reflects one key insight: that the work of video encoding can be divided into fast and slow
parts, with the “slow” work done in parallel across thousands of tiny threads, and only “fast”
work done serially. The slower the per-frame processing—eg, for pipelines that use 4K or
8K frames or sophisticated video compression, or that run a computer-vision analysis on
each frame—the more favorable the case for finegrained parallelism and for our approach …