Benchmarks

Single-node multimodal pipeline benchmarks

The benchmark workflow is here.

Primary result

Tuned batch-size results

Test environment: single node · 36 CPU cores · 64 GB RAM · 2080 Ti (modified VRAM) · 22 GB GPU memory

Elapsed time (seconds)
Tuned batch sizes · log scale · lower is better
Vane DataRay DataDaft
86.09127413
Document
1.15k1.77kOOM
Image
2.31k2.36kOOM
Audio
7.60k6.92k8.32k
Video

This single-node evaluation adapts the workload design from the Ray Data multimodal AI benchmark. Daft's OOMs on the image and audio workloads may reflect the test machine's limited memory, while the older 2080 Ti GPU also limits throughput. Given current hardware and compute-budget constraints, we have not reproduced the original benchmark's full cluster-scale setup; these results apply only to the recorded single-node environment.

WorkloadVane DataRay DataDaftvs Ray Datavs Daft
Document86.09 sBatch size: 2560127 sBatch size: 320413 sBatch size: 2032.2% lower79.2% lower
Image1147 sBatch size: 1001767.11 sBatch size: 100OOM35.1% lowerOOM
Audio2312 sBatch size: 1282363.08 sBatch size: 128OOM2.2% lowerOOM
Video7603 sBatch size: 326922 sBatch size: 328322 sBatch size: Not set9.8% higher8.6% lower
Reference

Default batch-size results

Results from the same recorded single-node environment before per-engine batch-size tuning.

Elapsed time (seconds)
Default batch size · log scale · lower is better
Vane DataRay DataDaft
187209732.99
Document
1.19k2.04kOOM
Image
2.64k2.88kOOM
Audio
7.60k6.92k8.32k
Video

This single-node evaluation adapts the workload design from the Ray Data multimodal AI benchmark. Daft's OOMs on the image and audio workloads may reflect the test machine's limited memory, while the older 2080 Ti GPU also limits throughput. Given current hardware and compute-budget constraints, we have not reproduced the original benchmark's full cluster-scale setup; these results apply only to the recorded single-node environment.

Build your first AI pipeline on multimodal data.