Diffusion Tensor Magnetic Resonance Imaging (DT-MRI) is an emerging technique that explores the structural connectivity of the human brain. The probabilistic fiber tractography based on DT-MRI data behaves more robustly than deterministic approaches in the presence of fiber crossings, but requires more prohibitive computational time. In this work we present a GPU-based probabilistic framework for brain fiber tractography. The framework includes two main steps: 1) Markov-Chain Monte-Carlo (MCMC) sampling, and 2) probabilistic streamlining fiber tracking. We implement the Metropolis-Hastings sampling for local parameter estimation on GPU. In the probabilistic streamlining fiber tracking, we find that fiber lengths are exponentially distributed, and propose a novel segmenting strategy to improve the load balance. On mid-range GPUs, we achieve performance gains up to 34x and 50x over CPUs for the two steps respectively.