An Interview with NVIDIA's David Kirk September 29, 2003 Brandon Bell

Introduction

Page:: ( 1 / 5 )

For the first time in years, NVIDIA is locked in a very real battle for graphics supremacy on the PC desktop. Up until a year ago, NVIDIA’s stiffest competition was itself, not only were they first to market with DirectX 7 hardware (with the original GeForce GPU), they also beat their competitors to market with DX8 hardware.

Then, in August 2002 ATI shocked the world with its RADEON 9700. Not only had ATI beat NVIDIA to DirectX 9 technology, they’d built a world-class part with breakneck performance thanks in part to its eight pixel pipeline architecture and 256-bit memory interface. NVIDIA’s DX9 part, GeForce FX 5800, was late to market due to several product delays. By the time it was released, over six months had passed since RADEON 9700’s launch; ultimately it was not released in significant quantities.

Since then, NVIDIA has introduced a top-to-bottom family of DirectX 9 products, ranging from the GeForce FX 5200 in the value segment, all the way up to the GeForce FX 5900 Ultra for hardcore enthusiasts. We’ve found that ATI and NVIDIA hardware compete very closely at the high-end, with each architecture winning their fair share of benchmarks. But more recent preliminary tests with DirectX 9 titles such as Half-Life 2 suggest that ATI has a significant lead over NVIDIA.

We were recently given the opportunity to conduct a phone interview with David Kirk, Chief Scientist at NVIDIA. In this position, Dr. Kirk has played a key role in developing the technologies that ultimately went into NVIDIA’s RIVA and GeForce line of video cards. There have been lots of questions surrounding the GeForce FX architecture that we wanted to ask about, more recently, NVIDIA has also shifted its focus from Cg, it’s high-level programming language for graphics, to Microsoft’s equivalent, HLSL.

FiringSquad: Could you give us more details on your recent decision to emphasize HLSL over Cg?

Kirk: You’re already seeing the ramifications of it as we begin to work more closely with Microsoft and optimize more and more for working with HLSL but let me give you a little history. Go back, say around two to three years when we started working on the FX architecture, we began to realize that complex programmable pipelines were going to require a high level language, and at that time there was no HLSL and we were working with other people in the industry including Microsoft to try and promote a programming language but we were not able to get any interest. So we started our own effort to develop Cg because we knew it would be required, whether or not anyone else thought it was important or not.

As we started out with Cg it was a great boost to getting programmers used to working with programmable GPUs. Now Microsoft has made a major commitment and in the long term we don’t really want to be in the programming language business and that’s not where our expertise is but its something we had to do, there was no other choice available. I think now that we have the opportunity to work collaboratively with Microsoft on HLSL for DirectX that’s a much more efficient way for us to work than to do it on our own. Now that doesn’t mean that we’re going to abandon Cg, there are other platforms that Microsoft does not support: OpenGL and non-Windows platforms, and also the professional workstation and content creation markets. There are a number of tools that have integrated Cg and we’ll continue to support those people and their markets.



SIDEBAR: Dr. Kirk first joined NVIDIA in 1997.

Cg (cont’d)/Detonator 50

Page:: ( 2 / 5 )

FiringSquad: So basically Cg isn’t going away exactly but it’s going to be less emphasized?

Kirk: Well there’s no need for us to focus on it on Windows and DirectX because we have such a good opportunity to work with Microsoft on it.

FiringSquad: I guess you didn’t see the rapid adoption of HLSL, when you came up with Cg?

Kirk: Well, going back as I said when we came up with Cg there was no HLSL. It’s not the rapid adoption, there wasn’t any. So once Microsoft and the rest of the industry have embraced high level shading languages there’s an opportunity for us to not be the sole developer on that technology. It’s much better for us to do it along with everyone else than to do it on our own.

FiringSquad: Could you give us some specific examples of the Detonator 50 drivers and where you guys think you can improve your performance and where you were inefficient before with Detonator 40 series and how you feel Detonator 50 is shaping up?

