NVIDIA’s Sirius Sets New ClickBench Performance Records with GPU Acceleration
NVIDIA has unveiled Sirius, a next-generation GPU-accelerated SQL execution engine that has achieved record-breaking results on the ClickBench benchmark. Developed in partnership with the University of Wisconsin–Madison, Sirius integrates seamlessly with DuckDB and demonstrates how GPU-native analytics can dramatically improve both speed and cost efficiency for modern data workloads.
A Powerful Collaboration Driving Innovation
DuckDB has rapidly gained popularity among organizations such as Microsoft, Databricks, and DeepSeek, thanks to its lightweight architecture and high performance. Building on this momentum, NVIDIA and the University of Wisconsin–Madison set out to enhance DuckDB’s analytical capabilities without redesigning its core architecture.
The result is Sirius — an open-source GPU-accelerated execution engine that uses NVIDIA CUDA-X libraries to offload computationally intensive SQL operations to GPUs. This approach allows organizations to unlock massive performance gains while continuing to use familiar DuckDB workflows.
GPU-Native Architecture Without Disruption
One of Sirius’s most notable advantages is its non-intrusive design. It functions as a GPU-native execution backend for DuckDB and does not require changes to DuckDB’s existing codebase.
Sirius relies on proven NVIDIA technologies such as cuDF and RAPIDS Memory Manager, enabling efficient processing of large datasets directly on the GPU. By reusing DuckDB’s query planner and storage components, Sirius combines DuckDB’s maturity with the parallel processing power of GPUs.
Record-Breaking ClickBench Results
Sirius delivered exceptional results on the ClickBench analytics benchmark, a widely recognized test for analytical database performance. When executed on NVIDIA’s GH200 Grace Hopper Superchip, Sirius surpassed all previously reported systems.
Key highlights include:
Significantly faster query execution times
At least 7.2× higher cost efficiency compared to CPU-only solutions
Consistent performance across complex analytical queries
These results demonstrate that GPU-accelerated SQL processing can outperform traditional CPU-based systems not only in speed, but also in operational efficiency.
What’s Next for Sirius
NVIDIA plans to further expand Sirius’s capabilities by:
Improving GPU memory management
Introducing GPU-native data readers
Transitioning toward a distributed, multi-node execution model
These enhancements are designed to support petabyte-scale analytics and make Sirius suitable for even the most demanding enterprise workloads.
Conclusion
Sirius represents a major step forward in GPU-accelerated analytics. By combining DuckDB’s simplicity with NVIDIA’s GPU ecosystem, Sirius proves that high-performance SQL processing can be achieved without sacrificing usability or requiring a complete system redesign.
As GPU data processing continues to evolve, Sirius sets a new benchmark for what modern analytical engines can deliver.
