The queries are extremely simple (in appendix after references) and, in the absence of indexes, require independent evaluation of all of the rows in the table.

The queries are extremely simple (in appendix after references) and, in the absence of indexes, require independent evaluation of all of the rows in the table. This maps extremely well to the CUDA architecture which has hundreds of memory-coupled independent processing elements. But this isn't anywhere near the typical usage scenario of SQL databases. If it was, scaling databases would be trivial by sharding tables across many servers. In real life, you have indexes and joins, which translate into random access patterns, which do not perform very well on CUDA.