Select Specific NVIDIA GPU for FFmpeg Decoding

When running FFmpeg on a system with multiple NVIDIA graphics cards, hardware-accelerated video decoding defaults to the primary GPU (GPU 0). This guide covers how to direct FFmpeg to a specific NVIDIA GPU for decoding using the modern -hwaccel_device flag, decoder-specific arguments, or the CUDA_VISIBLE_DEVICES environment variable, ensuring efficient workload distribution across multiple cards.

Step 1: Identify Your GPU Index

Before specifying a device, determine the index assigned to each NVIDIA card on your system. Run the following command in your terminal:

nvidia-smi

Look at the leftmost column labeled GPU. The indices are zero-based (e.g., 0, 1, 2). Note the index number of the GPU you intend to use for hardware decoding.

Method 1: Use the -hwaccel_device Flag (Recommended)

The modern, standard method for selecting an NVIDIA GPU in FFmpeg uses the CUDA hardware acceleration framework paired with the -hwaccel_device parameter.

Place -hwaccel cuda and -hwaccel_device <index> before the input file (-i):

ffmpeg -hwaccel cuda -hwaccel_device 1 -i input.mp4 -c:v libx264 output.mp4

In this command:

  • -hwaccel cuda activates NVIDIA CUDA-based hardware decoding (NVDEC).
  • -hwaccel_device 1 routes the decoding workload specifically to GPU 1.
  • -i input.mp4 specifies the source file.

Method 2: Use the -gpu Option with CUVID Decoders

If you are using legacy CUVID decoders directly (such as h264_cuvid or hevc_cuvid), you can specify the target card using the -gpu flag prior to the input:

ffmpeg -c:v h264_cuvid -gpu 1 -i input.mp4 -c:v libx264 output.mp4

In this command:

  • -c:v h264_cuvid forces the use of the NVIDIA CUVID H.264 decoder.
  • -gpu 1 forces the decoder to run on GPU 1.

Method 3: Use the CUDA_VISIBLE_DEVICES Environment Variable

You can restrict FFmpeg’s visibility to a specific GPU at the operating system level. When set, FFmpeg will only see the specified card and treat it as GPU 0.

On Linux/macOS:

CUDA_VISIBLE_DEVICES=1 ffmpeg -hwaccel cuda -i input.mp4 -c:v libx264 output.mp4

On Windows (PowerShell):

$env:CUDA_VISIBLE_DEVICES="1"; ffmpeg -hwaccel cuda -i input.mp4 -c:v libx264 output.mp4

On Windows (Command Prompt):

set CUDA_VISIBLE_DEVICES=1 && ffmpeg -hwaccel cuda -i input.mp4 -c:v libx264 output.mp4

Complete Pipeline: Decoding and Encoding on the Same GPU

To avoid copying frames back and forth across the PCIe bus, maintain the video frames in the VRAM of the designated GPU for both decoding and encoding:

ffmpeg -hwaccel cuda -hwaccel_output_format cuda -hwaccel_device 1 -i input.mp4 -c:v h264_nvenc -gpu 1 output.mp4
  • -hwaccel_output_format cuda keeps the decoded frames in GPU memory.
  • -c:v h264_nvenc -gpu 1 ensures the encoder also uses GPU 1 to process the stored frames directly.