Hardware implementation of non-bonded forces in molecular dynamics simulations

Molecular Dynamics is a computational method based on classical mechanics to describe the behavior of a molecular system. This method is used in biomolecular simulations, which are intended to contribute to the study and advance of nanotechnology, medicine, chemistry and biology. Software implementa...

Full description

Autores:
Caicedo Beltrán, Álvaro José
Tipo de recurso:
Trabajo de grado de pregrado
Fecha de publicación:
2011
Institución:
Universidad del Valle
Repositorio:
Repositorio Digital Univalle
Idioma:
eng
OAI Identifier:
oai:bibliotecadigital.univalle.edu.co:10893/16237
Acceso en línea:
https://hdl.handle.net/10893/16237
Palabra clave:
Modelos computacionales
Procesadores digitales
Diseño electrónico
Dinámica molecular
Simulación dinámica
Rights
openAccess
License
http://purl.org/coar/access_right/c_abf2
Description
Summary:Molecular Dynamics is a computational method based on classical mechanics to describe the behavior of a molecular system. This method is used in biomolecular simulations, which are intended to contribute to the study and advance of nanotechnology, medicine, chemistry and biology. Software implementations of Molecular Dynamics simulations can spend most of time computing the non-bonded interactions. This work presents the design and implementation of an FPGA-based coprocessor that accelerates MD simulations by computing in parallel the non-bonded interactions, specifically, the van der Waals and the electrostatic interactions. These interactions are modeled as the Lennard-Jones 6-12 potential and the direct-space Ewald summation, respectively. In addition, this work introduces a novel variable transformation of the potential energy functions, and a novel interpolation method with pseudo-floating-point representation to compute the short-range forces. Also, it uses a combination of fixed-point and floating-point arithmetic to obtain the best of both representations. The FPGA coprocessor is a memory-mapped system connected to a host by PCI Express, and is provided with interruption capabilities to improve parallelization. Its main block is based on a single functional pipeline, and is connected via Avalon Bus to other peripherals such as the PCIe Hard-IP and the SG-DMA. It is implemented on an Altera¿s EP2AGX125EF35C4 device, can process 16k particles, and is configured to store up to 16 different types of particles. Simulations in a custom C-application for MD that only computes non-bonded forces become up to 12.5x faster using the FPGA coprocessor when considering 12500 atoms.