NVIDIA Video Codec SDK 13.1: Zero-Copy Transcode, AV1 B-Frames, and Frame-Accurate Seek | NVIDIA Technical Blog

SDK 13.1 introduces Hierarchical Reference Mode for AV1 encoding, supporting up to 31 B-frames, which improves quality with minimal performance impact but increased video memory consumption. It also combines UHQ tuning info with iterative encoding. Decode improvements include per-macroblock decode statistics for H.264/HEVC, enabling GPU-accelerated video analytics without CPU overhead. Frame-accurate seek is now supported via a GOP-aware architecture in the NvVideoDecoder class, optimizing for AI and editing workflows by processing only necessary frames. The transcoder pipeline is redesigned with a modular, queue-based architecture and zero-copy transcode support using application-allocated CUarray buffers, leading to reduced SM and memory bandwidth utilization, higher throughput, and simplified deployment with a Docker-based environment.