Design space exploration of streaming multiprocessor architectures

In this paper, we present a comparison of two design-space exploration approaches. The comparison is in terms of (1) speed of simulation versus accuracy of performance numbers, and (2) connection to trajectories for detailed design. The two approaches are: the trace driven approach and the control data flow graph approach. The first approach leads to the shortest simulation time, but is insufficiently accurate in the performance numbers it provides. It also does not connect well to a trajectory for detailed design. The second method is leading to rather long simulation times, yet it can give fairly accurate performance numbers, and it produces results that can be readily taken as input for further design. The two approaches are somehow extreme in that several in-between methods can be conceived of. We also describe our search for an exploration trajectory which would be somehow "optimal" in terms of speed versus accuracy and closeness to a design trajectory. As expected, this trajectory appears somewhere in-between the two extremes mentioned above.