Serving World Models: Latency vs Throughput with NVIDIA Cosmos 3 Super on Nebius
Serving world models for video generation demands a different latency/throughput balance than LLMs - and the right setup makes all the difference. The Nebius Physical AI team tests four different ways to split up 8 GPUs for running NVIDIA Cosmos 3 Super, from using all 8 together to running 8 separate single-GPU setups. Join the upcoming livestream on October 28, 9am PT to see what the data shows…and bring your questions for the NVIDIA and Nebius teams: https://www.youtube.com/watch?v=i9FZyG9-Gx0
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
Add the livestream to your calendar: Serving World Models: Latency vs Throughput with NVIDIA Cosmos 3 Super on Nebius | Cosmos Labs | AddEvent
Read the Technical Blog: World models: Request latency model serving with 42.5% more video at 5.5x the latency
Explore the Cosmos 3 Model Collection: Cosmos3 - a nvidia Collection
Get Models and Code on GitHub: GitHub - NVIDIA/cosmos: NVIDIA Cosmos is an open platform of world models, datasets, and tools that enables developers to build Physical AI for robots, autonomous vehicles, smart infrastructure, and more. · GitHub