Kirk: I think that question is really hard to answer because what we’ve done with Detonator 50 is improve the quality of the instruction generation from the pixel shader (extrapolating pixel shader programs into hardware instruction set) and so the main place where I think you’ll see it is in applications that have shaders. Every shader will be optimized differently now than was before and will hopefully be optimized better. It’s not something you can say it’s going to help just here. Each instruction sequence in each shader is going to be progressively better with each optimizing compiler release.

I think the observation is when we first released GeForce FX we had not invested enough in the compilation and optimization technologies so we were not always generating very good programs for the hardware. Over time it will continue to improve as we go forward.

FiringSquad: Regarding Tomb Raider…is there anything in particular that you guys feel has been inaccurate in terms of the benchmarks that have been posted with Tomb Raider?

Kirk: Well I think the Tomb Raider as well as the Half-Life discussions; you see a lot of that on the web. And I think if I continue to comment on it, it just inflames it…the right way to look at all of these things is we really think benchmarks should be representative of gameplay and often in early release of benchmarks just like with early releases of drivers people are trying things out and I don’t really believe they end up being representative of actual game performance. So I think we need to give it a little time and you also need to just not think of a single benchmark, think of a collection of benchmarks as that being representative, because no one benchmark can really represent the hardware’s performance on all games.



SIDEBAR: NVIDIA recently unveiled GoForce, its family of products for handheld devices.

FP32 vs. FP24

Page:: ( 3 / 5 )

FiringSquad: Can you give us maybe a little background on why you guys decided on FP32 versus FP24?

Kirk: I think that’s a very interesting question. If you go back to how DX9 was developed it wasn’t clear as we were doing our development and Microsoft was doing their development and ATI was doing their development what the target precision was going to be for DirectX. If you look at processors FP24 doesn’t exist anywhere else in the world except on ATI processors and I think it’s a temporary thing. Bytes happens in twos and fours and eights -- they happen in powers of two. They don’t happen in threes and it’s just kind of a funny place to be.

FP24 is too much precision for pure color calculations and its not enough precision for geometry, normal vectors or directions or any kind of real arithmetic work like reflections or shadows or anything like that.

I think what ended up happening was during the course of DX9 development and discussions between the various parties the targeted precision changed several times and we took a snapshot when the precision being discussed was 32 and ATI took a snapshot when the precision was 24. In fact DX9 was released without any guidelines as to precision and a clarification was made later and the clarification that was made was very timed to ATI in that it did not make a statement that 24 was not enough.

Certainly one of the choices that Microsoft could have made is that it has to be 32 or nothing. They could have also made the choice that it has to be 16 or nothing. So it’s just kind of unfortunate for the whole industry, people really want to have predictable precision and predictable results and since the precision was under specified there was an opportunity for us and for ATI and for Microsoft to all make different choices and as you look forward you can ask ATI, you can ask Microsoft I think that you won’t see FP24 long term. I think that just like with CPUs you’ll see that 32-bit floating point as being the main calculation mode.

FiringSquad: Do you feel like you guys are at a disadvantage because of the fact that you don’t have a native 24-bit mode?

Kirk: I personally think 24-bit is the wrong answer. I think that through a combination of 16 and 32, we can get better results and higher performance.


SIDEBAR: You can also check out NVIDIA’s Computex presentation online, details are on the frontpage.

Hardware performance

Page:: ( 4 / 5 )

FiringSquad: One of the things that ATI has kind of said, or least they were suggesting at Shader Day is the fact that they can do more floating-point operations than you guys can. How would you respond to those types of statements?

Kirk: Well I guess the first response would be of course they would say that. But I don’t really see why you or they would think that they understand our pipeline, because in fact they don’t. The major issues that cause differing performance between our pipeline and theirs is we’re sensitive to different things in the architecture than they are so different aspects of programs that may be fast for us will be slow for them and vice versa. The Shader Day presentation that says they have two or three times the floating point processing that we have is just nonsense. Why would we do that?

