NVIDIA's chief scientist and VP of Architecture, Dr David Kirk, was recently in Australia. John Gillooly and James Wang put him through the Lunch-o-Tron regarding the future of 3D hardware.
Atomic: Intel has told us repeatedly that PCI-Express will first appear as an AGP replacement and NVIDIA has made public statements of support for the standard. What impact do you see PCI-Express having on 3D hardware?
Dr Kirk: I think PCI-Express is very exciting because AGP has essentially run out of steam with AGP 8x, there is no AGP 16x and, as I said in my presentation, we are trying to reduce the dependence on bandwidth by architecting to take advantage of more computation, but bandwidth still matters. And free bandwidth is still very high but we are still limited by AGP 8x and we can draw small polygons faster than we can transfer them.
Atomic: Will PCI-Express reduce reliance on dedicated GPU memory?
Dr Kirk: The ratio of system memory performance to GPU memory performance has been about a factor of 10 for as long as I have been in the industry. So any time you are willing to give away 90% of your performance you don't have any more reliance on GPU memory.
The need for it comes from the fact that the signalling technology in AGP 8x is really at its limits so we are moving to a new level from where we can build. Bandwidth is really important.
Atomic: Will PCI-Express, implementation comes with a wholesale shift to the technology when will a new generation of cards be announced?
Dr Kirk: It would be nice if the industry would just flip a switch and all change at once but unfortunately I don't think there is ever going to be a step-function transition like that. We will continue to support AGP 8x. Right now we are AGP 8x top-to bottom and when PCI-Express' introduction comes we will very quickly have top-to-bottom products for PCI-Express and AGP 8x as long as there is interest in the product. For example: we still sell TNT2 and Vanta cards because people still want to buy them. Millions of them a month, that's how much people want to buy them.
Atomic: The main reason cited for the NV30 delay is the problems with 0.13-micron technology at TSMC. NVIDIA has always aimed at the most advanced process to push their GPU to the limit and one-up the competition. This strategy has worked superbly until the NV30, which for the first time has dropped the ball. Is it because TSMC's foundry is second-rate fabrication technology compared to IBM or Intel?
Dr Kirk: I disagree
Atomic: You don't think this is some sort of bottleneck?
Dr Kirk: No. You said 'second rate'. That I didn't agree with. You are completely wrong.
Atomic: Sony and Toshiba already currently have working libraries for 0.065 SOI processes, while TSCM is struggling at 0.13.
Dr Kirk: Don't confuse marketing with reality. How much silicon are they are actually producing with that? It is perfectly easy to say I have working libraries. It is perfectly easy to say I can manufacture one device that works, but mass production is completely different. I would say that the gap in time between TSMC and the best is only six months.
Atomic: Has NVIDIA looked into gaining higher tech fabrication technologies from process leaders like IBM and Intel?
Dr Kirk: The future is the future, and everything is unpredictable, but TSMC and NVIDIA have had a great partnership and 0.13 as a process, particularly with copper wiring, presented a lot of technical problems for everyone. And again, I don't want to name names, but the very first pioneers who went into production on 0.13 had enormous problems with copper. Everybody had to conquer those problems in different ways because the design was different, but 0.13 was a big challenge. And that is why I mention that it is easy to say "oh yeah, we have libraries for the process" but actually getting it into production and getting good yields - those are hard. The next step after 0.13 is 90 nanometres and it will have different problems and we don't even know what they are yet.
You mentioned 65 nanometres; well, there isn't even anybody in production with 90 yet.
Atomic: Far down the track - how do you ultimately see the GPU in five, ten years from now? It's said that everything done in 3D right now is just a hack. Which hacks will remain and which hacks will be replaced with technically correct solutions a decade from now?
Dr Kirk: Well 3D graphics is all about cheating without getting caught.
Atomic: Will you get better cheating methods?
Dr Kirk: We'll just get better and better hacks. What is happening is the world is simulating radiant light bouncing around and reflecting and all that stuff. It is amazing as nature does it in real time. And we do a lot of approximations because simulating the amount of photons that are in this room bouncing around, we're not going to have computers that can do that for a long time.
So, we have to take short cuts. So, your question about which hacks are going to remain - one of them where you are going to start seeing a change right away is textures. I think that you are going to see a lot more synthesised and calculated patterns rather than textures, I don't think textures are going to go away, but I think the dependence on textures is going to be reduced.
Atomic: How will the GPU's internals look? A massive memory embedded chip with large numbers of virtual SIMD arrays perhaps?
Dr Kirk: I think you will have a massive number of processors that are working independently. The architecture will continue to take advantage of the streaming processor architecture where most computation operates on data that is already on the chip, most of the time. There are a lot of challenges. We were talking about four or eight pipelines, how about 2048 pipelines? How about using those pipelines when you are drawing one particular triangle. Essentially all of Dawn [NVIDIA's flagship GeForce FX demo star] is very small polygons. It would require a lot of changes in architecture to think about - how do you program a 2000 processor pipeline? I don't know, nobody knows. Those are the problems we have to solve in the architecture.
Atomic: Looking even further down the pipe, and with the current trends of large-scale integration, eventually the CPU will takeover the role of the GPU or vice versa?
Dr Kirk: I think we've shown why the GPU path is going to win. The only way that the CPU can do all this processing just as fast is by becoming a GPU.
General purpose computing is not very hard, how much faster do you need your Excel spreadsheet to go? What you really care about is having high definition video, being able to have 3D graphics of better and better quality. Media kinds of things that GPUs are really good at, those are the things that we want to go faster. How fast do you want your keyboard and mouse processing to go? It's fast enough already.
Atomic: When the first single-die, dual-core CPU/GPU arrives, whose label will it have, Intel's or NVIDIA's?
Dr Kirk: I don't know. The future is hard to predict. I think we'll continue to try and make exciting and compelling products; people don't get passionate about a CPU but will get very excited about a GPU.
Atomic: The floating-point unit on the GeForce FX has uses beyond 3D graphics. What about programs like Seti@home using the GPU during idle times?
Dr Kirk: Yeah, we are actually working with the SETI at home people for their new client, which is not just for processing radio waves. It is actually a framework where you can have it run
Issue: 133 | February, 2012