Why the next step is the GPU.
Analytics spent fifteen years squeezing the processor. The next order of magnitude comes from different hardware.
The argument in four steps.
The processor road has ended.
Columnar engines got faster for fifteen years by feeding a few processor cores as efficiently as physics allows. That road has stopped climbing.
Memory bandwidth is the wall.
An analytical query is a data-movement problem: read, filter, group, join. A server processor moves a few hundred gigabytes per second, shared by every core.
A GPU matches the work.
Thousands of cores and memory that moves about a terabyte per second, applied to counting, filtering, grouping and joining. The data stays on the card.
Why it did not happen before.
GPU databases proved the idea a decade ago and stayed proprietary and expensive. The cards are now everywhere and the tooling has matured.
What RayoDB does about it.
The data lives on the card.
Storage is arranged so the columns a query needs sit in the GPU’s own memory. The card is where the data belongs, not a cache.
Capacity is a count.
Each worker owns one GPU. More headroom means more workers; the coordinator splits the question and merges the answers.
Models run next to the data.
Forecasting, anomaly detection and search by meaning run as SQL functions on the same cards that hold the data.
One session, two statements.
1 · A week of events, per hour
SELECT toStartOfHour(ts) AS hour, count() AS events
FROM product_events
WHERE ts >= now() - INTERVAL 7 DAY
GROUP BY hour;| hour | events |
|---|---|
| 2026-08-27 00:00 | 1,284,913 |
| 2026-08-27 01:00 | 962,340 |
| … | … |
168 rows · sample data
2 · The next 24 hours, forecast in SQL
SELECT * FROM AI_FORECAST(hourly_events, hour, events, 24);| hour | forecast | low | high |
|---|---|---|---|
| 2026-09-03 00:00 | 1,301,540 | 1,214,002 | 1,388,118 |
| 2026-09-03 01:00 | 978,116 | 901,455 | 1,054,210 |
| … | … | … | … |
24 rows · sample data
Real RayoDB SQL, illustrative rows. No timings are shown, by policy.
See what the GPU changes on your data.
Early access is open to a small group of design partners with real workloads.