Support Vector Machines (SVMs) are supervised learning systems that have gained wide acceptance among the pattern recognition community. Learning is based on structural risk minimization over a training set and leads to a quadratic programming problem. Due to the sample size these optimization problems are very large and training remains one of the most computationally expensive stages in Support Vector Machine design. This paper addresses this problem by exploring different approaches to parallel training. Several algorithms are developed and evaluated including a non-optimal approach for parallel training based on the unbiased version of Piatt's Sequential Minimal Optimization (SMO) algorithm, an improvement to a previous biased non-optimal parallel SMO, and an optimal solution combining SMO with the Chunking approach. Experimental results show that non-optimal solutions can achieve a speed-up of 0(N2), according to the number of processors used, with a compromise in the increment of the number of Support Vectors and a decrement in accuracy. The SMO - Chunking optimal solution presents a much lesser speedup, which depends on the number of support vectors vs. total number of samples ratio.