Documents
Vane Local: 52.7% less time than Ray DataVane Ray: 50.4% less time than Ray Data
Benchmarks
Faster multimodal data processing.
Compare both Vane execution modes across four single-node workloads, with default and tuned batch sizes.
2.80×
Image speedup
Vane Ray · vs Ray Data, tuned batch sizes, single node
1831.16 s → 653.36 sElapsed time reduction: 64.3%
2.12×
Document speedup
Vane Local · vs Ray Data, tuned batch sizes, single node
215.24 s → 101.76 sElapsed time reduction: 52.7%
Compare the results
Test environment: Single node / 36 CPU cores / 64 GB RAM / 2080 Ti (modified VRAM) / 22 GB GPU memory
Tuned batch-size results
Ray 2.58.0Each engine uses tuned batch sizes. The highlights above use this configuration.
Vane LocalVane RayRay DataDaftElapsed time in seconds. Lower is better.
Each workload uses its own linear scale starting at zero.
Images
Vane Local: 63.9% less time than Ray DataVane Ray: 64.3% less time than Ray Data
Audio
Vane Local: 1.0% more time than Ray DataVane Ray: 3.2% more time than Ray Data
Video
Vane Local: 1.4% less time than Ray DataVane Ray: 1.5% more time than Ray Data
36 CPU / 64 GB RAM / 2080 Ti, 22 GB modified VRAM. Single node. Ray 2.58.0.
Exact elapsed times in seconds. Percentages compare each Vane mode with Ray Data in the selected configuration.
| Workload | Vane Local | Vane Ray | Ray Data | Daft |
|---|---|---|---|---|
| Document | 101.76 s52.7% less time vs Ray | 106.81 s50.4% less time vs Ray | 215.24 s | 413.00 s |
| Image | 661.20 s63.9% less time vs Ray | 653.36 s64.3% less time vs Ray | 1831.16 s | OOM |
| Audio | 2132.67 s1.0% more time vs Ray | 2177.74 s3.2% more time vs Ray | 2110.56 s | OOM |
| Video | 6938.71 s1.4% less time vs Ray | 7141.23 s1.5% more time vs Ray | 7035.15 s | 8322.00 s |
About this benchmark
Ray Data and Vane Ray use Ray 2.58.0. Vane Local runs locally; Vane Ray uses the Ray execution backend. Both modes are measured on the same single-node setup.
Ray Data has the lowest audio elapsed time in both configurations. Video results are close across Ray Data and the two Vane modes.
These results measure pipeline elapsed time; they do not measure model accuracy, context quality or token savings.
The retest records default and tuned configurations; exact per-engine batch sizes are not listed here.
This single-node evaluation follows the workload design of the Ray Data multimodal AI benchmark. Results apply to the documented single-node hardware, data and settings. Daft results are retained from the previous run and were not retested; OOM indicates out of memory in that run.
Benchmark code and workflow