Department of Computer Science Colloquium
Thursday, October 10, 2002, 4:15pm 
Upson Hall B17

Why is Graphics Hardware so Fast?

Pat Hanrahan
Stanford University

Recently NVIDIA has claimed that their graphics processors (or GPUs) are improving at a rate three times faster than Moore's Law for processors. A $25 GPU is rated from 50-100 gigaflops and approximately 1 teraop (8-bit ops). Alongside this increase in performance is new functionality. The most recent innovation is user-programmable vertex and fragment stages that allow GPUs to compute a wide range of new visual effects enabling movie-quality games. Announced chips have as many as 200 programmable floating point units operating in parallel. The result is that the latest generation of commodity graphics and game chips are powerful data-parallel computers. Why are these graphics processors so fast? Will the future performance of GPUs continue to increase faster than CPUs? And, if so, what are the implications for computing?