Approximate SRAM-Based Memory System for Intra-Frame Prediction in VVC Encoders

This paper introduces a memory system for the Versatile Video Coding (VVC) standard, aimed at optimizing energy consumption in intra-frame prediction through approximate storage techniques. We propose a system that employs data approximation due to dynamic voltage scaling Static Random Access Memory (SRAM) in two distinct buffers: the Neighbor Samples Buffer (NeighSB) and the Original Samples Buffer (OrigSB). Our findings indicate the feasibility of approximate storage exploitation of our memory system, with minor coding efficiency drops at controlled error rates. As NeighSB and OrigSB presented distinct resilience behaviors, we evaluate different approximation strengths to obtain the most suitable coding efficiency, enabling still effective energy savings. Finally, we investigate the energy savings achieved through different SRAM operation levels, yielding reductions of 42%-62%.