Parameterized Implementation of K-means Clustering on Reconfigurable Systems

Processing power of pattern classification algorithms on conventional platforms has not been able to keep up with exponentially growing datasets. However, algorithms such as k-means clustering include significant potential parallelism that could be exploited to enhance processing speed on conventional platforms. A better and effective solution to speed-up the algorithm performance is the use of a hardware assist since parallel kernels can be partitioned and concurrently run on hardware as opposed to the sequential software flow. A parameterized hardware implementation of k-means clustering is presented as a proof of concept on the Pilchard Reconfigurable computing system. The hardware implementation is shown to have speedups of about 500 over conventional implementations on a general-purpose processor. A scalability analysis is done to provide a future direction to take the current implementation of 3 classes and scale it to over N classes.