sirius

Sirius makes DuckDB run your SQL on an NVIDIA GPU. Load it as an extension and your existing queries get intercepted and executed on the graphics card with no rewrites, and anything it can't handle falls back to the CPU. It's Apache 2.0, built mostly by NVIDIA engineers with UW-Madison, and nobody sells a paid version.

The hardware gate is real: Linux, an NVIDIA Turing-or-newer GPU, recent CUDA drivers, and a kernel and filesystem that allow io_uring and direct I/O. No Macs, no AMD, no Windows. There are no packaged releases either. You build from source with Pixi, and your DuckDB version has to match the one Sirius was built against exactly.

This is for data teams already living in DuckDB who own or rent NVIDIA GPUs and want big analytical queries to finish faster. Everyone else should stay on plain DuckDB (duckdb/duckdb). For GPU dataframes instead of SQL, look at cuDF. DuckDB is the only database it supports today; StarRocks and Doris are roadmap items.

The catch: the headline ClickBench records came from the old engine, which has since been rewritten. And the CPU fallback is silent, so a query with window functions or nested types quietly runs on the CPU and you won't know you lost the speedup.