New mixed-precision codec enables cross-platform neural video streaming

Neural video codecs (NVCs) deliver impressive rate-distortion performance, making them ideal for bandwidth-constrained 5G and satellite direct-to-cell links. But real-world deployment has been blocked by cross-platform floating-point non-determinism—identical bitstreams can crash or desync when decoded on different GPU architectures. Prior integer-based fixes (INT8, INT16) either gut compression efficiency or destroy hardware acceleration. Now, researchers Kasidis Arunruangsirilert, Heming Sun, and Jiro Katto propose a mixed precision strategy: P-frames run in hardware-accelerated FP16 for real-time throughput, while I-frames and periodic feature-adapter resets are forced to IEEE-754-compliant FP32, creating deterministic synchronization boundaries that guarantee decoders stay aligned across systems.

The team validated the approach across 12 GPUs spanning four architectural generations, and report elimination of intra-generation fragmentation with substantially broader cross-die interoperability—achieving seamless cross-generation decodability at 1080p with negligible impact on compression efficiency. They also tested end-to-end latency over real-world networks including Wi-Fi 6, 5G NR (FDD/TDD), and Starlink direct-to-cell links, proving the practical viability of streamable learned video compression while documenting unique challenges in non-terrestrial networks. The paper is set for presentation at IEEE GLOBECOM 2026 in Macau.