> This page is for NVIDIA 3D Body Pose NIM, version 1.0.0 (Latest) (default).
> For other versions, use one of these documentation indexes:
> - 1.0.0 (Latest) (default): https://docs.nvidia.com/nim/maxine/body-pose/1.0.0/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/nim/maxine/llms.txt.

# Support Matrix

This page lists the models, GPU architectures, and software prerequisites supported by the 3D Body Pose NIM.

## Models

The following table lists the model packaged in this NIM.

| Model Name | Model ID | Publisher |
| ---------- | -------- | --------- |
| 3D Body Pose | body_pose | NVIDIA |

## Optimized Configurations

The following table lists GPUs with optimized configurations.

| GPU | Compute Capability | Precision |
| --- | ------------------ | --------- |
| A100, A30 | 8.0 | FP16 |
| A10G, A40 | 8.6 | FP16 |
| L4, L40S, RTX 6000 Ada Generation | 8.9 | FP16 |
| H100, H200 | 9.0 | FP16 |
| B200 | 10.0 | FP16 |
| NVIDIA RTX PRO 6000 Blackwell Server Edition | 12.0 | FP16 |

### GPU Memory

24 GB is sufficient. The A10G is the smallest part in the table above and the smallest one benchmarked, and the NIM runs on it with room to spare. Nothing below 24 GB has been validated, so treat that as the tested floor rather than a measured minimum.

## Other Architectures (Consumer RTX GPUs)

| Consumer GPU | Compute Capability | Precision |
| ------------ | ------------------ | --------- |
| RTX 3090 | 8.6 | FP16 |
| RTX 4090 | 8.9 | FP16 |
| RTX 5090 | 12.0 | FP16 |

The 3D Body Pose NIM is compatible with professional and consumer GPUs that have Tensor cores and are based on the following NVIDIA architectures: Blackwell, Hopper, Ada, and Ampere.

## Unsupported Architectures

| GPU | Compute Capability | Reason |
| --- | ------------------ | ------ |
| T4 and other Turing GPUs | 7.5 | Not a supported GPU architecture for this NIM. |

The 3D Body Pose NIM uses NVDEC hardware acceleration for video decoding. GPUs without NVDEC hardware support are not supported. Some GPUs support only a limited number of concurrent NVDEC sessions, which means the NIM can process only that same number of concurrent inputs on those GPUs. For details, refer to the [Video Encode and Decode Support Matrix](https://developer.nvidia.com/video-encode-and-decode-gpu-support-matrix-new).

## Input Media

| | |
| --- | --- |
| Container | MP4, or WebM/Matroska |
| Video codec | H.264 (preferred), H.265, AV1, VP8, VP9 |
| Pixel format | 4:2:0 chroma subsampling, 8-bit, SDR. HDR, 10-bit, 12-bit, 4:2:2, and 4:4:4 content is not supported. |
| Frame rate | Constant frame rate (CFR), enforced. A clip whose frame interval varies, including a constant-rate clip with a dropped frame, is rejected with `INVALID_ARGUMENT`. |
| Maximum size | 1 GiB, configurable using `NV_AI4M_MAX_INPUT_FILE_SIZE_BYTES` |

An input outside this table, or one whose codec this GPU's decoder does not decode, is rejected with `INVALID_ARGUMENT` before decoding. The message names the format found, for example:

```text
unsupported video format: VP9 4:4:4 8-bit (profile 1) -- this NIM accepts 4:2:0 8-bit only (SDR; HDR and 10/12-bit content are not supported). See Support Matrix -> Input Media.
unsupported video format: H.265 4:2:0 10-bit (profile 2) -- this NIM accepts 4:2:0 8-bit only (SDR; HDR and 10/12-bit content are not supported). See Support Matrix -> Input Media.
unsupported video format: AV1 4:2:0 8-bit -- this GPU's decoder does not decode AV1 (it decodes: H.264, H.265, VP8, VP9). See Support Matrix -> Input Media.
```

To transcode such an input to 4:2:0 8-bit H.264 at a constant frame rate, use FFmpeg. It runs on your own machine to prepare a clip before you send it; it is not part of the NIM and is not installed in the container. Check with `ffmpeg -version`, and if it is missing, install it from your distribution's package manager or from a build listed at [ffmpeg.org/download.html](https://ffmpeg.org/download.html).

```bash
ffmpeg -i in.webm -r 30 -c:v libx264 -pix_fmt yuv420p out.mp4
```

## Host Platform

The constraints on this page bind the machine that **runs the container**, not the machine you call it from. The NIM container is published for Linux on amd64 only, and requires an NVIDIA GPU from the architectures listed above.

The Python client communicates with the NIM using gRPC and can run on Windows, macOS, or Linux on Arm64. For example, an Arm workstation can connect to a NIM hosted on an AMD64 Linux machine. Only the host that runs the container requires Linux on AMD64 and a supported GPU.

## Software

### NVIDIA Driver and Prerequisites

The following are required on the host running the container:

| Prerequisite | Version | Download and Installation Steps |
| ------------ | ------- | ------------------------------- |
| NVIDIA driver (Linux) | 570.26+ | [NVIDIA Unix Drivers](https://www.nvidia.com/en-us/drivers/unix/) |
| Docker | latest | Ubuntu, CentOS, Debian: [Install Docker Engine](https://docs.docker.com/engine/install/); Rocky Linux: [Docker - Install Engine](https://docs.rockylinux.org/gemstones/containers/docker/) |
| NVIDIA Container Toolkit | latest | [Installation and configuration instructions](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html) |
| `acl` package, provides `setfacl` (SSL and mTLS only) | Any | Your distribution's package manager, for example `sudo apt install acl`. Needed only by the [SSL Enablement](/nim/maxine/body-pose/advanced-usage#ssl-enablement) recipe. |

3D Body Pose NIM uses the following NVIDIA software platforms:

| Components | Version |
| ---------- | ------- |
| CUDA | 12.8.1 |
| cuDNN | 9.8.0 |
| TRT | 10.9.0.34 |
| Triton Inference Server | 2.56.0 (`nvcr.io/nvidia/tritonserver:25.03-py3`) |
| GStreamer | 1.24.12 |
| DeepStream | 8.0 |