Endeavor: New PairHMM method speeds DNA variant detection 2x
A novel parallelization strategy cuts genome analysis bottleneck by up to 2.14x on CPUs.
Endeavor, introduced by Miguel Graça and Aleksandar Ilic, presents a breakthrough in DNA variant calling by redefining the PairHMM algorithm's parallelization. Traditionally, the PairHMM step is the primary bottleneck in large-scale genomic pipelines, consuming up to 70% of execution time. Existing CPU and GPU approaches fail to scale to long DNA sequences due to limited anti-diagonal data parallelism. Endeavor instead exploits row-level fine-grained parallelism without sacrificing accuracy, enabling efficient processing of sequences up to 100k basepairs for the first time.
Benchmarking on Intel/AMD CPUs shows Endeavor delivering up to 2.14x peak throughput improvement over GKL and at least 2x speedup over GATK HaplotypeCaller on real-world datasets. On the GPU side, NVIDIA and AMD implementations achieve up to 2.05x speedups on genome-scale data compared to prior state-of-the-art GPU methods. The work, published at HPDC '26, provides a portable, SIMD-based solution that dramatically accelerates variant detection, a critical step for understanding genetic diseases and genomic structure.
- Endeavor introduces row-level fine-grained parallelism for PairHMM, handling sequences up to 100k basepairs.
- On CPUs, it outperforms GKL by up to 2.14x and GATK HaplotypeCaller by at least 2x.
- NVIDIA and AMD GPUs achieve up to 2.05x speedups over state-of-the-art GPU methods.
Why It Matters
Faster DNA variant calling accelerates genomic research and clinical diagnostics, enabling scalable analysis of massive datasets.