Workshop: IA^3 2026 — 16th Workshop on Irregular Applications: Architectures and Algorithms
Authors: Souvadra Hati (Georgia Tech), Benjamin Glick (NVIDIA Corporation), and Richard Vuduc (Georgia Tech)
Abstract: The problem of k-mer counting (KC) is to quantify the frequency of length-k substrings in a biological sequence. It is an essential and often limiting step in bioinformatics workloads. Despite the recent popularity of GPUs, the fastest distributed-memory KC solutions remain CPU-based. Existing GPU implementations have poor memory-bandwidth utilization and weak communication-computation overlap. To address these issues, we present Squid, a GPU-native distributed-memory, bulk-synchronous parallel k-mer counter. Squid introduces a transpose-based k-mer generation kernel that reaches up to 71% of peak high-bandwidth memory bandwidth and is 6× faster than prior approaches. On identical hardware, Squid runs 22× faster end-to-end than the previous state-of-the-art GPU counter, and scales to 256 GPUs on a modern GB200-NVL72 cluster at ≥85% intra-rack parallel efficiency. Squid is also the first GPU k-mer counter to outperform the CPU-based DAKC and HySortK, running on 1024 CPUs of Perlmutter, by up to 2× and 6.6×, respectively.
Back to IA^3 2026 — 16th Workshop on Irregular Applications: Architectures and Algorithms Archive Listing
Back to Full Workshop Archive Listing