Engineering at Meta is the official engineering blog of Meta, providing in-depth technical articles and insights directly from Meta's engineers. The blog covers a wide range of topics including artificial intelligence, virtual and augmented reality, infrastructure evolution, and open-source projects. It differentiates itself by offering exclusive behind-the-scenes looks at how Meta builds and scales its technology to serve billions of users worldwide. The content is written by the engineers themselves, providing a unique perspective on solving real-world problems at massive scale.
Meta has transitioned its Instagram Live ingest infrastructure from the proprietary RUSH protocol to the open IETF standard Media over QUIC (MoQ). This deployment aims to unify Meta's streaming transport stack, improve latency and resilience, and leverage a standards-based approach for future live and interactive video services.
Meta has extended its Private Processing confidential computing infrastructure to its AI glasses, utilizing Trusted Execution Environments (TEEs) to process and store user data in the cloud. This architecture aims to enable stateful, personalized AI features while ensuring that sensitive data remains inaccessible to Meta and third parties.
Meta has expanded AV1 video codec support to the majority of mobile devices for its real-time communication (RTC) applications, Messenger and WhatsApp. To overcome computational and bandwidth-constrained network limitations, Meta deployed an ultra-low-complexity encoder preset, integrated the dav1d decoder, and utilized a custom ML-based device eligibility framework. The architecture also leverages adaptive codec complexity switching and advanced error-resilience tools like temporal layers and Long-Term Reference (LTR) frames.
Meta has deprecated its internal FFmpeg fork after collaborating with the open-source community to add key features to the main project. The newly added capabilities, inspired by Meta's internal work, include efficient threaded multi-lane transcoding and real-time quality metric computation during encoding. The company states it will continue to maintain internal-only patches for its custom Meta Scalable Video Processor (MSVP) transcoding ASIC due to its hardware-specific nature.
Meta has developed Andromeda, a personalized ads retrieval engine leveraging NVIDIA Grace Hopper Superchip and Meta Training and Inference Accelerator (MTIA) hardware, to enhance its Advantage+ automation for ad recommendations. This system utilizes deep neural networks and hierarchical indexing to improve personalization and ad performance, achieving a +6% recall improvement and +8% ads quality improvement on selected segments within Instagram and Facebook applications. Andromeda is designed to scale with the exponential growth of ad creatives generated by generative AI tools, increasing efficiency and return on ad spend for advertisers.
