area, and retain improved product yields. S3's 0.18 micron process for Savage 2000, he said, uses 0.18 micron transistors, but on 0.22 micron metal, and cannot truly be called 0.18um. Guess we'll be waiting for Coppermine to get a taste of that after all.

Who needs the Z?

I asked about the inclusion of full 24-bit Z and W buffers. Recently, we've heard that a W-buffer, commonly used for perspective correction, and also be used for depth buffering, but that a Z was more accurate at large distances, and the W was more accurate up close. Scott said that in "normal" implementations we see today, that's correct, since they use an integer Z buffer and a floating point W. VSA-100 has support for both floating and fixed values for both the W and Z buffers, so there is no inherent loss in accuracy. Since W information has to be sent down the bus for perspective correction, it makes sense to use the W for depth information as well, bypassing the need to send extra coordinates down in the form of a Z buffer.

Currently, DirectX 6.0 supports the W buffer, while there is no equivalent support in OpenGL. We asked if there were plans for a GL extension, and he replied that it was a distinct possibility, since use of the W-buffer wouldn't be an exclusive 3dfx functionality.

In terms of final product speed, all VSA-100 products will be shipped at either 166MHz or 183MHz, depending heavily on memory prices. There is no inherent advantage to running different core/memory speeds. 3dfx is also hoping to push the RAMDAC up to 400MHz by time of launch, but that has yet to be seen.

Asked about overclocking, Scott mentioned that the number of powering grounds on VSA-100 boards is substantially higher than in Voodoo3 architecture, which should make overclocking much easier and successful than what we've seen for the current generation of 3dfx products. This should be a welcome option for hard-core users without the budget for a V5 6000. Just for reference, you'll have to boost a Voodoo5 5000 or 5500 up to 250MHz in order to reach the 1 Gigapixel mark. Try that on 6.0 or 5.5ns RAM!