Abstract : This article presents a multi-GPU adaptation of a specific Monte Carlo and classification based method for pricing American basket options, due to Picazo. The first part relates how to combine fine and coarse-grained parallelization to price American basket options. A dynamic strategy of kernel calibration is proposed. Doing so, our implementation on a reasonable size (18) GPU cluster achieves the pricing of a high dimensional (40) option in less than one hour against almost 8 as observed for runs we conducted in the past, using a 64-core cluster (composed of quad-core AMD Opteron 2356). In order to benefit from different GPU device types, we detail the dynamic strategy we have used to load balance GPU calculus which greatly improves the overall pricing time we obtained. An analysis of possible bottleneck effects demonstrates that there is a sequential bottleneck due to the training phase that relies upon the AdaBoost classification method, which prevents the implementation to be fully scalable, and so prevents to envision further decreasing pricing time down to handful of minutes. For this we propose to consider using Random Forests classification method: it is naturally dividable over a cluster, and available like AdaBoost as a black box from the popular Weka machine learning library. However our experimental tests will show that its use is costly.