How do we turn GPU back to the centralized model? Viewed that GPUs all receive the same string as a parameter, we cannot transmit different information on which data to process to each GPU.
A natural workaround would be to implement a web server with a .php script that sends different data each time. To avoid all GPUs accessing the web server at the same time, plugins should distribute connection attempts with exponential waiting. The entire Gnutella mechanism becomes a trigger to start the centralized computation.