Comparison of Single and Double Floating Point Precision Performance for Tesla Architecture GPUs

Abstract

The paper compares the single and double floating-point precision performance of NVIDIA Tesla GPUs. Double precision is crucial for the accuracy of some applications containing scientific computation. Three representative applications have been chosen in order to compare the performances: matrix multiplication, incompressible Navier-Stokes equations, and the steady-state heat conduction problem. The expected values of performance decrease lie between two for bandwidth-limited applications and eight for compute-limited applications. Based on the conducted experiments, the type of each application has been identified and the decreases in performance have confirmed the expected theoretical values.