Streaming at 250 kbps: raise your expectations
summary
Shot-based Dynamic optimizer picks QP and resolution per shot using Netflix’s VMF quality metric and a convex-hull of pre-encodes, delivering 28–37% codec-local bitrate wins and about 55% less bits for the same perceptual quality at mobile rates like 250 kbps. Implementation scales by pruning redundant operating points (1.2 million pre-encodes drops by an order of magnitude), adding checkpoints, using an internal spot market (≈12,000 instances), and encoding 18.5M shots (~1.5M CPU days) with average wall-clock ~7 days per movie.
AI-written summary · about
chapters (14)
- 0:00Start
- 2:02Optimal Video Encoding a Scale
- 4:20Dynamic optimizer framework
- 5:29Convex hull of optimal pre-encodes
- 6:15Find optimal path
- 7:48Dynamic optimizer improvements
- 8:40Parallel Encoding
- 10:27Implementation challenges
- 10:52Number of encodes
- 11:44Number of tasks
- 12:29Internal Spot market
- 18:42Summary
- 19:14Teaser: Codec Comparison
- 19:50Video Algorithms
related talks
17mDelivering better video encodes for legacy devices
27mHighly-Efficient SVT-AV1-based Solutions for VOD Applications
11mFrom Artifacts to Art: A Data-Driven Guide to Android Video Encoding
19mRethinking the Video Transcoding Pipeline: How Pre-Filtering Holds the Key to Efficiency and Quality
31mA Little (video) Goes a Long Way
22mProviding better video experiences for the next billion users.
03mImproving ABR at Netflix
41m8K Codec Low-Latency Livestreaming, and AI Streaming