Video @Scale 2022 LIVE Q&A #1
summary
Edge compute hosts GPU backends to offload heavy AR ML (GANs, avatar generation) and reduce device work, using geographically distributed edge servers plus compression to cut inter-region bandwidth. End-to-end latency targets land around 200 ms target with anything under 400 ms acceptable (p70 ≈ 400 ms from one-to-one call data); platform measures ~200 ms via 60 ms per-direction P2P render plus buffering. Transport starts with WebRTC then QUIC/RUSH and explores custom protocols; depth is handled by sending tracking data to the backend, rendering composites there and streaming back.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:52AR and video scalability challenges
- 3:15Edge server latency for AR
- 5:03Protocol choices for real-time video
- 7:32Latency causes in live streaming
- 10:01WebRTC vs custom protocols for AR
- 11:06Real-time interaction setup
- 15:43Optimizing AR latency
- 17:52Closing remarks
related talks
23m3 Second End to End Latency At Scale
16mCDN Challenges of HTTP-based Low Latency Live Streaming Delivery
29mThe Cat and Mouse Game of Infra Cost vs. Latency
21mPlayback on Edge - Or how I learned to stop worrying and love serverless.
30mBeyond Cloud Computing: Towards the Edge
41mReal Time Architectures Real Fast
18mScaling APIs to Billions: Extreme-Scale Reliability
37mFrom Hours To Milliseconds: Scaling AI Inference 10x With Serverless on Kubernetes