Hi! Not sure if its a bug or current limitation of karma xpu. Scaled with transform node volume in solaris is rendering much slower vs scaled volume in sop and then imported to solaris. It looks like that transform in solaris (karma xpu mode) didnt change voxel size and make scaled volume huge. In cpu mode render time is same. Attached is example file - switch between two variants to see render time difference. This creates problem when instancing volumes with high scale variance.

there is definitely something fishy going on, please submit a bug

it almost looks like internally Volume Step Rate may be scaled for such volumes, even though they likely will not contain more information since they are just scaled up so it's usually unnecessary to sample them at higher rate
and as you said, it behaves differently from CPU, which seems to have more correct behavior

Submitted. Also If I rembeber right I saw something similar with redshift so maybe its a limitation of gpu renderers?