NVCR: An extensible native runtime for deploying neural video codecs

Neural video coding repositories usually provide a model and evaluation scripts, but a deployable application also needs persistent temporal state, compatible compiled artifacts, a streaming interface, and a defined byte stream. NVCR (Neural Video Codec Runtime) provides these services through a native C + + 20 runtime with stateful sessions, codec-adapter and execution-provider interfaces, target-qualified artifact selection, and bounded access units. Its sole production integration is Microsoft’s Deep Contextual Video Compression–Real Time (DCVC-RT) with TensorRT half-precision inference (FP16) on NVIDIA hardware. We compare this implementation with a Python/PyTorch DCVC-RT implementation on RTX 4070 and Jetson Orin Nano Super using six resolutions, four quantization settings, three prediction-group lengths, and ten independent throughput runs per condition. Each job measures 100 frames after ten warm-up frames. Process throughput includes startup, model loading, warm-up, file input/output (I/O), and teardown. Pooling equal-work runs within each resolution gives NVCR/Python encode and decode throughput ratios of 1.96–3.05 and 1.37–3.01 on RTX 4070, and 1.51–3.50 and 1.27–3.21 on Orin. These ratios compare complete implementations. Across 144 matched quality points, with one quality pass per implementation and condition, NVCR-minus-Python differences range from − 0 . 1 2 9 to + 0 . 1 2 7 dB in 6:1:1-weighted YUV peak signal-to-noise ratio (PSNR), and from − 1 . 4 5 % to + 2 . 0 7 % in entropy bits per pixel. Separate RTX 3050 and RTX 5060 runs demonstrate the deployment workflow on additional NVIDIA laptops. The results support this integration; reuse across learned codec families or independent production providers remains unvalidated. Standard media-container and FFmpeg integration remain future work.