Implementation of a median filtering algorithm using the RADram system illustrates the performance gain achievable due to the parallelism available in image-processing algorithms. The above graph shows the performance of RADram execution wall-time as more reconfigurable functional units are operated in parallel compared to that of a normal processor/memory system. Note that the time to run just the median algorithm decreases as more functional units are simultaneously executed. However, the wall-time remains the same. This shows that on the RADram system, the time to setup the computation exceeds the processing time. This performance is obtained only through a customized application circuit (the sorting of nine values) made possible by reconfigurable logic.
The performance of RADram while executing a Dynamic Programming algorithm is indicative of potential performance gains in a broad class of integer problems. A largest common sub-sequence algorithm was implemented on RADram and compared to a normal processor/memory implementation. As shown in the graph above, execution time for the normal version increases as O(n2) as expected. However, the execution time of the RADram system grows as O(n1.2). Clearly, as the problem size increases, so does the amount of RAD memory which must be used, and hence a greater amount of parallelism is achieved. This phenomenon is unique to RADram. Other parallel systems do not automatically increase processor resources as the problem size increases.