An interview with ATI - Dave Baumann

Dave, in your vision, what would be the biggest challenge here, as we feel that software developers hate the fact that there will be two standards (more standards = more programming = more expensive = a higher investment in the game). Initially I expected (well hoped) GPU Physics to be an integral part of DirectX 11, and while that's partly true (compute shaders could help here) it's far from optimal from what I heard. What is your feel and take on this, because in the end it's all about improving that gaming experience, and hate or love it... Physics certainly brings a new level of effects into the gaming arena.

OpenCL and Compute Shaders are merely interfaces to particular processor types and, on their own, are unlikely to be the answer for the majority of developers because they have to do all the work from the ground up in their game / engine for physics. Generally speaking what developers want are tools in order to help them get it done and they don't necessarily want to get distracted by underlying standards that may limit them. Instead they want tools that help them optimally target the platform.

This is exactly why we are working with Havok for Physics - because they clearly have a world class physics middleware toolkit and they have a huge portfolio of titles that have and will used it. By working with Havok and OpenCL we are helping to take barriers away from developers, hopefully making their lives easier by enabling the best middleware with a backend that is processor (brand and type) agnostic.

There's no interview possible without mentioning and asking where you guys are heading in the future. We already moved from fixed graphics pipeline, to unified programmable shaders. There is of course also a new party tapping into the GPU market. It's Intel's Larrabee. Larrabee will be fully programmable. In fact it takes me back to the 90ies in the early stages of GPU development. What Intel is doing is pretty radical, completely software programmable, super scalar... a bit of a bold brute force method one might even say. What is your gut feeling about Larrabee looking at it from an engineering and architectural point of view? Do you see ATI's future GPUs moving into this direction as well?

The path of GPU evolution has been fairly consistent, from pure fixed function texture mappers / rasterisers, to more flexible fixed function devices that encompass more of the 3D processing pipeline, to programmable compute devices. Following the evolution from DirectX 8 capable processors the programmable compute elements of the GPU are obviously the parts that are both increasing in physical area on the die itself and increasing in flexibility in terms of what they can achieve. In no way do I see the direction of that evolution altering over time.

The trick, though, is about getting the best mix of ingredients in order to get the balance right for the period of time the processor is available. We've tended to talk about the "Sweet Spot Strategy" in terms of market positioning and time to market, but equally you can apply that to the design of the architecture in regards to what it is going to be judged on and where end users are going to get the most benefit.

We started discussing Stream processing back when we had the ATI Radeon ™X1800 launch, and we were the first to have a Folding@Home GPU client running on those processors, however over the following generations we have increased our scope and capabilities of ATI Stream to the point where we have a fairly comprehensive set of capabilities and very good performance. It is clear that at this point that nobody is buying graphic boards based on the state of GPGPU applications - gaming performance is still the clear purchasing factor. This is where the ATI Radeon™ HD 4000 series hit the other Sweet Spot, in terms of providing a sufficiently forward looking architecture to promote the use of the processors beyond graphics, but without dedicating too much of the architecture to the point that it impacts gaming performance. It strikes me that our competitor didn't ride that line well enough for the current generation of processors. On the other end of the spectrum it's likely that Intel's "Larrabee" may lean far towards the compute side of things.

Once we, and other IHVs, have OpenCL fully available and we start seeing more consumer-oriented applications making use of it, the focus on Stream capabilities and performance will increase. It's true to say that we are already seeing a rapid growth in interest and adoption of OpenCL by ISVs, now that it has been ratified.

Most likely it will take a long time for this to be the primary factor in the choice of performance and enthusiast graphics boards. Curiously, stream applications serving as an influencing factor may occur quicker in the mainstream and entry level spaces as OEMs and end users begin to look at the productivity benefits of it.

2 GigaFlops per GPU in 2010?

Ha! No comment!

Dave, after seventeen questions I figured to call it a day. I want to thank you for your time to answer the questions we asked. If there's anything you like to add to this interview, now's a good time to do so :)

I think you tapped me out actually - my fingers need a rest from all this typing!
Just thanks for giving me this opportunity to sound out. Hope your readers enjoy it.

We'd like to thank Dave Baumann for his time and dedication to answer some of our questions.

We now like to invite you the reader towards this web page, here we started a contest in collaboration with two AMD partners and give away 3x Radeon HD 4890 graphics cards and 3x Radeon HD 4830 graphics cards.

• If you like this article please digg it.
• Leave comments or read feedback
• Sign up to receive a notice when we publish a new article
• Or go back to Guru3D's front pages