I think the GPU could be a great fit for OLAP, but when it comes to the nasty OLTP use cases the CPU will absolutely dominate.
Strictly serialized transaction processing facilities demand extremely low latency compute to achieve meaningful throughput. When the behavior of transaction B depends on transaction A being fully resolved, there are no magic tricks you can play anymore.
Consider that talking to L1 is at least 1,000x faster than talking to the GPU. Unless you can get a shitload of work done with each CPU-GPU message (and it is usually the case that you can), this penalty is horrifyingly crippling.