Some Ideas for this Issue:
culling vertices (cpu): don’t batch what isn’t necessary to display
combining vertices (cpu): since you’re using voxels, many vertices can be combined when they’re on the same plane
tessellation (gpu): you can sometimes expand vertices from (2) back into the proper level of detail; can possibly use quads here instead of triangles
lighting (pre-cpu): you can bake lighting data into your voxel data
data structure (cpu): you should consider using data structures that provide for simplifying data at a distance. For example, once you hit between 400 and 800 voxels or so away, all you generally see are ‘points’. Consequently, it’s unnecessary to batch up a full 8 vertices when 1 will suffice. And at some point (I haven’t measured), you wouldn’t even need single voxels; you could use mutliple voxels /w a color average instead. Batch calls should be tied to your data structure sizing.
ref: https://www.reddit.com/r/VoxelGameDev/comments/2cd0w6/what_are_the_ways_to_optimize_a_voxel_game/