A GPU-Based Gibbs Sampler for a Unidimensional IRT Model.

Area of Science:

• Psychometrics
• Computational Statistics
• High-Performance Computing

Background:

• Item Response Theory (IRT) models are widely used for large-scale statistical analysis.
• Fully Bayesian estimation of IRT models is computationally intensive and memory-demanding, limiting practical applications.
• Previous attempts using Message Passing Interface (MPI) on distributed clusters faced communication overhead issues due to data dependencies.

Purpose of the Study:

• To address the computational limitations of Bayesian IRT model estimation.
• To explore the use of massive core-based graphic processing units (GPUs) as a practical and cost-effective solution.
• To demonstrate the performance advantages of GPU-based computation over traditional methods.

Main Methods:

• Implementation of IRT model estimation using Compute Unified Device Architecture (CUDA) on GPUs.
• Comparison of computational performance against serial Central Processing Unit (CPU) and MPI-based parallel approaches.
• Evaluation of speedups and efficiency gains offered by the GPU-accelerated method.

Main Results:

• The CUDA GPU approach demonstrated significant performance advantages over serial CPU implementations.
• Compared to MPI, the CUDA GPU method proved more effective in overcoming computational bottlenecks for IRT models.
• The GPU-based approach offers a practical, cost-effective, and convenient alternative for complex IRT analyses.

Conclusions:

• Massive core-based GPUs (CUDA) provide a superior framework for accelerating Bayesian IRT model estimation.
• The findings suggest that GPU computing is a preferred method for handling computationally demanding psychometric analyses.
• This approach enhances the feasibility of applying advanced IRT models in various research and practical settings.