GPU visual art at work: Propagation of errors on a GPU
By serendipity this visual computing experiment led to a visualization of colorful structures that evolve from the parallel propagation of errors on a graphics processing unit (GPU). Authors wanted to transfer color values in subsequent volume slices using an OpenCL kernel. Due to concurrent memory access this transfer operation was erroneous and produced interesting visual patterns in a point-based visualization. Meanwhile they fixed this “error” with another kernel but we wanted to share the initial results as a piece of visual computing art. The visual patterns reveal the structure of block-wise kernel execution on the GPU. This real-time application is running on a NVIDIA GTX 580 on a volume of 256³ points.
Conclusion
We described the setup of this visual computing experiment and gave an explanation of the e ects which occurred by serendipity as a result of erroneous operations. The visualization of the propagation of the error revealed the typical block-wise structure of kernel executions on GPUs. As a side effect, we can watch the GPU working to produce a volume of millions of errors by propagation and recombination. You can watch this experiment running in real time on video above.
Detailed description of this experiment and relevant source code snippets:
http://visualcompute.noumentalia.de/media/2012/06/gpuerror.pdf
For further information: http://www.visualcompute.com/
Category: Computer Science, Video
Subscribe
If you enjoyed this article, subscribe to receive more just like it.
