The Kohonen Feature Maps are commonly employed to process large input data but their effective working abilities can be achieved only after a time-consuming process of learning. Performed tests have shown that the sequential program, solving a typical problem, uses more than 95 percent of its time to localize the winners. The aim of the paper is to present and compare different ways of the algorithm parallelization. We compare two different classes of parallel implementations - the network parallelization and the learning set parallelization. During performed experiments two different ways of experimental evaluation are used: standard evaluation based on such metrics as speedup and efficiency and the approximation method based on the granularity