Notes from building a GPU database.
Engineering notes from the RayoDB team: why analytics is changing hardware, and what building for it teaches.
AI functions belong inside SQL
Forecasting and anomaly detection usually live in a pipeline next to the database. On a GPU-native engine they are table functions in the query — which changes who gets to use them.
What "GPU-native" actually means
Storage layout, residency, kernel fusion, one GPU per worker: the design choices that separate a database built for the GPU from a CPU engine with an accelerator — and their trade-offs.
Why we built an analytics database on GPUs
The CPU era of analytics has a ceiling, and we hit it running video analytics at NPAW. This is the story of how an internal engine became RayoDB.