How NVIDIA is Advancing Video Curation with Generative AI
summary
NVIDIA describes an end-to-end video curation service running on DGX Cloud and orchestrated with Ray Data, processing >100 PB of raw video down to tens of PB of curated clips. Pipeline splits, transcodes, filters, deduplicates and captions at GPU scale (initial estimate 1M video-hours/day → ~10,000 GPUs). Optimizations include TensorRT LLM, FP8 quantization and GPU hardware decode/encode, yielding ~7x speedup on the VLM stage and ~3.5x end-to-end throughput with auto-balanced Ray actors; datasets are configurable and iteratively reannotated.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 0:17Video curation challenges
- 1:18Defining data curation
- 5:07Curation pipeline overview
- 11:09Performance optimization
- 17:19Data deduplication
- 21:26Visual understanding
- 23:31Conclusion and Q&A
related talks
26mPetabyte-Scale Video Curation on CoreWeave
40mFueling the AI Revolution with Gaming
14mFrom Reservation to 1,600 GPUs: A Petabyte-Scale Ray Data Pipeline
17mThe State of Generative Media
47mBest Practices for Real-time Intelligent Video Analytics
22mAI Enhanced GPU Video Coding: Achieving Joint High Compression Efficiency and Throughput
21mAnalyst Notebook
22mMovie Gen Video: State-of-the-art Video Generation