Read mapping — aligning sequenced DNA fragments to a reference genome — is a major bottleneck in genome analysis because of expensive dynamic-programming alignment, and existing filters handle the dominant paired-end reads poorly by evaluating each read of a pair independently. GenPairX is a hardware-algorithm co-designed accelerator that introduces a filter jointly considering both reads in a pair for far higher filtering effectiveness, plus a lightweight alignment algorithm that replaces most costly dynamic programming, backed by two specialized hardware mechanisms. It achieves orders of magnitude higher throughput-per-watt than leading CPU-based mappers and 1.43× over the best accelerator-based mapper, with no loss of accuracy.

Published at IEEE HPCA 2026.