NVIDIA announced Cosmos, a platform for world foundation models and development tools for physical AI. Cosmos addresses challenges in sensor adaptation and reliable deployment, offering modular tools for training, reinforcement learning, optimization, and inference.
What happened
NVIDIA has introduced Cosmos, a platform designed to address the challenges faced by physical AI developers in adapting models to sensors and deploying them reliably. Cosmos provides world foundation models and development tools, including modular, open tools for training, reinforcement learning, optimization, and inference. The Cosmos Framework supports models such as Cosmos 3, NVIDIA Alpamayo, GR00T, Qwen, and pi0.5. Cosmos 3 Nano simulation recipes generate synchronized camera video and LiDAR observations for various scenarios, and Cosmos NIM microservices deliver accelerated inference with quantization, speculative decoding, and KV-cache-aware routing, achieving roughly 2-3x speedups on H200 GPUs.
What to weigh
- GPU memory requirements vary by model and task.
- Cosmos NIM microservices deliver significant speedups on H200 GPUs.
FAQ
What models does Cosmos support?
Cosmos supports models such as Cosmos 3, NVIDIA Alpamayo, GR00T, Qwen, and pi0.5.
What is the speedup on H200 GPUs?
Cosmos NIM microservices deliver roughly 2–3x speedups on H200 GPUs.
What are the GPU memory requirements?
GPU memory requirements vary by model and task, with Cosmos3-Nano generation needing 58 GiB in BF16 and 44 GiB in FP8 with offloading.
Source: NVIDIA