Power Efficiency of Ray Tracing

Master's thesis work carried out by Alexander Toresson during the winter of 2012-2013 at Intel Sweden AB

Download now

Abstract: This thesis compares the energy efficiency of state-of-the-art CPU and GPU solutions for ray tracing; GPU-based OptiX and CUDA RT and CPU-based Embree, and investigates the reasons behind the differences. The comparison is done by measuring the energy in kilojoules that needs to be spent to attain a specific level of image quality. Recently, there have been successful efforts to leverage modern programmable GPUs for ray tracing. However, in professional so-called render farms, CPUs are still dominant.

Path tracing is a popular method based on ray tracing to produce digital images with correct global illumination. The main drawback with this approach is that it is generally very computationally expensive to produce an image. It is often necessary to trace hundreds or even thousands of rays per pixel to produce a good image.

In motion picture productions, studios have big render farms to accelerate rendering. In these settings, power and cooling comes with a significant cost, and thus power efficiency is of utmost importance. We show that the relative energy efficiency between the CPU and GPU solutions examined are within a factor of two, where CPU-based Embree is on par with GPU-based OptiX.

Keywords: energy, Joule, CPU, Intel, GPU, NVIDIA, CUDA, path tracing, Embree, OptiX

Errata: None so far

This page was last updated 2013-05-06.