Scalable and accurate multi-GPU-based image reconstruction of large-scale ptychography data.

Area of Science:

• Materials Science
• Biophysics
• Computational Imaging

Background:

• Advances in synchrotron light sources enable nanoscale ptychographic imaging.
• Large datasets from ptychography experiments strain computing resources.
• Existing multi-GPU approaches lack optimization for heterogeneous intranode interconnects.

Purpose of the Study:

• To develop an optimized intranode multi-GPU implementation for large-scale ptychographic reconstruction.
• To address performance bottlenecks in the conjugate gradient (CG) solver for maximum likelihood reconstruction.
• To create a scalable and accurate software tool for ptychographic data analysis.

Main Methods:

• Developed a novel hybrid parallelization model for multi-GPU communication.
• Implemented the model in a tool named PtyGer (Ptychographic GPU(multiple)-based reconstruction).
• Focused on maximum likelihood reconstruction using a conjugate gradient (CG) method.

Main Results:

• PtyGer demonstrates efficient handling of terabyte-scale ptychography datasets.
• The hybrid parallelization model optimizes communication across heterogeneous GPU interconnects (PCIe, NVLink).
• Achieved outstanding intranode GPU scalability without compromising reconstruction accuracy.

Conclusions:

• PtyGer provides an efficient solution for large-scale ptychographic reconstruction problems.
• The optimized intranode multi-GPU implementation significantly enhances computational performance.
• The tool offers a scalable and accurate approach for advanced imaging applications.