GPUs (Graphical Processing Units) have recently been recognized for their impressive computational potential. Taking advantage of this mostly untapped resource requires strong collaboration between computational scientists and domain scientists. A team of programmers at Nvidia (Dr. Scott Legrand) and the San Diego Supercomputing Center (Dr. Ross Walker) has ported and optimized PMEMD for GPUs. The resulting program, CUDA-enabled pmemd, was used to run molecular dynamics simulations on the GM2 activator protein (GM2-AP).
The protein GM2-AP is responsible for extracting the GM2 ganglioside from lipid bilayers and preparing them for hydrolysis by the HexA protein. The breakdown of this process results in a potentially lethal GM2 buildup in nervous tissue commonly leading to Tay-Sachs disease. Crystallographic studies and electron paramagnetic resonance (EPR) spectroscopy studies suggest conflicting predictions of the GM2-AP mechanism. Through NIH funding, the Fanucci group in the Department of Chemistry has been studying this protein using experimental methods. The Roitberg group – also in the chemistry department and Quantum Theory Project – has conducted computational molecular dynamics studies to help the interpretation of the experimental data. The molecular dynamics runs need to be as long as possible to properly sample the space of possible configurations. The Roitberg group develops and uses the program Amber and PMEMD, which are fast, scalable molecular dynamics engines geared towards biomolecules.
Molecular dynamics simulations carried out on this protein using both CUDA-enabled pmemd and its traditional CPU counterpart clearly show the advantages that can be gained by using GPUs for the computations. Using the Tesla machines installed at the UF HPC, running a two-thread simulation with traditional CPUs on this protein yielded a computational efficiency of 0.7 ns per day (that is, a 0.7 ns of a molecular dynamics trajectory was propagated in 24 hours of simulation time). Running the same simulation on a single Nvidia Tesla machine yielded a comparable length of 26.9 ns per day. This represents about an 80-fold speed-up comparing a single GPU to a single CPU.
While GPU calculations are still in their infancy, their potential in the future of this field is uncontested.