Behind the Scenes with AMD's Eric Demers - An Interview with Rage3D

Company: AMD
Author: James Prior
Editor: Charles Oliver
Date: July 11th, 2011

Teasing the Next Gen

R3D: So, the next generation graphics core just got disclosed, not a high level of detail.

Eric: Well, the core architecture, yeah.

AMD Graphics Core Next Gen

R3D: First things the guys at Rage3D will want to know - alright, how does that make my gaming better?

Eric: So, how does the new compute architecture help graphics? Well, it's pretty cool, there's a couple of different ways. First the distinction between what is compute and what is graphics is blurring at a rate that you can't imagine. It's not just about DirectCompute, but there's more and more compute; either through DirectCompute or through OpenCL; there's more and more compute being used with graphics. Look at DiRT 3, they're doing DirectCompute for physics computations on all the water, the flag movement effects, doing some of their shadow effects, they're doing all their post processing effects as compute. More and more of the frame is being interspersed with classic DX11 pipeline and compute. As we go forward I see that blurring completely, I find it hard to even see a separation. I see a separation between one API and the other. What the hardware does fundamentally is effectively very much the same.

R3D: What you're saying is, we've gone from disguising compute as graphics, to graphics as compute?

Eric: Yes. Now, there are unique elements that are tuned for graphics. Like having a scan convertor, having color and Z buffering; those things continue in this architecture. While we focused [in the keynote and deep dive] on the compute side of it, to introduce that so that developers could see, and how that would work with the Fusion System Architecture, fundamentally all of those things are there, available. This new architecture is going to improve compute and it is going to improve gaming at the same time.

Eric: Now, there are specific aspects of the architecture, such as virtual memory support, that aren't easily exposable with current OS support. But, PRTs are our way around that [limitation]. We said 'Let's make a customized version of that, just for graphics’. So, Partially Resonant Textures enables, for example, hardware mega-texturing; enables hardware texture streaming; allows much larger volumetric data. It allows a bunch of different stuff that is very difficult to do, that gives stutters when you are running applications; all of those, now, can get improved for applications that use them. We have to create that as an OpenGL extension, and we'll have to create the equivalent of a DX extension to expose under DX, but fundamentally that kind of capability in hardware which leverages the compute elements of FSAIL is exposable in graphics. We're going to do it in a bunch of other places as well; we haven't given all of our secrets out yet! There's no doubt in my mind that we look at it holistically, that compute and graphics are all intermingled and we're going to advance both technologies.

Eric: We're going to advance and push both APIs, we're going to advance and continue to push compute, and I think, in the end, gaming, specifically, will benefit from all of that.

R3D: Is PowerTune a part of your strategy for the next graphics products?

Eric: We haven't announced future products at this point, we've just shown Trinity, that's all we've effectively said. It's a pretty safe bet to say that PowerTune is a really key asset for us. Obviously it was oriented towards the previous architecture, and in a newer architecture it would probably need an update to that. The idea of being able to push harder on an architecture when it has thermal and power [budget] available to it, is a really great thing. In fact, even our CPU's are going to be using some of the same. Turbo boost and those kinds of things are very similar kind of capabilities. So, to think that's going away would be foolish, but to think that we're going to improve it, make it better, is a reasonable thing. Certainly, this architecture is fully supportive of anything we would put the power tune branding to, in it.

Eric: Yes, going forward for us, being able to maximize the utilization or our available power and thermal is really key. If you're running 30% slower than you need to because some application once ran into that limit, you're artificially limiting your hardware. We want to be able to let you push it all the way to the limit, because otherwise we're just shortchanging you a little bit, right? PowerTune allows that on Cayman, and gets you very close, or closer, to running at that TDP [limit]. Room for improvement? Definitely, going forward we're going to have more of that.

R3D: It occurs to me then, that with your new ability to run different workloads simultaneously, then PowerTune at a Compute Unit level would be smart move.

Eric: PowerTune is already at the SIMD level in the current architecture, it actually monitors each of the SIMDs and looks at the issue rate and texture miss rate; it looks at all types of metrics inside each of the SIMDs to figure it out. All of that will map to the Compute Unit and probably new things will be also added to those measurements. So there's not fundamental architecture change that is required. We just need to retune what is being monitored for the Compute Unit. Fundamentally this [the next generation compute core] is a more efficient unit, I think we'll get even more performance out of it but we'll have to wait until the products. Anything I say should be taken with a grain of salt, at the end it's the product that matters, when you guys get those products and do the reviews you'll be telling us if we did a good job or not.

Eric Demers, AMD CVP & CTO, Graphics

At this point we had to end the interview as it was time for Eric to get his stuff and catch his flight, but we did chat with him as he walked to get his bags. The subject of the IQ wars, and Anistropic Filtering came up, and it turns out that AMD, internally, take criticism badly; that is, they take it personally and to heart. They are working hard to provide the best quality products they can, and they are not going to stop trying to deliver.

AMD's Graphics Core Next Gen architecture is scheduled to appear in products due in Q4, 2011, as both Eric Demers and Rick Bergman confirmed for us. It is likely they will use TSMC's 28nm process, and have a dual architecture split like Northern Islands. If so, then this time Cayman VLIW-4 architecture hits the mainstream, and the next generation architecture debuts in the high end - similar to AMD's deployment of VLIW-4 in the Radeon HD 6000 series. This makes sense for compatibility with APUs, if Trinity is using VLIW-4 then the discrete GPUs will need to as well for best compatibility and performance.

Rage3D would like to thank Eric Demers and Dave Erskine for their time and expertise, and their hospitality at the inaugural AMD Fusion Developer Summit. We'll bring you more on AFDS and the Llano Lynx platform shortly - watch the Rage3D front page!