Humanity is currently witnessing a biodiversity decline. In order to understand the causes and thereby finding means to prevent and possibly reverse the effects, we need to be able to measure the trends of species populations at scale. Traditional monitoring is labour-intense, but new alternative avenues are offered by environmental DNA (eDNA) methods. By sampling and sequencing media such as the air around us, we can in theory monitor the entirety of life on earth - as all living organisms shed DNA into the environment. However, the sequenced fragments of eDNA (reads) need to be classified to their source organisms before any analysis and inference can be made. There are different ways of classifying. At grand scales with billions of reads, each having to be attributed to one of millions of possible source species, there is in practice only one option – k-mer based classification. This involves making ultrafast comparisons of shorter words (k-mers) within each read to reference k-mers, obtaining hits to a specific taxon for each k-mer, and making a joint call to classify each read from the results. The de facto standard software for this type of classification is Kraken 2 [1]. It is a CPU and RAM-intense algorithm that has been effective for when reference datasets were smaller. In the last years we have seen an increasing flow of new reference genomes. A year ago, we constructed a Kraken 2 reference database from one representative genome for all species that had one (or more). The total reference dataset was 14 TB and would have required a node with 9 TB of RAM to construct using the default settings of Kraken 2. Since then, the number of Eukaryotic species with a reference genome has increased 25%. Eukaryotic genomes are on average much bigger than prokaryotic genomes and in practice govern the size of the database. Kraken 2 has lately introduced an algorithm where the database can be partitioned, so that nodes with less memory can be used. However, the speed of database construction and subsequent classification of our dataset (~45 TB) is still a concern. Recently, an alternative to Kraken 2 called MetaCache-GPU [2] has emerged, which has a GPU-accelerated algorithm for database construction and classification. The authors claim vastly improved speeds. We have whole analysis pipelines tailored to Kraken 2 as the read classifier, and we know little about MetaCache-GPU and what it outputs and how it could fit into our research. We’re applying for this small GPU-project at Dardel to evaluate MetaCache-GPU for our use. [1] Wood, D. E., Lu, J., & Langmead, B. (2019). Improved metagenomic analysis with Kraken 2. Genome biology, 20(1), 257. https://doi.org/10.1186/s13059-019-1891-0 [2] Kobus, R., Müller, A., Jünger, D., Hundt, C., & Schmidt, B. (2021). MetaCache-GPU: Ultra-fast metagenomic classification. arXiv preprint arXiv:2106.08150. https://doi.org/10.48550/arxiv.2106.08150