Clock Speeds, Dual-Head, and Memory

FiringSquad: The T-Buffer and FXT1 are both active and deliberate steps to move towards "Hollywood on the desktop." What do you see as the next immediate limitation or roadblock to overcome in order to further advance towards this goal?

I hate to keep coming back to fill rate but like I said earlier we still have five orders of magnitude to go! We need higher fill rate to get to more sophisticated levels of many of the techniques that we already use like better multitexturing for lighting and shading, more T-Buffer samples, and deeper color depths. We're also quickly approaching a system-level bottleneck with AGP 4X. As scenes get more and more complex even that bus will soon be choked with vertex traffic so we'll need some help from the folks who create PC bus standards to keep improving the PC as a gaming platform.

FiringSquad: Have you decided on what clock speed to release Voodoo4/5 at(166MHz or 183MHz)? Are there plans for faster speeds or a move to a segmented core/memory architecture?

The VSA-100 based products will have a synchronous core/memory clock. We are not prepared to announce the clock speeds at this time.

FiringSquad: We reported earlier on the possibility of dual-monitor or multi-output video/3D with multi-chip VSA-100 solutions. Is this still under consideration for a future product?

I hope we did not mislead you, we said it was possible. We never alluded to the probability. It is still possible with the VSA-100, but we have not announced any plans for that kind of solution.

FiringSquad: It seems the inherent drawback with multi-chip solutions is scalability - on one hand VSA-100 can use up to 32 chips, but on the other, you'd need 32x32MB RAM to get a 32MB frame buffer on such a product. Is 3dfx looking into ways to alleviate this problem?

Well first off we all have to address this unified frame buffer issue again. What you said isn't correct. The frame buffer is distributed across the memories attached to all the VSA-100 chips: logically it's one big frame buffer. We have to repeat texture data for each chip but that's it. We carefully considered this before designing the product and concluded that any drawback from repeating texture data is vastly offset by the benefits of scalability like leveraging software across a huge range of products, offering compelling products to serious gamers and professionals at many price points, and solving the memory bandwidth issue with highly-available, reasonably-priced mainstream memory.

I think that relying on very scarce and very expensive memory types to get performance gains with your chip is a drawback. I think the lack of DDR cards available in retail is a testament to that. I believe that the ability to use more of a highly available and economical memory type is a benefit of a scalable design and greatly increases our company's flexibility.