Prerequisites#

This page outlines the required hardware and software for each Cosmos component.

Hardware#

Predict2.5#

Use NVIDIA GPUs with Ampere architecture or newer, including RTX 30 Series and A100 GPUs. Refer to the Predict2.5 Model Matrix for the required number of GPUs and memory for inference and post-training.

Predict2#

Use NVIDIA GPUs with Ampere architecture or newer, including RTX 30 Series and A100 GPUs. Refer to the Predict2 Model Matrix for the required number of GPUs and memory for inference and post-training.

Predict1#

Use NVIDIA H100-80GB or A100-80GB GPUs for inference and post-training. Refer to the Predict1 Model Matrix page for the required number of GPUs and memory for inference and post-training.

Transfer2.5#

Use NVIDIA GPUs with Ampere architecture or newer, including RTX 30 Series and A100 GPUs. Refer to the Transfer2.5 Model Matrix for the required number of GPUs and memory for inference and post-training.

Transfer1#

Use NVIDIA GPUs with Ampere architecture or newer, including RTX 30 Series and A100 GPUs.

Reason2#

Reason2 is supported on the Hopper and Blackwell GPU architectures; other GPUs may work, but have not been tested. Reason2 has been validated on the following GPUs:

GPU

CUDA Version

Functionality

NVIDIA H100

12.8

inference/post-training/quantization

NVIDIA GB200

13.0

inference

NVIDIA DGX Spark

13.0

inference

NVIDIA Jetson AGX Thor (Edge)

13.0

Inference via Transformers

The following are the minimum GPU and memory values for each Reason2 model:

  • Cosmos-Reason2-2B: 24GB

  • Cosmos-Reason2-8B: 32GB

Reason1#

Use NVIDIA GPUs with Ampere architecture or newer, including RTX 30 Series and A100 GPUs. The following are the minimum GPU and memory values for inference and post-training:

  • Inference: 1 GPU with 24GB memory

  • Post-training: 4 GPUs with 80GB of memory

Software#

Predict2.5, Transfer2.5#

Other Model Families#