Manindra Parhy from NVIDIA presents a scalable AI-driven approach to video compression that addresses the unsustainable growth of traditional video codecs. While standards like H.264, HEVC, VVC, AV1 and the recent AV2 have incrementally improved compression efficiency, each new codec requires dedicated silicon that cannot be reused, creating cumulative area growth and diminishing returns in visual quality gains.
The proposed solution leverages existing AI accelerators already present in smartphones, laptops and GPUs—ranging from 35-1660 INT8 TOPS today and projected to exceed 100 TOPS even on low-end devices by 2028-2029.
Rather than fully end-to-end neural compression requiring 200+ TOPS (impractical for mobile) or poorly performing hybrid block-based approaches, this scalable AI extension adds an enhancement layer to traditional codecs, operating within 5-50 TOPS. The enhancement bitstream can operate in post-processing mode for artifact reduction and super-resolution, or residual bitstream mode to recover lost textures with approximately 10% bitrate overhead.
Visual demonstrations show significant improvements over AV2 alone, with reconstructed facial details, cleaner edges, and restored textures in tree foliage and drum surfaces. This pragmatic path forward combines three decades of codec innovation with growing AI compute capabilities, aiming for industry adoption in 2-3 years and eventual transition to fully learned compression.
