New GPU codec decodes genomic data at 260GB/s with instant random access

Genomic archives grow faster than decompression can keep up—the European Nucleotide Archive holds tens of petabytes, and tools like gzip are fundamentally sequential. Existing GPU decompressors (e.g., nvCOMP DEFLATE at ~50GB/s) decode whole files with no random access, while CPU tools like samtools support region seeks but at CPU speed. Yakiv Shavidze's new paper extends ACEAPEX, an absolute-offset parallel LZ77 codec, with a full device-resident GPU decode pipeline that handles both entropy and match resolution entirely on the GPU. On FASTQ data, this pipeline reaches up to 260GB/s—closing the gap from earlier work that only accelerated the match phase.

The paper also introduces position-invariant random access: a compact coordinate index enables decoding an arbitrary read in just 0.362ms, roughly 6x faster than a warm samtools faidx lookup, with an index 6.3x smaller than a .fai file. A range-decode strategy decouples output size from VRAM, sustaining 165.7GB/s on a 50GB genome where whole-file decode would run out of memory. The author also benchmarks Meta's open DietGPU ANS on H100 at 592GB/s decode, suggesting a fully open high-throughput stack is viable. All code is MIT-licensed.