PoreMap: Unlocking High-Throughput Raw Nanopore Signal Mapping with Near-Memory Processing

Nanopore sequencing enables long-read, real-time DNA analysis and supports adaptive sampling, where sequencing signals are analyzed during runtime to selectively eject nontarget strands. This capability depends on fast mapping of raw nanopore signals to a reference genome. Raw-signal mapping avoids basecalling overhead and enables earlier sequencing decisions, but the seeding and chaining steps in signal-domain mapping are memory-bound and computationally intensive, especially for large references such as the human genome. In this paper, we present PoreMap, a near-memory processing (NMP) accelerator for high-throughput raw nanopore signal mapping. PoreMap introduces an NMP-friendly seed-and-extend algorithm based on k-d tree vector range search, along with a scalable architecture that exploits the high bandwidth and parallelism of HBM. To efficiently support irregular tree traversal and leaf search, we further design specialized PIM instructions and scheduling strategies that operate within the memory hierarchy. Compared to state-of-the-art accelerators, PoreMap achieves up to 6.8X higher performance and 12X better energy efficiency, while enabling real-time mapping for large reference genomes including the human genome.