The computational power of GPUs, coupled with increasing programmability, is making the GPU a compelling platform for high-performance computing. GPUs excel at regular, structured computation, but irregular computation -- where processors consume an irregular, runtime-dependent amount of input or produce an irregular, runtime-dependent amount of output -- is a challenging problem in a parallel computing environment. These characteristics are common in today's real-time graphics pipelines but are typically handled in hardware. We anticipate these problems will become more relevant as we move toward the next generation of graphics systems that have at their core *programmable real-time graphics pipelines* and must instead support these workloads in more general-purpose ways. Fortunately, the research projects that are beginning to address these challenges can also address other complex, demanding application domains.