> ## Documentation Index
> Fetch the complete documentation index at: https://livepeerfoundation-d4522ba3.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Hardware & GPU support

> NVIDIA GPU compatibility, NVENC session limits, VRAM requirements by workload, and driver versions for orchestrators.

Livepeer orchestrators use **NVIDIA** GPUs for video transcoding (NVENC/NVDEC) and AI inference
(CUDA / Tensor cores). **AMD and Intel GPUs are not supported.**

## Supported GPUs

| GPU family                | Transcoding | AI inference | Notes                                                      |
| ------------------------- | ----------- | ------------ | ---------------------------------------------------------- |
| GeForce RTX 40xx (Ada)    | Yes         | Yes          | Best consumer option; AV1 encode                           |
| GeForce RTX 30xx (Ampere) | Yes         | Yes          | Widely used; good price/performance                        |
| GeForce RTX 20xx (Turing) | Yes         | Yes          | Supported but older                                        |
| GeForce GTX 16xx (Turing) | Yes         | Limited      | No Tensor cores — AI slower/unsupported for some pipelines |
| GeForce GTX 10xx (Pascal) | Yes         | Limited      | Legacy; NVENC Gen 6; no Tensor cores                       |
| Tesla T4                  | Yes         | Yes          | Data center, 16 GB, common in cloud                        |
| Tesla V100                | Yes         | Yes          | Data center, 16/32 GB                                      |
| A100                      | Yes         | Yes          | Data center, 40/80 GB, highest throughput                  |
| A10 / A10G                | Yes         | Yes          | Cloud-optimized (AWS G5), 24 GB                            |
| L4                        | Yes         | Yes          | Ada data center, 24 GB, good for AI                        |
| L40 / L40S                | Yes         | Yes          | 48 GB, high-end AI + transcoding                           |
| H100                      | Overkill    | Yes          | 80 GB, primarily LLM / large-model inference               |

## NVENC session limits

Consumer GPUs cap concurrent NVENC encode sessions, which limits simultaneous transcode streams per
GPU.

| GPU class                 | Default NVENC sessions |                  |
| ------------------------- | ---------------------- | ---------------- |
| GeForce GTX 10xx          | 2                      | Can be patched   |
| GeForce GTX 16xx          | 3                      | Can be patched   |
| GeForce RTX 20xx          | 3                      | Can be patched   |
| GeForce RTX 30xx          | 3–5 (by model)         | Can be patched   |
| GeForce RTX 40xx          | 3–8 (by model)         | Can be patched   |
| Tesla / Quadro / A-series | Unlimited              | No session limit |

The community [nvidia-patch](https://github.com/keylase/nvidia-patch) removes the limit on consumer
GPUs and is widely used by orchestrators.

<Warning>
  Patching modifies a system binary, is unsupported by NVIDIA, must be re-applied after driver
  updates, and may be disallowed on some managed cloud GPU instances.
</Warning>

## VRAM by workload

| Workload                        | Minimum VRAM | Recommended | Notes                                          |
| ------------------------------- | ------------ | ----------- | ---------------------------------------------- |
| Video transcoding only          | 4 GB         | 8 GB        | NVENC/NVDEC uses minimal VRAM                  |
| Batch AI (single warm model)    | 8 GB         | 16 GB       | SDXL needs \~7 GB                              |
| Batch AI (multiple warm models) | 16 GB        | 24 GB+      | Each warm model consumes VRAM simultaneously   |
| LLM inference (quantized)       | 8 GB         | 16 GB       | Via Ollama runner, quantized weights           |
| LLM inference (full precision)  | 24 GB+       | 48 GB+      | Large models at full precision                 |
| Real-time AI (ComfyStream)      | 12 GB        | 16 GB+      | Latency-sensitive; headroom improves stability |

## Driver & toolkit versions

| Component                | Minimum | Notes                                                       |
| ------------------------ | ------- | ----------------------------------------------------------- |
| NVIDIA driver            | 525+    |                                                             |
| CUDA toolkit             | 12.0+   |                                                             |
| NVIDIA Container Toolkit | Latest  | Required for Docker (AI runner, containerized orchestrator) |

```bash theme={null}
nvidia-smi          # driver version
nvcc --version      # CUDA version
docker run --gpus all nvidia/cuda:12.0-base nvidia-smi   # Docker GPU access
```

## GPU selection guidance

| Goal                     | Pick                                                                                     |
| ------------------------ | ---------------------------------------------------------------------------------------- |
| Transcoding only, budget | GTX 1660 Super (6 GB); patch the NVENC limit for more sessions                           |
| Transcoding + AI         | RTX 4070 Ti Super (16 GB) or RTX 3090 (24 GB) — 24 GB runs 2–3 warm models + transcoding |
| AI-heavy / LLM           | RTX 4090 (24 GB), or A100 / L40S in a data center                                        |

## Related

<CardGroup cols={2}>
  <Card title="Configure your orchestrator" icon="sliders" href="/network/guides/orchestrator-configure">
    GPU selection and session-limit flags.
  </Card>

  <Card title="Add AI inference" icon="microchip" href="/network/guides/orchestrator-add-ai">
    Match a pipeline to your available VRAM.
  </Card>
</CardGroup>