Rongchai Wang
Dec 15, 2025 18:16
NVIDIA’s Sirius, a GPU-accelerated SQL engine for DuckDB, sets new performance benchmarks on ClickBench, showcasing cost-efficiency and speed through innovative GPU-native execution.
NVIDIA, in collaboration with the University of Wisconsin-Madison, has introduced Sirius, a GPU-accelerated SQL engine that has set a new performance benchmark on ClickBench. The open-source Sirius engine integrates with DuckDB, a database renowned for its simplicity and speed, to enable high-performance analytics by leveraging GPU technology.
Partnership and Innovation
The adoption of DuckDB has surged among companies like DeepSeek, Microsoft, and Databricks, thanks to its versatile and efficient nature. Recognizing the potential for enhanced performance, NVIDIA and the University of Wisconsin-Madison developed Sirius to provide GPU-accelerated analytics without the need to reconstruct database systems from the ground up. By utilizing NVIDIA’s CUDA-X libraries, Sirius accelerates query execution, offering significant improvements in performance and throughput compared to traditional CPU-based systems.
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, including cuDF and RAPIDS Memory Manager, to build its execution engine. This integration allows Sirius to reuse DuckDB’s advanced subsystems while enhancing them with GPU acceleration, facilitating efficient execution of SQL operations.
Record-Setting Performance
On the ClickBench analytics benchmark, Sirius demonstrated record-breaking performance. Running on NVIDIA’s GH200 Grace Hopper Superchip, Sirius outperformed other top systems, achieving superior cost-efficiency and speed. In tests, Sirius delivered at least 7.2 times higher cost-efficiency than CPU-only systems, underscoring its capability to handle complex queries with ease.
Future Developments
Looking forward, NVIDIA and its partners aim to advance GPU data processing capabilities. Efforts will focus on enhancing GPU memory management, developing GPU-native file readers, and evolving the execution model into a scalable, multi-node architecture. These advancements aim to streamline data processing and extend the scalability of Sirius to handle petabyte-scale datasets.
For more detailed insights, visit the original [source name](https://developer.nvidia.com/blog/nvidia-gpu-accelerated-sirius-achieves-record-setting-clickbench-record/).
Image source: Shutterstock
