Meta Reveals Use of FFmpeg for Video at Scale

Meta's video infrastructure processes over 1 billion new uploads daily, with video accounting for more than 60% of the time users spend on Facebook and Instagram. To handle this, the company executes FFmpeg and its ffprobe utility tens of billions of times every day for tasks like transcoding and media properties analysis. For years, Meta maintained its own internal fork of FFmpeg to add critical features for at-scale processing. This fork included capabilities like threaded multi-lane encoding and real-time quality metric computation, which were necessary to handle their massive workload efficiently but diverged significantly from the public version of the open-source project. Recently, Meta collaborated with FFmpeg developers and supporting companies like FFlabs and VideoLAN to integrate its unique, at-scale features directly into the main FFmpeg project. This strategic shift was driven by the need to incorporate new codecs, formats, and reliability improvements from the upstream version without the engineering overhead of maintaining a separate, increasingly outdated fork. One of Meta's key contributions was influencing a major refactoring of FFmpeg's threading model, which landed in versions 6.0 and 8.0. This enhancement, directly influenced by the design of Meta's internal fork, allows for parallelized video encoding, significantly improving efficiency by deduplicating processes when creating multiple outputs for a single video. Beyond software, Meta has also integrated support for its own custom ASIC, the Meta Scalable Video Processor (MSVP), directly into the FFmpeg framework. This allows them to use the same FFmpeg APIs to orchestrate transcoding across both standard hardware (from Intel, NVIDIA, AMD) and their own specialized, power-efficient chips. This reliance on a foundational open-source tool is standard across the industry. Netflix uses FFmpeg for efficient encoding and has developed its own FFmpeg-based filters for tasks like neural network-based video downscaling. Similarly, YouTube has historically used FFmpeg for transcoding a vast array of user-uploaded formats.