NVIDIA Details Video Summarization Microservice Performance on H100, RTX, L40S GPUs
NVIDIA has released performance metrics for its Video Summarization microservice, detailing end-to-end latency and maximum concurrent request capacity. The data is provided across various video lengths and GPU platforms, including H100, RTX Professional 6000 SE, and L40S. This information helps streaming professionals size GPU infrastructure for AI-driven video processing workloads.
Key Takeaways
- NVIDIA's Video Summarization microservice uses RTVI-VLM for vision inference and Nemotron 3 Nano for summarization, operating at FP8 model precision.
- An H100 GPU configuration with a 4x4 topology summarizes a 10-minute video in 20.1 seconds, achieving 125 concurrent requests for a 10-minute video at target latency.
- RTX Professional 6000 SE, with a 4x4 topology, summarizes a 10-minute video in 25.9 seconds and handles 79 concurrent requests for a 10-minute video.
- The L40S GPU, using a 4x4 topology, processes a 10-minute video summary in 39.1 seconds, supporting 41 concurrent requests for a 10-minute video.
- Performance metrics are provided for video lengths ranging from 1 minute to 720 minutes, tested in a warehouse safety monitoring scenario.
Why It Matters
NVIDIA's detailed performance benchmarks provide concrete data for optimizing GPU infrastructure in AI-driven video processing. This allows enterprises to make informed decisions on hardware deployments for real-time video summarization, impacting efficiency and cost for applications like content moderation, security, and media analysis. The focus on specific GPU models and configurations highlights NVIDIA's push to integrate its hardware deeper into the video AI pipeline. Going forward, watch for adoption rates of these specific GPU configurations in enterprise video AI deployments and any subsequent updates on real-world performance at scale.
Additional Context
NVIDIA has been actively enhancing its video AI capabilities, with the Video Summarization microservice (VSS) representing a key component. In May 2026, NVIDIA released a Metropolis Blueprint for VSS, which aims to transform large volumes of video into searchable, actionable intelligence (NVIDIA Developer Blog, May 2026). This blueprint emphasizes a modular design, advanced fusion search, and integration with AI agents like Codex and OpenClaw, enabling automated deployment and interaction via chat interfaces. The VSS architecture provides a reference for building video analytics AI agents that perceive, reason, and act on live and recorded video streams. VSS also supports the use of OpenAI-compatible Vision-Language Models (VLMs) and Large Language Models (LLMs), alongside its own optimized VLMs like CR1, CR2, and Qwen, for flexible model selection (NVIDIA VSS documentation, current). Furthermore, the VSS microservice offers both REST API and Model Context Protocol (MCP) interfaces, facilitating integration into diverse workflows and AI orchestration systems (NVIDIA VSS documentation, current).
Read full article at docs.nvidia.com