NVIDIA Vera
We ran the full TPC-H benchmark with the scale factor 1,000 dataset using the duckdb-tpch project.
Our setup included two server platforms: an x86 baseline (Intel Xeon 6) and NVIDIA Vera.
We also compared two DuckDB versions: the latest stable version (v1.5.5) and the upcoming release's alpha version (v2.0.0-alpha), with the latter shipping several optimizations that make complex workloads such as TPC-H faster.
Platforms
x86 Baseline: Intel Xeon 6
We ran the benchmark on a server with an Intel Xeon 6 CPU.
- CPU: Intel Xeon 6 CPU, 96 cores with hyper-threading
- Memory: 768 GB
- Disk: 1 NVMe SSD with 3 TB+ storage, formatted to ext4
- Operating system: Ubuntu 26.04
DuckDB flags:
SET threads = 96;
SET block_allocator_memory = '500G';
SET allocator_background_threads = true;
NVIDIA Vera
We conducted the benchmark on a dual-socket system but restricted DuckDB to a single socket.
- CPU: NVIDIA Vera CPU, 88 cores with Spatial Multithreading
- Memory: 768 GB
- Disk: 1 NVMe SSD with 3 TB+ storage, formatted to ext4
- Operating system: Ubuntu 26.04
DuckDB flags:
SET threads = 88;
SET block_allocator_memory = '500G';
SET allocator_background_threads = true;
DuckDB Versions
We ran DuckDB using the Python client and benchmarked two DuckDB versions:
- DuckDB v1.5.5
- DuckDB v2.0.0-alpha (Python package:
1.6.0.dev379)
Benchmark Methodology
We performed an initial warmup run and discarded its result, then conducted three runs and took the median value of the QphH@SF composite scores.
Benchmark Results
| Setup 1 | Setup 2 | Setup 3 | Setup 4 | |
|---|---|---|---|---|
| CPU | Intel | Intel | Vera | Vera |
| DuckDB | 1.5.5 | 2.0.0-alpha | 1.5.5 | 2.0.0-alpha |
| Run 1 | 1,643,610.87 | 2,120,588.82 | 2,488,209.48 | 3,073,292.99 |
| Run 2 | 1,700,471.78 | 2,022,795.01 | 2,389,726.36 | 3,131,306.27 |
| Run 3 | 1,673,422.48 | 2,109,782.19 | 2,482,589.73 | 3,052,102.87 |
| Median score | 1,673,422.48 | 2,109,782.19 | 2,482,589.73 | 3,073,292.99 |
The results show two key findings. First, the Vera platform has a 46–48% performance advantage on TPC-H SF1,000 over the x86 baseline (Xeon 6). Second, DuckDB v2.0.0-alpha shows a 24–26% improvement over DuckDB v1.5.5 on both CPUs – and this is something all users will benefit from once DuckDB v2.0.0 is out.
We'll share more details in an upcoming blog post. Stay tuned!