DuckDB Benchmarks Its v2.0 Preview on NVIDIA Vera, Reporting a 46% TPC-H SF1,000 Edge Over an Intel Xeon 6 Baseline
DuckDB's team ran a v2.0 development build on NVIDIA's Vera CPU and an Intel Xeon 6 server, reporting median TPC-H SF1,000 scores of 3,073,292.99 versus 2,109,782.19 QphH@Size.
Overview
The DuckDB team published a benchmark of a development build of DuckDB v2.0 running on NVIDIA’s Vera CPU on October 16, 2026. According to the DuckDB blog, written by Gábor Szárnyas, the Vera platform showed “a 46% performance advantage on TPC-H SF1,000 over the x86 baseline (Intel Xeon 6).” The post says the work was done in collaboration with the NVIDIA team, and it frames the numbers as the team’s own runs rather than audited results.
What the Test Covered
The post says the team ran “the full TPC-H benchmark (not just the read queries!) on scale factor 1,000” on two server platforms: an Intel Xeon 6 baseline and the NVIDIA Vera. Per the DuckDB blog, both machines ran Ubuntu 26.04 and had 768 GB of memory and a single NVMe SSD with 3+ TB of storage formatted as ext4.
Because DuckDB v2.0 was not yet released at the time of writing, the team used a preview build, version 1.6.0.dev379, with the duckdb-tpch benchmark harness. The post says the team set block_allocator_memory to ‘500G’ and enabled allocator_background_threads to get the best performance.
The two systems were configured differently:
- Intel Xeon 6 baseline: The DuckDB blog says the CPU had 96 cores with hyper-threading, yielding 192 threads, but that DuckDB performed best with 96 threads, so the team set
threadsto 96. - NVIDIA Vera: The post says the chip packs 88 cores with Spatial Multithreading. The team used a dual-socket system but restricted DuckDB to a single socket with the
numactlcommand, and setthreadsto 88.
According to the post, Vera runs arm64 binaries, so it ran DuckDB’s regular build without modifications. NVIDIA’s product page likewise describes the CPU as having 88 custom Olympus cores. The DuckDB post also cites up to 1.2 terabytes per second (TB/s) of memory bandwidth.
Results
The team performed a warmup run, discarded its result, then ran three complete runs in sequence and took the median of the QphH@Size composite scores, according to the DuckDB blog.
| Run | x86 QphH@Size | Vera QphH@Size |
|---|---|---|
| Run 1 | 2,120,588.82 | 3,073,292.99 |
| Run 2 | 2,022,795.01 | 3,131,306.27 |
| Run 3 | 2,109,782.19 | 3,052,102.87 |
| Median | 2,109,782.19 | 3,073,292.99 |
The post states that if the results were audited through the official TPC-H channels, “both would stand a chance to be in the top 3 among the TPC-H Top Performance Results on SF1,000.” The team attributes Vera’s advantage to three factors: high memory bandwidth enabling fast scans, low memory latency keeping cores occupied during processing, and the Olympus cores’ large caches allowing efficient access to the hash tables built during aggregations and joins.
What We Don’t Know
- The figures are unaudited results produced in collaboration with the hardware vendor; the post does not report a price or power comparison between the two systems.
- The test used a single socket of a dual-socket Vera system. The team writes that it is looking forward to continuing the benchmark with DuckDB v2.0.0’s stable build, “running on the full dual-socket setup and incorporating different benchmarks.”
- The post does not give a release date for DuckDB v2.0.
Context
This follows an earlier v2.0 performance preview from the project. As previously reported, the DuckDB 1.5.6 announcement included a Windows test in which v2.0.0-dev finished a TPC-H SF300 run in 129 seconds versus 822 seconds for v1.5.6.