Jason T. Chittenden, Robert H. Leary, and Brett Matzuka
Pharsight Corporation
Objectives: Recent advances in the general computing capabilities of readily available graphics processing units (GPUs) make them appropriate tools for the solution of nonlinear mixed effects problems. The architecture of the GPU presents challenges for their general use for population pharmacokinetic/pharmacodynamic (popPK/PD) applications. The solution of ordinary differential equation (ODE) models can cause the execution paths of the concurrently executing threads to diverge, destroying the computational efficiency and obliterating the benefits of the GPU. Thread divergence is aggravated by adjustable step-size ODE solvers and different event times (observations, doses, time-lags, reflux, etc…) between subjects. In this work we demonstrate an implementation that circumvents these issues and lays the groundwork for professional application of GPUs to popPK/PD problems.
Methods: A novel version of a stiff ODE solver was implemented as a GPU kernel. The entire process of generating individual predictions is run on the GPU, with the CPU handling the parameter estimation process with a Quasi-Random Parametric Expectation Maximization (QRPEM) algorithm. An OpenMP implementation of the same ODE solver on the CPU was used for comparison.
The test problem was a single dose, intravenous bolus, saturating clearance, simulated dataset. The problem was run with various numbers of samples to test the scaling of the GPU and CPU implementations.
Results: The GPU implementation was between 5 and 20 times faster than the CPU implementation on a single core. The CPU and GPU both exhibited linear scaling with number of subjects for large problems, but the GPU showed linear scaling with small problems.
Conclusions: Computing the objective function for popPK/PD problems on the GPU is a feasible and possibly worthwhile opportunity. For particularly large and/or computationally intensive problems the cost of custom coding could result in tremendous time savings. Commercially viable GPU solutions for popPK/PD are needed to make these benefits generally accessible to the pharmacometrics community.
References: Available on request.
Reference: PAGE 22 (2013) Abstr 2801 [www.page-meeting.org/?abstract=2801]
Poster: Estimation methods