My GPU Database Lost to a Single CPU Thread. The Bug Was One Constant, 128x Too Small

This is a submission for DEV's Summer Bug Smash: Smash Stories powered by Sentry.

I maintain a GPU SQL engine — a DuckDB community extension that runs aggregates and GROUP BY on Apple Silicon via Metal. Last night I ran its own benchmark and got a result that should not be possible for a GPU database:

gpudb-groupby-bench rows=10000000 groups=1024

[CPU] single-threaded std::unordered_map median wall = 14.65 ms

[Metal] Apple M4 Max median wall = 25.39 ms

My GPU Database Lost to a Single CPU Thread. The Bug Was One Constant, 128x Too Small

Related reading

GPU Incident at 3am: eBPF Tracing from Page to Root Cause in 60 Seconds

An AMD GPU Beat My Mac on Llama 8B. The Same GPU Lost on Phi-3.

Nothing Crashed. Nothing Errored. My CPU Sat at 390% for an Hour.

[Aug 17-22] We Swapped GPUs, Then Kept Finding Blind Spots in Our Own Checks

Running 35B–400B LLMs on a GPU-less Cluster to Mine 10,000 Papers — and the 4…

Solving the GPU Pinning Saga and Gemma's Meta-Commentary