GPU-Accelerated SQL Analytics: How Starburst and NVIDIA Deliver Industry-Benchmark Speedups on GPU Infrastructure

Starburst has integrated NVIDIA's cuDF, an open-source data processing toolkit, into its SQL query engine to enable GPU-accelerated query processing, significantly improving performance without requiring changes to SQL or application code. This integration allows for up to 6x speedup in TPC-H benchmark queries and 3.4x speedup in ClickBench queries, with some individual queries executing over 11 times faster. The enhanced throughput is achieved at a cost comparable to or lower than CPU-only instances, challenging the assumption that GPU infrastructure is more expensive. This development reflects a technical collaboration between Starburst and NVIDIA, aiming to make GPU-accelerated SQL a staple in enterprise data platforms. The integration specifically targets the physical operator level, allowing data to remain on the GPU across multiple execution operators, thereby reducing data transfer overheads and compounding performance benefits. While still in active development, with plans for further optimization and expanded operator coverage, the current implementation demonstrates significant potential for reducing latency in agentic AI systems and improving efficiency in analytics workloads.