Skip to content

stable (8d17afe)

Pre-release
Pre-release

Choose a tag to compare

@mike-wendt mike-wendt released this 02 Oct 20:43
8d17afe

Getting started

A quick path from nothing installed to a GPU-accelerated query. Full details, requirements, and the Python API:
docs/README.md.

Requirements: Linux (amd64 or arm64), NVIDIA GPU Turing or newer (compute capability 7.5+), and either CUDA 12.x
(driver >=525.60.13) or CUDA 13.x (driver >=580.65.06) matching the file you download below.

1. Install DuckDB v1.5.6

Sirius extensions must match the exact DuckDB version they were built against:

curl https://install.duckdb.org | DUCKDB_VERSION=1.5.6 sh

2. Download the Sirius extension for your platform

3. Decompress it

Downloads are gzip-compressed to fit under GitHub's release-asset size limit:

gunzip <downloaded-file>.duckdb_extension.gz

4. Allow unsigned extensions

Sirius isn't published through DuckDB's official signed extension repository yet, so you need to explicitly allow
unsigned extensions before loading it:

duckdb -unsigned

For client APIs, pass the same config at connect time instead, e.g. Python:
duckdb.connect(config={"allow_unsigned_extensions": "true"}).

Warning: unsigned extensions execute native code with the host process's privileges.

5. Load the extension and try it

LOAD './sirius-linux_amd64-cuda12.duckdb_extension';

-- Generate TPC-H data on the fly (scale factor 1 = ~1 GB) and query it, on GPU,
-- automatically. No query rewrites needed; unsupported operators fall back to CPU.
INSTALL tpch;
LOAD tpch;
CALL dbgen(sf=1);

SELECT l_returnflag, sum(l_quantity)
FROM lineitem
GROUP BY l_returnflag
ORDER BY l_returnflag;

About Sirius

Sirius is a GPU-accelerated execution engine for DuckDB. This stable release is a maintainer-promoted snapshot, not a
versioned release, it will be replaced the next time a maintainer promotes a new commit. File issues at
https://github.com/sirius-db/sirius/issues or join the discussion on Slack.