Next-generation acceleration and code optimization for light transport in turbid media using GPUs

Area of Science:

• Biomedical Optics
• Computational Physics
• Medical Imaging

Background:

• Monte Carlo (MC) simulations are crucial for accurate light transport modeling in biological tissues.
• Widespread use of MC for inverse problems like photodynamic therapy (PDT) planning is hindered by long computation times.
• Optimizing MC code for Graphics Processing Units (GPUs) presents challenges due to memory access bottlenecks.

Purpose of the Study:

• To develop a highly optimized Monte Carlo code package for simulating light transport on GPUs.
• To overcome performance limitations in GPU-accelerated MC simulations.
• To enhance the feasibility of MC simulations for complex biomedical applications.

Main Methods:

• Developed an optimized Monte Carlo (MC) code package, GPU-MCML, for NVIDIA Fermi GPUs.
• Implemented an optimization scheme utilizing fast shared memory to mitigate global memory access bottlenecks.
• Applied various optimization techniques to harness full GPU potential for light transport simulations.
• Tested the code on a 7-layer skin model and a four-GPU cluster.

Main Results:

• Achieved a ~600x acceleration of the MCML code on a Fermi GPU compared to a high-end CPU.
• Demonstrated linear performance improvement with an increasing number of GPUs in a cluster setup.
• The GPU-MCML package is open-source and available in optimized and simplified versions.

Conclusions:

• The developed GPU-MCML code significantly accelerates light transport simulations.
• GPU acceleration and optimization strategies make complex MC simulations more practical for biomedical research and applications.
• The open-source release promotes wider adoption and further development of GPU-accelerated MC methods.