Designing coprocessors for hybrid compute systems
Volker Hampel, Peter Sobe, Erik Maehle · Proceedings - IEEE International Parallel and Distributed Processing Symposium · 2008
A hybrid compute system (HCS) combines standard CPUs and reconfigurable devices, usually FPGAs, in one system. These systems have become more attractive again, due to a closer and hence faster coupling of both computational components. From our work with several designs for the same application, we have found the communication between a CPU and a FPGA-based coprocessor to relate either to pipelining or to a bulk-wise transfer with buffered data processing. We identify conditions which determine whether the pipelined or the buffered style should be used in a design. A Reed/Solomon encoding coprocessor has been implemented for each of the communication architectures to serve as an example of how these conditions materialize and how they influence the performance.