Rethinking Analytical Processing in the GPU Era (arxiv.org)

The GPU era is reshaping data analytics, and recent hardware and software advances have eliminated many traditional barriers to GPU adoption in this field. A new open-source project, Sirius, exemplifies this shift by serving as a GPU-native SQL engine that can seamlessly accelerate diverse data systems. By treating the GPU as the main processing unit and leveraging high-performance libraries like libcudf, Sirius enables significant speedups without altering the user-facing interface, thanks to its use of the Substrait query representation standard. Sirius’s design allows it to integrate effortlessly with existing databases, delivering substantial performance improvements at comparable hardware costs. For example, when paired with DuckDB, it achieves a 7x speedup on the TPC-H benchmark in a single-node setup, and this boosts up to 12.5x faster when combined with Apache Doris in a distributed environment. These results highlight the growing practicality and efficiency of GPU-native analytical processing, marking a pivotal step toward broader adoption of GPU-accelerated database engines in real-world applications, benefiting the AI/ML community that relies heavily on fast, scalable data querying.