NMF-mGPU: non-negative matrix factorization on multi-GPU systems.

Area of Science:

• Bioinformatics
• Computational Biology
• High-Performance Computing

Background:

• Non-negative Matrix Factorization (NMF) is crucial for extracting patterns from high-dimensional bioinformatics datasets.
• Traditional NMF implementations face computational time challenges with large datasets, even on parallel systems.
• Graphics Processing Units (GPUs) offer powerful parallel processing capabilities for computational tasks.

Purpose of the Study:

• To develop an efficient and user-friendly implementation of NMF leveraging GPU computing power.
• To address the memory limitations of GPUs in NMF algorithms.
• To accelerate the analysis of large-scale biological data.

Main Methods:

• NMF-mGPU utilizes CUDA for GPU computation, optimizing for various CUDA architectures.
• Implements blockwise data transfer for GPUs with limited memory.
• Supports multi-GPU synchronization using Message Passing Interface (MPI) for distributed systems.

Main Results:

• NMF-mGPU achieves significant speedups, approximately 120x faster than single-core processors on a four-GPU system.
• Demonstrates over 4x speedup compared to single-GPU devices, indicating super-linear performance.
• Efficiently processes large matrices by managing GPU memory constraints.

Conclusions:

• GPUs provide a cost-effective and high-performance alternative to traditional clusters in bioinformatics.
• NMF-mGPU facilitates the extraction of biological insights from massive experimental datasets.
• The tool is accessible to researchers with minimal GPU programming expertise and available on GitHub.