GPU render engines (very important to create a 3D configurator) are becoming more and more popular and offer more and more functions. You`ve certainly thought about integrating GPUs into your workflow. The driving force behind a migration to GPU rendering has always been speed. More and more often the following question is asked: “How much faster is rendering GPUs compared to rendering CPUs?”
The following video illustrates a parallel comparison between a 17 HexaCore processor and the GeForce GTX 970 GPU:
With this article we want to give you a better understanding of the image quality in relation to the render times of different rendering engines with different renderers. We will also show you a way to better compare CPU rendering with GPU rendering.
Benchmarking.
Most likely, you`ve already come into contact with benchmarks for CPU rendering, The concept is simple: set up a benchmark scene and evaluate the render times, the only variable being the CPU model to be tested. The scene file, lighting, shaders and render settings remain constant throughout the test. Benchmarking is comparatively more complex when comparing GPU and CPU rendering. For example, the V-Ray of the Chaos Group has a popular GPU rendering engine (V-Ray RT for GPUs). Although V-Ray RT is very similar to the traditional CPU-based V-Ray Advanced in terms of usability, it is a completely different rendering engine with its own unique settings that delivers slightly different results within a scene.
For this reason, many users will try to achieve the visual image quality of the CPU renderer. But realizing this intention by observation alone is too subjective and unreliable for benchmarking processes. In terms of an image with GPU rendering, it can be said that the results on the CPU look comparatively good. Simultaneously, the render times could be used to compare them with the render time of the CPU. For a meaningful comparison, an empirical approach is required.
Calculate image quality.
Anyone involved in the processing of digital photographs is certainly familiar with the noise in their images. Particularly individual areas of digital photography such as astrophotography suffer from excessive noise. Astronomers only want to work with images of the highest quality when they stack image data. To remove bad images from their data sets, an image quality analysis must be performed for each image.
Ton achieve this, the Department of Ohysics/Astronomy at the University of Manitoba has developed the so-called Image Quality Calculator. The tool, a plug-in for ImageJ, can analyze the quality of 3D renderings. To demonstrate the tool`s capabilities for noise analysis in 3D renderings, V-Rays RT was used to gradually increase the render time for GPUs and to plot the image quality values for 30 rendered images.
The result was a curve one would expect. The amount of measurable noise decreased continously over time. This curve also mimics the observation that the degree of visual improvement is more difficult to estimate with increasing render time.
Calculation of the Image Quality Factor.
The factor is calculated from the sum of the differences between adjacent horizontal and vertical pixels. This is done on three different scales (original size, half size and quarter size). Images with high contrast and sharp edges should have a higher quality factor with this method.
Specify the guideline value for the image quality.
By accurately measuring the image quality between two different engines, we can better judge the performance of GPU and CPU rendering. To do this, we must first determine a guideline for image quality that meets the final requirements of the underlying project. For example, when an animation is created, the image quality benchmark is based on render settings that produce a acceptable noise- and flicker-free animation sequences.
The underlying assumption is that users switch from V-Ray Advanced to V-Ray RT GPU.
Comparison of results.
With the IQ value (and the GPU rendering engine) there is now a target that can be adjusted. In order to compare the CPU with the GPU, it was determined how long the GPU engine took to achieve the image quality of the reference value. In a test, the CPU rendering took exactly 19 minutes and 11 seconds to achieve the benchmark image quality, while the graphics card took 3 minutes and 4 seconds in total. Thus, the graphics processor (2688 CUDA cores) was 6.2 times faster than the CPU.
Choosing the right hardware.
Regardless of whether you rely on CPUs or GPUs in your rendering pipeline. You should be thoroughly informed before making a decision and purchasing hardware.
Have you already decided on a rendering engine? Or are you still thinking? If the latter is the case, we recommend that you read the following article.
Leave A Comment