Efficient video processing at scale using MSVP
Journal
Proceedings of SPIE - The International Society for Optical Engineering
Journal Volume
12674
ISBN
9781510665620
Date Issued
2023-01-01
Author(s)
Reddy, H. M.
ALBERT CHEN
Lan, J.
Katsavounidis, I.
Anandharengan, B.
Lalgudi, H. G.
Alaparthi, S.
Hua, G.
Chuang, H. C.
Wu, P. H.
Lei, Z.
Mastro, A.
Petersen, C.
Chaudhari, G.
Prakash, P.
Regunathan, S.
Reddy, S.
Venkatapuram, P.
Rao, V.
Nom, K.
Bjorlin, A.
Zeile, M.
Lewis, A.
Singh, A.
Sunil, A.
Chen, C. C.
Lin, C. F.
Chen, C.
Sundar, D. P.
Jayaraman, D.
Ucar, H.
Li, H.
Singh, J.
Liu, J. C.C.
Rachamreddy, K. R.
Sriadibhatla, K.
Datla, K.
Berg, L. V.D.
Feng, L.
Jampani, P.
Moola, R.
Mallya, R.
Jha, S.
Pan, S.
Srinivasan, S.
Vaduganathan, V.
Zha, X.
Wang, Z.
Sengottuvel, A. K.
Alluri, B.
Oshin, B.
Kanumetta, C.
Sahin, E.
Athaide, J. M.
Kurapati, K. C.
Manthati, K.
Thottempudi, K.
Chennamsetti, R. R.
Jagannath, K. R.
Arvapalli, S.
Kala, T.
Wang, T.
Chopda, P.
Gandhi, K.
Ramesh, A.
Gupta, R.
Fadnavis, S.
Qassoud, A.
Friedt, C.
Li, F.
Gao, H.
JIH-HSIANG LEE
Dixit, M.
Ugaji, S.
Karuturi, T.
Xie, X.
Narasimha, A.
Jakka, B.
Dodds, B.
Yang, J.
Skandakumaran, K.
Modi, M.
Modi, P.
Stejerean, C.
Ronca, D.
Wang, H.
Pham, N.
Lu, L.
Shen, H.
Ning, J.
Narayanan, K.
Chen, L.
Avidan, N.
Arnold, W.
Xu, F.
Patil, G.
Balan, V.
Grandhi, S. D.
Abstract
Videos uploaded to Meta's Family-of-Apps are transcoded into multiple bitstreams of various coding formats, resolutions and qualities to provide the best quality of experience across a wide variety of devices and connection bandwidth constraints. On Facebook alone, there are more than 4 billion video views per day. To address video processing at this scale, we need a video processing solution that can deliver the best video quality possible, with the shortest amount of encoding time, while being scalable, programmable and energy efficient. In this paper, we present Meta Scalable Video Processor (MSVP), a custom ASIC developed by Meta, that can transcode videos at the similar quality as software encoders, but with much higher throughput and lower power consumption. Each MSVP ASIC can deliver a peak SIMO (Single Input Multiple Output) transcoding performance of 4K 15fps with the highest quality configuration, and can deliver up to 4K 60fps with the standard quality configuration. This performance is achieved at ~10W of PCIe module power. We achieved a throughput gain of ~8x for H.264 when compared against libx264 SW encoding at medium preset. For VP9, we achieved a throughput gain of ~50x when compared against libvpx speed 2 preset. Key components of MSVP transcoding include video decoding, frame rescaling, encoding and quality metric computation. In this paper, we will first provide a high-level overview of the MSVP architecture. Then, detailed design of individual components will be introduced. Finally, benchmarking results that compare the compression efficiency and performance per Watt against software encoders will be presented.
Subjects
AVI | H.264/AVC | video on demand | Video processing | Video transcoding | VP9
SDGs
Type
conference paper