NMF-mGPU

NMF-mGPU implements non-negative matrix factorization (NMF) on NVIDIA GPUs to accelerate extraction of interpretable components from high-dimensional biological datasets.

Key Features:

• Non-negative Matrix Factorization (NMF): Implements the NMF algorithm for extracting interpretable components from data matrices.
• CUDA acceleration: Executes computations using CUDA (Compute Unified Device Architecture) to perform GPU-based linear-algebra operations.
• Blockwise memory management: Manages limited on-board GPU memory by blockwise transferring large input matrices from system main memory to GPU memory.
• Multi-GPU synchronization: Supports synchronization across multiple GPUs using MPI (Message Passing Interface) for parallel execution.
• CUDA architecture optimization: Optimized for various CUDA architectures to utilize different NVIDIA GPU generations.
• Performance scaling: Demonstrates approximately 120 times faster processing on a four-GPU setup compared to a single conventional processor and more than four times the speed of a standalone GPU device.

Scientific Applications:

• Dimensionality reduction: Extraction of interpretable components from high-dimensional biological datasets.
• Large-scale experimental data analysis: Processing of extensive biological experimental datasets to support downstream biological interpretation.
• High-performance bioinformatics computing: Accelerated matrix factorization for computationally intensive bioinformatics tasks.

Methodology:

Implements NMF via CUDA for GPU linear-algebra operations, performs blockwise transfers of large input matrices from system main memory to GPU memory, synchronizes multiple GPUs using MPI, and is optimized for various CUDA architectures.