Serving World Models: Latency vs Throughput with NVIDIA Cosmos 3 Super on Nebius | Cosmos Labs
From NVIDIA Developer ยท ยท 57 min

The player loads from YouTube only when you click. It uses YouTube's privacy-enhanced mode.
If the video does not play, the owner may have turned off embedding. Watch it on YouTube instead.
Description
Serving world models for video generation requires a different balance of latency and throughput than serving large language models. In this livestream, Nebius Physical AI team members Stewart Tong and Timothy Le examine NVIDIA Cosmos 3 Super benchmarks through vLLM-Omni on NVIDIA HGX B200 and HGX H200 systems.
Their work compares four serving topologies, ranging from one eight-GPU replica to eight single-GPU replicas. See how replica count, parallelism, and concurrency affect request latency and validated video-seconds per node-hour. Learn which settings to hold constant for video generation benchmarks and how to check generated clips for file integrity and basic motion.
What You'll Learn: - Why faster video requests can mean less output per node - How to measure latency and throughput with video-validation checks - When more replicas help and extra concurrency adds delay - What NVIDIA HGX B200 and HGX H200 benchmarks show about allocating GPUs - Have questions about serving NVIDIA Cosmos 3 Super for world-model video generation? Drop them in live โ the NVIDIA and Nebius teams will address them during the discussion.
Learn more about NVIDIA Cosmos 3 serving and reproduce the benchmark:
๐ Read the technical blog โ https://nebius.com/blog/posts/serving-world-models-latency-node-throughput ๐งช Reproduce the benchmarks โ https://github.com/nebius/nebius-physical-ai ๐ Explore the NVIDIA Cosmos platform โ https://github.com/nvidia/Cosmos ๐ฅ Join the NVIDIA Omniverse community โ https://discord.com/invite/nvidiaomniverse