Asset Optimization#

The high-quality, ground truth asset should be optimized for specific deployments. Considerations should include target hardware and use cases (e.g., configurators, marketing imagery, design review) when optimizing the asset.

In this example, we are building a configurator that streams on GDN. The lowest hardware target is the “L20” and has the following specifications:

  • L40G GPU running in 2:1 mode (½ GPU)

  • 4 Core / 8 Thread CPU

  • GPU RAM 12 GB

  • System RAM 14 GB (28 GB by Sep 2024)

The primary goal is to reduce geometry to fit into system memory and texture sizes to fit in GPU memory. To reduce geometry, you want to leverage the Scene Optimizer Framework’s mesh decimation operation. To reduce texture sizes, there is a sample script you can apply in the script editor.

The mesh decimation and texture size reduction described below takes the asset from:

  • 39M triangles

  • 23.4M verts

  • 10.1 GB GPU memory used

  • 12.1 GB system memory used

To:

  • 5.7M triangles

  • 4.4M verts

  • 3.8 GB GPU memory used

  • 7.7 GB system memory used

This optimization allows for the configurator to run on the “L20”s 12GB GPU RAM and 14 GB system RAM. If the configurator is not optimized, the machine runs out of memory.

To view the L20 instance startup state, you may create a debug seat on GDN dev portal and remote into the machine.

  • 25% of the RAM is taken up by system resources and the OS which leaves 9.5GB

  • Starting USD Viewer with no asset

  • 54% of the system RAM consumed - leaving 6.4GB

  • 25% of GPU RAM consumed - leaving 9GB