An extensible framework for video ASIC development and validation at Facebook scale

SPIE Optics + Photonics (SPIE)

Abstract

Video consumption across social platforms has increased at a rapid pace. Video processing is a compute-heavy workload, and domain-specific accelerators (ASICs) allow more efficient scaling than general purpose CPUs. One of the challenges for video ASIC adoption is that videos ingested in datacenters are user-generated content and have a long-tail distribution of uncommon features. Software stack can handle the outliers gracefully, but these uncommon features may pose a challenge for the ASIC with undesirable effects for the unsupported/unhandled end cases. To avoid undesirable effects in the production, it is critical to proof our system against the long-tail conditions early in the product cycle of the ASIC development. Similarly, critical signals like BD-rate quality and outlier detection are needed from production traffic early in the product cycle. To address these needs, we propose an extensible framework that allows a continuous development strategy using production traffic, through progressive evaluation in various product phases of the video ASIC development cycle. A similar framework would benefit other ASIC accelerator programs in reducing time to deploy on large-scale platforms.

Latest Publications

Boosted Dense Retriever

Patrick Lewis, Barlas Oğuz, Wenhan Xiong, Fabio Petroni, Wen-tau Yih, Sebastian Riedel

NAACL - 2022