Graphics Processors Design
X-bit labs: Does the approach to GPU design change with different generations of GPUs?
David Cummings: GPUs continuously evolve on all levels, from process nodes and architecture to materials. Our HD 4000 series differs significantly from our HD 3000 series, even though they both utilize the 55 nm process.
X-bit labs: What are the primary constraints for GPU performance today, power consumption, memory bandwidth, anything else?
David Cummings: Power, and its byproduct, heat, are probably the two biggest challenges. They were the two biggest reasons behind our move away from simply creating bigger and bigger chips and instead focusing on smaller, highly efficient performance segment GPUs that could easily be adapted to enthusiast, mainstream and value designs. That choice has led to the success we’ve enjoyed with the ATI Radeon HD 3000 series and the HD 4000 series.
ATI Radeon HD 4800 also known as RV770 graphics processor
X-bit labs: ATI was heavily criticized for not implementing hardware FSAA resolve into the R600/RV670 generation processors. But as we see now, Nvidia also decided not to improve render back ends (RBEs) in the GeForce GTX 200 (G200) generation and Intel wants to get rid of hardware RBEs in Larrabee. Perhaps, ATI will do the same eventually and the move on the R600/RV670 was right, but was made too early?
Dave Nalasco: One lesson we learned from our previous generation of GPUs was the level of importance our customers place on maximizing performance at high image quality settings. Specialized hardware is more efficient than general purpose processors for handling tasks like anti-aliasing, and this was a big area of focus for the HD 4000 series design. Besides providing very fast and powerful fixed function MSAA, we also managed to include a high degree of flexibility that allows us to drive image quality higher, with features like Adaptive Anti-Aliasing and Custom Filter Anti-Aliasing (CFAA). DirectX 10.1 support allows us to expose this programmability to developers through an industry standard API.
X-bit labs: What level of computing performance is needed to safely remove fixed-function hardware from GPUs? When would you expect this level of performance to be achieved? Is two, three or four generations the right guess?
Dave Nalasco: In general, I don’t expect this will happen in the foreseeable future. Although we will certainly see GPUs handling a much broader range of applications beyond graphics and video processing over the next few years, they still have to perform these important specialized functions. Well-designed fixed function hardware will continue to provide more speed and efficiency for these tasks than general purpose processors.
X-bit labs: How much time is it spent to create a “shrunk” version of a graphics processor, e.g., transit from 65nm (R600) to 55nm (RV670) with minimal design shifts?
Dave Nalasco: There is no simple answer to that question. It depends on many factors, including the maturity of the new process technology, the amount of tweaking required to achieve performance and power targets, and the level of priority and design resources we assign to the project.