gpuPairHMM: High-Speed Pair-HMM Forward Algorithm for DNA Variant Calling on GPUs.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Increasing DNA sequence data volume necessitates faster core algorithms.
• Pairwise sequence alignment using Pair-HMMs is crucial for DNA variant calling but computationally intensive.
• Existing GPU acceleration methods for Pair-HMMs suffer from inefficient memory access.

Purpose of the Study:

• To develop a significantly faster GPU-based solution for Pair-HMM computations.
• To address the limitations of previous GPU acceleration approaches for sequence alignment.

Main Methods:

• Introduced gpuPairHMM, a novel GPU parallelization scheme for the Pair-HMM forward algorithm.
• Utilized wavefronts and warp-shuffles to minimize memory accesses and instructions.
• Implemented on multiple generations of CUDA-enabled GPUs (Volta, Ampere, Ada, Hopper, Blackwell).

Main Results:

• Achieved close-to-peak performance on various modern GPUs.
• Outperformed prior GPU implementations by at least 11.7x.
• Demonstrated superior performance over CPU and FPGA implementations by 14.2x and 19.8x, respectively.

Conclusions:

• gpuPairHMM offers a substantial speedup for Pair-HMM based sequence alignment.
• The novel parallelization scheme effectively overcomes memory access limitations in GPU computing.
• This advancement significantly enhances the efficiency of DNA variant calling tools.