Skip to content

gpu.md

Latest commit

 

History

History
50 lines (35 loc) · 2.17 KB

gpu.md

File metadata and controls

50 lines (35 loc) · 2.17 KB

GPU Support

Livepeer supports decoding and encoding on NVIDIA GPUs on Linux. GPU transcoding can be enabled by starting Livepeer in -transcoder mode with the -nvidia <device-list> flag. The <device-list> is a comma-separated numerical list of GPU devices that you wish to use for transcoding. If you are unsure of your GPU device, use the nvidia-smi utility. For example, to select devices 0, 2 and 4:

./livepeer -transcoder -nvidia 0,2,4

Limitations

Currently the following limitations are observed:

  • Device validity Ensure valid devices are selected when starting up the node. Currently there is no start-up check to ensure device validity.

  • YUV 4:2:0 input format The pixel format of the source video must be in YUV 4:2:0 format (planar or interleaved). Anything else will return an error.

  • CUDA Availability If running the Livepeer binary, the CUDA shared libraries are expected to be installed in /usr/local/cuda. If the CUDA location differs on your machine, run the node with LD_LIBRARY_PATH=</path/to/cuda> environment variable.

So far, Livepeer has been tested to work with the following driver versions:

CUDA Nvidia
10.0.130
10.1 418.39 , 430.50
10.2.89 440.33.01
11.0 450.36.06

  • Driver Limits "Retail GPU cards may impose a software limit on the number of concurrent transcode sessions allowed on the system in official drivers.

  • Linux Only We've only tested this on Linux. We haven't tried other platforms; if it works elsewhere, especially on Windows or OSX, let us know!

Running Tests

A number of GPU unit tests are included. These may help verify your GPU setup. To run these tests, the Livepeer source code must be obtained; see the install documentation for details on setting up a build environment. Then the Livepeer unit test suite can be run with the NV_DEVICE environment variable. For example, to run the unit tests on GPU 1:

NV_DEVICE=1 bash test.sh

A more intensive set of GPU tests is available in the LPMS repository, which is vendored within go-livepeer. Refer to the LPMS README for details on how to run these tests.