Date of Award
Master of Science in Electrical Engineering - (M.S.)
Electrical and Computer Engineering
Alexandros V. Gerbessiotis
This work presents a practical implementation of a uni-processor system design. This design, named D2-CPU, satisfies the pure data-driven paradigm, which is a radical alternative to the conventional von Neumann paradigm and exploits the instruction-level parallelism to its full extent. The D2-CPU uses the natural flow of the program, dataflow, by minimizing redundant instructions like fetch, store, and write back. This leads to a design with the better performance, lower power consumption and efficient use of the on-chip resources. This extraordinary performance is the result of a simple, pipelined and superscalar architecture with a very wide data bus and a completely out of order execution of instructions. This creates a program counter less, distributed controlled system design with the realization of intelligent memories. Upon the availability of data, the instructions advance further in the memory hierarchy and ultimately to the execution units by themselves, instead of having the CPU fetch the required instructions from the memory as in controlled flow processors. This application (data) oriented execution process is in contrast to application ignorant CPUs in conventional machines. The D2-CPU solves current architectural challenges and puts into practice a pure data-driven microprocessor. This work employs an FPGA implementation of the D2-CPU to prove the practicability of the data-driven computer paradigm using configurable logic. A relative analysis at the end confirms its superiority in performance, resource utilization and ease of programming over conventional CPUs.
Sathe, Anish Arvind, "Configurable computer systems can support dataflow computing" (2004). Theses. 530.