ASIC-RDX and Compute-Compression Efficiency Optimization for UGC Video Processing
summary
Facebook deploys a hybrid ASIC–software RDX pipeline that runs an ASIC in L1 and software or ASIC encoders in L2, achieving production gains: prior RDX showed up to 15% gain over state-of-the-art, and the ASIC-first hybrid convex hull sits only 2% compression gap vs. software-only while cutting encoder compute and power dramatically. A linear-programming recipe optimizer picks per-video codec/preset mixes using watch-time distribution and device codec support to maximize global compression under compute constraints.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:10UGC video encoding efficiency
- 2:34RDX compute-compression
- 6:27Joint compute-compression optimization
- 14:27Q&A with Shankar & Srinath
- 23:40Closing remarks
related talks
20mBoosting Video Codec With RISC-V Vector Extension
22mAI Enhanced GPU Video Coding: Achieving Joint High Compression Efficiency and Throughput
15mUsing The Power of Edge Compute To Enhance Streaming Delivery
10mAdventures with VP9
11mHow Do We Compress Giant AI Models to Run at the Edge?
13mLow-level wizardry in dav1d
18mOn-device Video Playback Upsampling
17mEfficient & Frugal Image Data Processing Pipelines