Skip to content- Sources: primary
- Summary: The GenDB-V paper claims 22x to 61x speedups over Umbra and DuckDB from an AI-generated, formally verified engine, but Neumann calculates that at SF10 memory bandwidth alone puts a floor of about 280 ms under Umbra's roughly 540 ms total. The agents built per-query auxiliary structures at load time, such as suppliers sorted by s_acctbal for Q2, and the plain-array mode hard-codes dbgen's row numbering to make the Q18 join cheap. The verified fallback for data that breaks those assumptions is a slow nested loop that never runs with official TPC-H data.
- Why it matters: Treat AI-optimized database benchmark results with suspicion, and check whether a speedup survives inserts, deletes or data not produced by the benchmark's generator.
- Background: TPC-H is a decision-support benchmark of 22 analytical queries over data produced by a generator called dbgen at a chosen scale factor. Its predecessor TPC-D was broken by vendors who precomputed answers, and TPC-H rules forbid that kind of tuning, but unofficial results often ignore the rules. Umbra is a compiled-query research engine from Neumann's group.