Eight fewer cores and 46 percent more throughput

DuckDB has published a benchmark of NVIDIA's new Vera processor, run with the NVIDIA team, using the full TPC-H workload at scale factor 1,000 rather than the read queries alone. Vera is aimed at agentic workloads and carries 88 Olympus cores, Spatial Multithreading and up to 1.2 terabytes per second of memory bandwidth. Because it runs arm64 binaries, DuckDB's regular build worked on it unmodified.

The baseline was an Intel Xeon 6 with 96 cores and hyper-threading for 192 threads, though DuckDB performed best with threads set to 96. Vera was measured on a single socket of a dual-socket system, restricted with numactl and pinned to 88 threads. Both machines ran Ubuntu 26.04 with 768 GB of memory and one NVMe SSD of 3 TB or more, formatted ext4. Version 2.0 had not been released, so the team used a preview build, 1.6.0.dev379.

After a warm-up run that was discarded, three complete runs were taken and the median of the QphH composite scores recorded. The x86 median was 2,109,782.19 and Vera's was 3,073,292.99, which the authors state as a 46 percent advantage for Vera on this workload. They add that were the results audited through the official TPC-H channels, both machines would stand a chance of placing in the top three for scale factor 1,000.

Three reasons are offered for the gap. The high memory bandwidth makes scans fast. The low memory latency keeps the cores continuously occupied. And the large caches on the Olympus cores allow efficient access to the hash tables built during aggregations and joins. The team says it intends to repeat the exercise on the stable 2.0.0 build, on the full dual-socket machine, and with other benchmarks.