Petabyte-Scale Video Curation on CoreWeave
summary
Pipeline curates 600 TB of video on CoreWeave using Anyscale managed Ray + Ray Data on CKS, decoding with FFmpeg, sampling frames, and running VLLM inference on NVIDIA RTX Pro 6000 / V40 GPUs. Ray Data reduces GPU time to about 75% versus Ray Core; the cluster (256 nodes, 1,600 GPUs) produces 70 million captions in 95 minutes while Chaos + Lota sustain 110–120 GB/s reads, integrate NVMe caching, and eliminate thundering‑herd and cross‑region egress costs.
AI-written summary · about
chapters (12)
- 0:00Introduction and overview
- 1:30Video captioning with VLLM
- 3:35Challenges in petabyte-scale processing
- 4:57CoreWeave AI Object Storage
- 6:43Video curation pipeline workflow
- 8:06Ray Core vs Ray Data comparison
- 9:28Scaling to 600 terabytes
- 10:11Storage performance impact
- 11:59LOTA cache and thundering herd
- 13:03Design principles for efficiency
- 14:07Collaboration and results
- 14:51Q&A session
related talks
34mHow NVIDIA is Advancing Video Curation with Generative AI
35mHow Bytedance Builds Large-Scale Data Processing Pipelines for Multimodal Models with Ray
38mNetflix's Ray Platform: From Deep Learning to GenAI
33mArchitecting Scalable Multi-Modal Video Search
31mHow Vivix Scales Video Ad Classification with Ray
32mAnyscale's Ray Data: Revolutionizing Batch Inference
30mHow to Scale Unstructured Data Processing with Ray Data
28mScaling Multimodal Data and Reinforcement Learning with Ray