Speaker
Description
Modern genomic sequencing generates massive datasets requiring efficient sequence alignment algorithms. We investigate the Graphcore Intelligence Processing Unit (IPU), an AI-optimized accelerator, for the Needleman-Wunsch sequence alignment algorithm. Through comprehensive benchmarking on GC200 IPU hardware with 1,472 tiles and 918 MB on-chip memory, we demonstrate peak performance of 36 GCUPS for optimal diagonal computations—2.4× faster than NVIDIA A30 GPU baseline (15 GCUPS). However, memory constraints limit practical applications to sequences under 9,400 nucleotides. We identify critical memory-performance trade-offs: fine-grained parallelism achieves 1.25 GCUPS but handles only 2,000-length sequences, while coarse-grained approaches process 22× larger sequences at reduced performance (0.25 GCUPS). Memory profiling reveals 13.8× overhead from compiler-generated code. We provide practical guidelines for integrating IPUs into genomic pipelines, identifying suitable applications (database searches, quality control) and unsuitable scenarios (chromosome-scale alignment). Our findings demonstrate IPUs complement rather than replace GPUs in bioinformatics infrastructure.