Support Matrix

Models

Model Name	Model ID	Model Type	Publisher
claire	claire_v2.3.1	Regression	NVIDIA
james	james_v2.3.1	Regression	NVIDIA
mark	mark_v2.3	Regression	NVIDIA
multi	multi_v3.2	Diffusion	NVIDIA

Optimized configurations

GPU	GPU Memory (GB)	Precision	Batch Size (Regression)	Batch Size (Diffusion)
A10G	24	FP16	35	32
A30	24	FP16	55	41
L4	24	FP16	30	16
L40S	48	FP16	80	60
RTX 4090	24	FP16	80	35
RTX 5080	16	FP16	75	22
RTX 5090	32	FP16	100	52
RTX 6000 Ada	48	FP16	70	50
RTX PRO 6000 Blackwell	96	FP16	120	130
B200	192	FP16	255	144

Profiles can be automatically selected when you run the Audio2Face-3D NIM.

To verify auto profile selection, run:

$ docker run -it --rm --network=host --gpus all \
-e NGC_API_KEY=$NGC_API_KEY nvcr.io/nim/nvidia/audio2face-3d:2.0

If a pre-generated profile is found for your GPU, you’ll see this message in the logs:

"timestamp": "2025-02-12 05:53:29,945", "level": "INFO",
"message": "Matched profile_id in manifest from TagsBasedProfileSelector"

When running the Audio2Face-3D NIM, automatic profile selection will attempt to choose a compatible pre-generated profile for your GPU.

More Information on Automatic Profile Selection

This process works as follows:

1. The service detects your GPU and maps it to a profile key in the model manifest.

2. Some GPUs without a dedicated profile are mapped to a compatible alternative (for example, RTX 30 series GPUs are mapped to the A10G profile).

3. The service then starts using the selected TRT engines:

   • Downloaded into /opt/nim/workspace and copied to /tmp/a2x (default behavior), or

   • Loaded / generated locally into /tmp/a2x when NIM_DISABLE_MODEL_DOWNLOAD=true.

GPUs that have no mapping in the manifest (such as A100 or H100) are not supported with pre-generated profiles. To run on these GPUs, set NIM_DISABLE_MODEL_DOWNLOAD=true and generate TRT engines locally (see Getting Started).

Note

Blackwell-class GPUs have dedicated pre-generated profiles (for example RTX PRO 6000 Blackwell and B200). Auto selection will match those profiles directly.

Note

Audio2Face-3D does not have multi GPU support. If you want to use more GPUs, you will have to start multiple instances of Audio2Face-3D for each GPU and manage how requests are routed to each instance.

Non-optimized configurations

The GPU memory usage is dependend on the number of concurrent connections (number_of_streams) configured when starting up Audio2Face-3D.

Note

This approximation was observed on an RTX 4090 and RTX 3080 Ti, without generating the TRT engines.

GPU	GPU Memory (GB)	Precision
Any NVIDIA GPU with sufficient GPU memory and compute capability	0.15 * number_of_streams + 9	FP16

Software

Operating System	Ubuntu 24.04
NVIDIA CUDA	>=12.8, <13.0 (12.9 recommended)
NVIDIA Driver	NVIDIA GPU driver R570+ (R580 recommended) with CUDA 12.8.0+ support
TensorRT	>=10.13, <11.0 (included in the container)
NVIDIA Container Toolkit	latest version
Docker	latest version

• Any Linux distribution should work but has not been tested by our teams

• Your Docker environment must support NVIDIA GPUs. Please refer to the NVIDIA Container Toolkit for more information.

Note

NVIDIA Fabric Manager Requirement

NVIDIA Fabric Manager is required on systems with multiple GPUs connected using NVLink or NVSwitch technology. This typically applies to:

• Multi-GPU systems with NVLink bridges (e.g., DGX systems, HGX platforms)

• Systems with NVSwitch fabric interconnects

• Hosts running NVIDIA B200, H100, A100, V100, or other datacenter GPUs with NVLink

Fabric Manager is not required for:

• Single GPU systems

• Multi-GPU systems without NVLink/NVSwitch (PCIe-only configurations)

For installation instructions, refer to the official NVIDIA Fabric Manager documentation. See also CUDA Error 802 “system not yet initialized” on B200 Multi-GPU Systems in the Troubleshooting guide for B200-specific details.