Interview with NVIDIA´s Andrew Humber

March 5, 2003 / by aths / page 1 of 1

  Deutsche Version anzeigen

   Ten Questions to NVIDIA, relating to the new GeForceFX

Our interview partner is Andrew Humber, European PR Manager at NVIDIA (I think we don't have to introduce NVIDIA here).

1) You started at STB, then moved to 3dfx, and are now working for NVIDIA. Can you tell us if and what 3dfx technology made its way into the new GeForceFX GPU?

NVIDIA: GeForceFX is a combination of technology invented at NVIDIA and technology invented at 3dfx. When NVIDIA purchased the 3dfx technology assets, over 100 engineers from 3dfx were offered positions at NVIDIA and those engineers were immediately deployed across all of the products in development at that time. The team formed from the combined group took the best elements of each architecture and created GeForceFX.

2) As anti-aliasing is considered more and more important by our readers, we examined the new 6xS and 8x modes provided in the latest drivers. We have also released our guesses of what the respective subpixel masks might look like (http://www.3dcenter.org/artikel/2002/11-20_b.php). Can you confirm or deny whether our suggestions are correct? Will the GeForceFX´s 4x multi-sampling anti-aliasing subpixel mask be an ordered grid or a rotated grid?

NVIDIA: The 4xS mode for GeForceFX will use a rotated grid pattern. The standard 4x mode will use an ordered grid.

3) After the introduction of multi-sampling anti-aliasing which made high-performance FSAA possible even at high resolutions, NVIDIA seems to introduce more super-sampling mixed modes again. What are the reasons for NVIDIA to follow this "new old" strategy?

NVIDIA: NVIDIA offers many different anti-aliasing modes to give the gamer more choices. By adding "hybrid" modes that mix super-sampling and multi-sampling, we are able to offer more modes than we could otherwise. More choices is good for the gamer because he or she chooses where on the performance versus quality scale they would like to play games. For some older games that are fixed resolution or for any game that is CPU-bound, gamers can choose the higher quality modes such as 6xS or 8xS and still enjoy great frame rates. For newer games that are graphics-bound or graphics-memory bound, we believe gamers like to play with AA set to Quincunx or 2x.

4) We are curious: Why did NVIDIA chose to improve the efficiency of the 128 Bit memory interface rather than broaden the physical size? Can we expect a 256 Bit interface in 2003?

NVIDIA: Memory bandwidth is only part of the performance equation. Look at Matrox and their 256-bit memory interface as an example. How you use the memory bandwidth is more important. Performance is not just a function of raw memory bandwidth. Our efficiency with available bandwidth improved radically, which leads to better performance. With our color-compression, z compression and crossbar memory architecture we use the available bandwidth more efficiently. We also have other factors like DDR-2, new vertex shaders, new pixel shaders and .13 micron manufacturing technology that reflect back to higher performance. Bottom-line, theoretical numbers are nice to discuss, but it is the real world performance that really counts.

5) One of the GeForceFX´s new features is something NVIDIA calls "Adaptive Texture Filtering". We read that the chip adjusts the sample mode on a per pixel basis. But the most impressive part appears to be that the chip is actually monitoring texture content to decide on what sample mode to use. Can you please tell us something more about this feature?

NVIDIA: The GeForceFX family has a special adaptive texture filtering mode that the end user can turn on or off using sliders in the control panel. Conservative settings will perform the traditional texture filtering that the application or the user requests. Moving the slider to the Aggressive side will engage the adaptive texture filtering algorithms that can deliver higher performance by analyzing the texture data and making intelligent trade-offs to increase performance while maintaining high image quality.

6) The GeForceFX allows for very long Pixel Shader programs. This is certainly a nice feature, but quite a lot of people consider this to be a waste of resources considering such long programs will probably never run with good performance, in real-time. Is it possible to give us some performance numbers such as the number of Pixel Shader operations GeForceFX can handle per second?

NVIDIA: The more instructions you have, the more realism a developer can add to a game – ultimately, the developer or artist will determine the optimum amount of instructions to get the best blend of performance and image quality/realism. There will be always be a trade off, but by giving this power to the developer, we are working hard to ensure that the wave of next generation games look as good as they possibly can.

7) As expected, the GeForceFX still supports all OpenGL extensions of its predecessors. Also, the register combiners can work in combination with the new fragment programs. Because these two kinds of fragment proccessing have not much in common, we would like to know if register combiners are emulated in the Pixel Shader processor, or whether GeForceFX uses extra hardware for its register combiners?

NVIDIA: This is proprietary information I´m afraid.

8) In the near future, we will see Doom III and a number of other games based on the new id engine. We already know that that Doom III will highly stress stencil buffers. Does the GeForceFX offer any features increasing performance such as "Fast Stencil Clear", "Early Stencil Test" or a "Hierarchical Stencil Buffer"?

NVIDIA: GeForceFX is very fast at stencil operations, but we are not marketing any stencil-specific implementations.

9) For the GeForceFX, NVIDIA has developed a new color compression technology that can compress color information up to a ratio of 4:1. Is color compression used in all cases, or only with activated anti-aliasing? If it is used even for non-anti-aliased modes, does this mean that the color compression will also work on textures that are not already compressed using, for example, DXTC? Specifically, are textures created by render-to-texture compressed as well, using color compression?

NVIDIA: The GeForceFX colour compression is activated only for anti-aliased rendering and only applies to the frame buffer. The GeForceFX uses DXTC and S3TC texture compression for textures.

10) Will NVIDIA support TCPA-technology in future drivers and hardware?

NVIDIA: NVIDIA will be a leader in TCPA-related technology and products.