Revisiting sorting for GPGPU stream architectures

Abstract

This poster presents efficient strategies for sorting large sequences of fixed-length keys (and values) using GPGPU stream processors. Compared to the state-of-the-art, our radix sorting methods exhibit speedup of at least 2x for all generations of NVIDIA GPGPUs, and up to 3.7x for current GT200-based models. Our implementations demonstrate sorting rates of 482 million key-value pairs per second, and 550 million keys per second (32-bit). For this domain of sorting problems, we believe our sorting primitive to be the fastest available for any fully-programmable microarchitecture.