Databases on GPGPU are still in a very early stage of development. I would even say that, unless it is top secret, it is rather only a field for academic research at the moment – and that is where we can find the most information about it. Recently I got back to looking for something about applications of CUDA or OpenCL to PostgreSQL. I found a rather general, but interesting, presentation by Krzysztof Kaczmarski from Warsaw University of Technology, in which he compares GPU and CPU in OLAP Cubes Creation. It was presented at some conference in 2011, so since then a lot have changed. One of the things that struck me was the fact that a GPU was in some cases twice slower than a CPU. The stated reason for that was a low PCI-e bandwidth, which at the computer used in tests had throughput of 3GB/s. If you compare it with the current state of art PCI-e 3.0 x 16 cards, it is 5 times slower, for details confront Wikipedia.org. Hence, maybe rerunning the tests now on the most advanced graphic card would change the results a bit.
The final topic in this post is a kind of an announcement. If anybody has a bit more time you can read about the workshop on GPUs in Databases, which will take place on the 8th of September 2015 in Poitiers, France.