We develop a parallel-tempered feature allocation algorithm to infer tumor heterogeneity from deep DNA sequencing data. The feature allocation model is based on a binomial likelihood and an Indian Buffet process prior on the latent haplotypes. A variation of parallel tempering technique is introduced to flatten peaked local modes of the posterior distribution and yields a more efficient Markov chain Monte Carlo algorithm. Simulation studies pro- vide empirical evidence that the proposed method is superior to competing methods at a high read depth. In our application to Glioblastoma multiforme data, we found several distinctive haplotypes that indicate the presence of multiple subclones in the tumor sample.