FiringSquad: Could you give us specific examples of where maybe you feel you guys, you mentioned you guys can do some things better than they can, can you give us some specific examples of that?

Kirk: Well one example is if you’re doing geometric calculations with reflections or transparencies and you need to do trigonometric functions. Our sine and cosine takes two cycles theirs takes eight cycles, or seven cycles I guess. Another example is if you’re doing dependant texture reads where you use the result of one texture lookup to lookup another one. There’s a much longer title time on the pipeline than there is in ours. So it just depends on the specific shader and I feel that for the calculations I mentioned are pretty important for effects and advanced material shaders and the types of materials that people use to make realistic movie effects. So they will get used as developers get more used to programmable GPUs and we’ll have less of a performance issue with those kinds of effects.

FiringSquad: Do you feel that fact that you guys, your hardware came out later -- does that also contribute to the initial performance that’s coming out in terms of the DX9 titles that have been benchmarked with?

Kirk: Yeah, I would say that one of the issues is that since our hardware came out a little bit later some of the developers started to develop with ATI hardware, and that’s the first time that’s happened for a number of years. So if the game is written to run on the other hardware until they go into beta and start doing testing they may have never tried it on our hardware and it used to be the case that the reverse was true and in this case now it’s the other way around. I think that people are finding that although there are some differences there really isn’t a black and white, you know this is faster that is slower between the two pieces of hardware, for an equal amount of time invested in the tuning, I think you’ll see higher performance on our hardware.

FiringSquad: So you do think that the initial numbers that have kind of come out really aren’t indicative of final performance.

Kirk: [Yes] I believe. Of course, as I say if other people are commenting on our architecture without any knowledge, I don’t why [their comments on our architecture], why do you think it would be right? You know in our case I haven’t worked with all of the developers, but the ones that I have worked with have seen marked improvements when they start to actually work with our hardware and optimize for our hardware.



SIDEBAR: What happened to the San Francisco defense in yesterday’s game with Minnesota?

Half-Life 2 performance/Conclusion

Page:: ( 5 / 5 )

FiringSquad: Do you feel that in terms of the Half-Life 2 performance numbers that were released recently…do you feel that maybe you guys were, I don’t want to say given a bad rep, but maybe an unfair deal?

Kirk: Well again, not wanting to start a flame war back and forth, my feeling is if they had issues with speed, it’s really not appropriate to say that it doesn’t run at all. (Our question had mentioned this --FS) It’s just that so far in their state of optimization it doesn’t run fast. But after we’ve had a chance to work together on [inaudible] that will be able to provide a very good game experience with Half-Life on the full GeForce FX family. There’s no question in my mind that we’ll get there, it’s just a matter of time.

FiringSquad: So would you say that it’s a software type of issue that can be corrected?

Kirk: Yeah I don’t think there’s any issue with the hardware.

FiringSquad: Going forward what do you guys see in terms of upcoming technologies there are some new memory technologies that are coming up, is there anything else in particular that catches your interest?

Kirk: Well, lets see that’s a hard question. I would say that the advantage that we have right now is that people are getting used to much larger memory working sets, much larger amounts of frame buffer on the graphics cards and it creates a different operating style. I think where we go from that next is going to depend on the adoption of the developers and how much they take advantage of what we’ve provided them with this generation.


We’d like to thank Dr. Kirk for giving us the opportunity to speak with him. It is always enlightening to hear from the engineers that are behind the hardware we’re so passionate about, especially when they can shed some light on the design decisions that go into their latest products.

NVIDIA has a number of upcoming developments in the pipeline besides Detonator 50; we’ll certainly be keeping an eye on all of them in the near future.





SIDEBAR: Questions or comments? Feel free to chat about this article in the news comments!