high performance computing on graphics processing units: hgpu.org

The Jacobian evaluation problem is ubiquitous throughout scientiOc computing. In this article, the possibility of massively parallel computing of Jacobian matrix is discussed. It is shown that the computation of the Jacobian matrix shares the same parallelism with the computation being differentiated, which suggests that once we know how to parallelize a computation, its Jacobian computation can be parallelized in a similar manner. A simple problem is used to demonstrate the potential of massively parallel Jacobian computation with GPU. Significant speedup is observed.