1: \begin{abstract}

2: In this paper, we present a GPU-accelerated direct-sum boundary integral method

3: to solve the linear Poisson-Boltzmann (PB)  equation.

4: In our method, a well-posed boundary integral formulation

5: is used to ensure the fast convergence of Krylov subspace based linear algebraic solver such as the GMRES.

6: The molecular surfaces are discretized with flat triangles and centroid collocation.

7: To speed up our method, we take advantage of the parallel nature of the boundary integral formulation and parallelize the schemes within CUDA shared memory architecture on GPU.

8: The schemes use only $11N+6N_c$ size-of-double device memory

9:  for a biomolecule with $N$ triangular surface elements and $N_c$ partial charges.

10: Numerical tests of these schemes show well-maintained  accuracy and fast convergence.

11: The GPU implementation using one GPU card (Nvidia Tesla M2070) achieves 120-150X speed-up to the implementation using one CPU (Intel L5640 2.27GHz). With our approach, solving PB equations on well-discretized molecular surfaces with up to 300,000 boundary elements will take less than about 10 minutes, hence our approach is particularly suitable for fast electrostatics computations on small to medium biomolecules.

12: 

13: \vspace*{1cm}

14: 

15: 

16: 

17: {\it Keywords: Poisson-Boltzmann, electrostatics, boundary integral, parallel computing, graphic processing units (GPU)}~

18: 

19: \end{abstract}

20: