Tutorial: Maximum Box Sizes for Refinement

Maximum box sizes for various amounts of GPU memory (VRAM).

Maximum box sizes for various amounts of GPU memory

The most important factors influencing memory usage for the 3D refinement jobs are the box size of the volume and particles used, as well as the computational batch size. Below is a tabulation of the approximate maximum volume box sizes that are supported by various GPU RAM sizes, valid for the following job types.

• Homogeneous Refinement

• Helical Refinement (BETA)

• Local Refinement (BETA)

Note that non-uniform regularization incurs additional memory cost that has not been benchmarked. The box sizes below are tabulated in the case where non-uniform regularization is disabled.

GPU VRAM (GB)	Approximate Max Volume Box Size (px)
4	682
8	872
11	976
12	1004
16	1110
24 and above	1126

In most 3D refinement and reconstruction jobs, the computational batch size can be changed by setting the "GPU batch size of images" or "Computational minibatch size" parameter. Decreasing this may alleviate GPU out-of-memory issues.

You can also use the Downsample Particles job to reduce the box size (and maximum resolution) of particles before refinement in order to reduce memory usage during refinement.