High-Performance Design for the AV1 Multi - Alphabet Arithmetic Decoder

Abstract

In the course of the last years, there has been increasing use of video streams over the Internet. Streaming videos over networks is expensive, as it takes finite resources and physical infrastructure. Hence, companies have been investing in new standards for video encoding with higher compression rates. The AV1 video encoding format was launched in 2018 by the Alliance for Open Media as an open-source and royalty-free format. This paper presents a hardware architecture named Multi-Alphabet Arithmetic Decoder (MaAD) for the first step of the decoder: the entropy decoding, more specifically, its kernel step, the arithmetic decoding block. AV1 uses a multi-alphabet approach of up to sixteen symbols, where the majority of them have different and dynamic probabilities of occurrence. The architecture was synthesized to ST 65nm library, requiring 34.3K gates, and achieved the frequency of 467MHz, with an estimated throughput of 766 Mbits/s, capable of 8K video processing at 60 frames per second. To the best of the authors' knowledge, MaAD is the first-ever academic design with detailed implementation targeting the AV1 arithmetic decoding step.