Latency
8 ms versus 3.8 seconds
This result refers to one analytical database query run against the same data in an internal comparison. It is not an end‑to‑end application, auction, incident, or model response time.
Benchmark notes
Extraordinary results require visible context.
The performance figures on this site come from internal Oliver testing and early technical evaluations. They are directional, workload‑dependent, and not independently audited. We publish them with qualifiers because no benchmark result applies equally to every data shape, query, or deployment.
- up to 162×CPU performance observed in evaluated analytical query shapes
- up to 1000×GPU performance observed in evaluated analytical query shapes
- 50–300×lower compute use observed, depending on workload and configuration
How to read the claims
Same query and data. Different workloads, different results.
Shape
Hard queries widen the gap
Performance varies with joins, grouping, selectivity, cardinality, schema irregularity, data temperature, and concurrency. Oliver is designed to perform especially well on difficult analytical shapes.
Hardware
CPU and GPU are separate results
The CPU and GPU figures should not be read as one universal speedup. Hardware, memory, storage, caching, and deployment topology all materially affect the result.
Efficiency
Compute reduction is workload dependent
Lower compute and storage use comes from columnar compression, elastic stateless cells, ingest‑time resolution, and avoiding repeated agent work. The contribution of each factor varies.
Technical evaluation
Test Oliver on your workload.
During a technical evaluation, we review the dataset, query suite, hardware, concurrency, cache state, ingest path, and measurement window with your team. Reproducible methodology and raw results are shared for the workload being evaluated.