gpuMaxBatchBytes when it is left out: 256 MiB. On Sponza (Apple M5) it
takes the load's peak from 1,483 to 711 MiB (−771 MiB) and settles 11 MiB
lower; the first trees cost about +165 ms at 60 Hz as first measured, when
each batch waited for two updates, and nothing now (166 ms against 171)
that batches go on between updates. docs/internals/design/bvh-builders.md,
"Build peak memory".
gpuMaxBatchByteswhen it is left out: 256 MiB. On Sponza (Apple M5) it takes the load's peak from 1,483 to 711 MiB (−771 MiB) and settles 11 MiB lower; the first trees cost about +165 ms at 60 Hz as first measured, when each batch waited for two updates, and nothing now (166 ms against 171) that batches go on between updates. docs/internals/design/bvh-builders.md, "Build peak memory".