Rongchai Wang
Dec 15, 2025 18:16
NVIDIA’s Sirius, a GPU-accelerated SQL engine for DuckDB, units new efficiency benchmarks on ClickBench, showcasing cost-efficiency and pace via modern GPU-native execution.
NVIDIA, in collaboration with the College of Wisconsin-Madison, has launched Sirius, a GPU-accelerated SQL engine that has set a brand new efficiency benchmark on ClickBench. The open-source Sirius engine integrates with DuckDB, a database famend for its simplicity and pace, to allow high-performance analytics by leveraging GPU know-how.
Partnership and Innovation
The adoption of DuckDB has surged amongst firms like DeepSeek, Microsoft, and Databricks, due to its versatile and environment friendly nature. Recognizing the potential for enhanced efficiency, NVIDIA and the College of Wisconsin-Madison developed Sirius to offer GPU-accelerated analytics with out the necessity to reconstruct database techniques from the bottom up. By using NVIDIA’s CUDA-X libraries, Sirius accelerates question execution, providing important enhancements in efficiency and throughput in comparison with conventional CPU-based techniques.
Architectural Highlights
Sirius operates as a GPU-native execution backend for DuckDB, requiring no modifications to DuckDB’s codebase. It leverages NVIDIA’s high-performance libraries, together with cuDF and RAPIDS Reminiscence Supervisor, to construct its execution engine. This integration permits Sirius to reuse DuckDB’s superior subsystems whereas enhancing them with GPU acceleration, facilitating environment friendly execution of SQL operations.
File-Setting Efficiency
On the ClickBench analytics benchmark, Sirius demonstrated record-breaking efficiency. Operating on NVIDIA’s GH200 Grace Hopper Superchip, Sirius outperformed different high techniques, attaining superior cost-efficiency and pace. In assessments, Sirius delivered not less than 7.2 occasions greater cost-efficiency than CPU-only techniques, underscoring its functionality to deal with complicated queries with ease.
Future Developments
Trying ahead, NVIDIA and its companions intention to advance GPU knowledge processing capabilities. Efforts will give attention to enhancing GPU reminiscence administration, growing GPU-native file readers, and evolving the execution mannequin right into a scalable, multi-node structure. These developments intention to streamline knowledge processing and lengthen the scalability of Sirius to deal with petabyte-scale datasets.
For extra detailed insights, go to the unique [source name](https://developer.nvidia.com/weblog/nvidia-gpu-accelerated-sirius-achieves-record-setting-clickbench-record/).
Picture supply: Shutterstock
