LumiBench: A Benchmark Suite for Hardware Ray Tracing

Abstract

Ray tracing as a graphics rendering method is becoming increasingly popular in real-time applications, supported by dedicated accelerator cores in the latest generation GPUs. However, the high computational intensity of the ray tracing algorithm still limits the visual effects that can be produced while maintaining a high frame rate. To improve ray tracing hardware, it is important to understand the underlying characteristics of ray tracing workloads and identify performance bottlenecks. In this paper, we present LumiBench, the first benchmark suite for evaluating ray tracing hardware performance in modern GPUs designed to execute on the Vulkan-Sim GPU simulator. LumiBench features a diverse set of scenes and shaders that are representative of real applications but simple enough to be simulated in a reasonable amount of time. We first evaluate LumiBench against Rodinia to highlight the difference between ray tracing versus general purpose workloads and demonstrate the need for a dedicated benchmark suite. Then, we characterize the workloads included in LumiBench, which are organized into several clusters targeting different aspects of the ray tracing pipeline, and provide insights for future architectural research.