Emulation of the dataflow computing paradigm using field programmable gate arrays (FPGAs)

Building a perfect dataflow computer has been an endeavor of many computer engineers. Ideally, it is a perfect parallel machine with zero overheads, but implementing one has been anything but perfect. While the sequential nature of control flow machines makes them relatively easy to implement, dataflow machines have to address a number of issues that are easily solved in the realm of control flow paradigm. Past implementations of dataflow computers have addressed these issues, such as conditional and reentrant program structures, along with the flow of data, at the processor level, i.e. each processor in the design would handle these issues. The design presented in this thesis solves these issues at the memory level (by using intelligent-memory), separating the processor from dataflow tasks. Specifically, a two-level memory design, along with a pool of processors was prototyped on a group of Altera FPGAs. The first level of memory is an intelligent-memory called Dataflow Memory (DFM), carrying out dataflow tasks. The second level of memory called the Instruction Queue (IQ) is a buffer that queues instructions ready for execution, sent by the DFM. The second level memory has a multiple bank architecture that allows multiple processors from the processor pool to simultaneously execute instructions retrieved from the banks. After executing an instruction, each processor sends the result back to the dataflow memory, where they fire new instructions and send them to the IQ. This thesis shows that implementing dataflow computers at the intelligent-memory level is a viable alternative to implementing them at the processor level