Skip to main content

Using the GPU

Some types of content in Remotion can benefit from a GPU being available on the machine that is used for rendering. That is:

  • WebGL content (Three.JS, Skia, P5.js, Mapbox etc.)
  • 2D Canvas graphics
  • GPU-accelerated CSS properties such as box-shadow, text-shadow, background-image: linear-gradient(), background-image: radial-gradient(), filter: blur(), filter: drop-shadow()

If a GPU is available, it should be enabled by default while in the Remotion Studio or Remotion Player. However, in headless mode, Chromium disables the GPU, leading to a significant slowdown in rendering time.

Use --gl=angle for local renders

Since Chrome 98, the GPU can be used in headless mode. Adding --gl=angle (or {chromiumOptions: {gl: "angle"}} for the Node.JS APIs), we find that a video rendered on a <canvas> on macOS is many times faster compared to rendering without the flag.

However, there seems to be memory leak from Chrome that may kill a long render, therefore we don't set angle as default. We recommend to render long videos that use the GPU in multiple parts.

Use --gl=swangle for Lambda

Since Lambda does not have a GPU, the content has to be rendered using software. You can still render all types of content without having a GPU, it will just be slower. The best way to do so is by passing --gl=swangle if you are rendering using the CLI, or passing {chromiumOptions: {gl: "swangle"}} if using the Node.JS APIs.

Considerations

For rendering content that can benefit from a GPU, you might want to choose a cloud rendering solution to which a GPU can be attached to over AWS Lambda (which does not have a GPU). Most bigger cloud providers have some GPU-enabled VPS offerings. Apple M1 VPS instances might also be able to accelerate graphics rendering and be more economical than VPS instances with desktop graphic cards.

Using the GPU on Lambda

AWS Lambda instances have no GPU, so it is not possible to use it.

What are your experiences?

We'd love to learn and document more findings about the GPU. Let us know and we will amend this document!

See also