DuckDB Posts Unaudited Vera TPC-H SF1000 Numbers

DuckDB published on 16 October 2026 (Gábor Szárnyas) a blog post benchmarking a DuckDB v2.0-dev preview build on NVIDIA Vera, using the full TPC-H workload at scale factor 1,000. The work was done in collaboration with NVIDIA. The numbers below are DuckDB’s own, unaudited figures from that post — not an official TPC-audited result.

Hardware and build (as DuckDB describes them)

DuckDB says NVIDIA Vera comes with 88 “Olympus” cores, Spatial Multithreading, and up to 1.2 terabytes per second (TB/s) of memory bandwidth, and that Vera runs arm64 binaries so DuckDB’s regular build runs unmodified.

Both the Vera machine and an x86 baseline ran Ubuntu 26.04, with 768 GB of memory and a single NVMe SSD (3+ TB, ext4). Because DuckDB v2.0 was not released yet, the team used preview build 1.6.0.dev379.

Important limits called out by DuckDB:

The Vera run was on a dual-socket system but DuckDB was restricted to a single socket with numactl, and threads = 88
The x86 baseline was an Intel Xeon 6 configured with threads = 96 (DuckDB says the CPU had 96 cores with hyper-threading / 192 threads, but it performed best at 96 threads)

So this is a single-socket Vera comparison against a 96-thread Xeon 6 baseline, on a preview build — not a dual-socket audited TPC submission.

Exact QphH@Size scores DuckDB reports

DuckDB discarded a warmup run, then ran three complete TPC-H runs and took the median composite QphH@Size scores:

	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 score	2,109,782.19	3,073,292.99

DuckDB’s own summary: “the Vera platform has a 46% performance advantage on TPC-H SF1,000 over the x86 baseline (Intel Xeon 6).” The post also says that if these results were audited through official TPC-H channels, “both would stand a chance to be in the top 3” among TPC-H Top Performance Results on SF1,000. That ranking claim is DuckDB’s hypothetical — these results are not audited.

DuckDB attributes Vera’s advantage to three factors in its own words: high memory bandwidth for fast scans; low memory latency keeping cores busy; and large Olympus-core caches for hash tables in aggregations and joins.

Tuning flags DuckDB used

For this large, concurrent SF1,000 throughput run, DuckDB says it set:

SET block_allocator_memory = '500G';
SET allocator_background_threads = true;

The harness was duckdb-tpch at commit cf096cea8a0a054cefe0e695c842ca5bca6ec04d, with dataset generation at SF=1000, per the published setup script.

Why it matters

NVIDIA has been pushing Vera (and Vera Rubin rack designs) as CPU/memory bandwidth for agentic and data-heavy nodes. An open analytics engine publishing a full TPC-H SF1,000 pass — refresh streams included, not only read queries — on a preview build is useful colour for that hardware story, provided readers keep the unaudited / single-socket / preview caveats.

TSN previously covered DuckDB’s unreleased v2.0 Agent Mode work and NVIDIA’s Vera Rubin OCP rack disclosures; this post is the database-on-Vera microbenchmark thread, not a product GA for DuckDB 2.0.

What’s still unknown

When DuckDB v2.0.0 stable ships and whether scores move on that build
Dual-socket Vera numbers (DuckDB says it wants to run the full dual-socket setup next)
Whether anyone submits an audited TPC-H SF1,000 result on Vera

What to watch

Watch DuckDB’s follow-up for stable v2.0.0 and dual-socket runs, and watch official TPC-H Top Performance Results if NVIDIA or a partner later audits a submission.

What this means

DuckDB’s own unaudited benchmark, done with NVIDIA’s collaboration.
Preview build 1.6.0.dev379 on a single Vera socket (88 threads) vs Xeon 6 at 96 threads.
Median QphH@Size: Vera 3,073,292.99 vs x86 2,109,782.19 — DuckDB’s stated 46% advantage.