Hello maru, I think you can easily reproduce this. Create a teapot, add 2+ turbosmooth iterations. It has to be complex geometry. A simple box object won't produce this issue.
Apply a material with gloss 0.8, ior 999, anisotropy 0.4, and c triplanar + c bitmap with some imperfection map plugged into aniosotropic rotation. It's just a simple metal shader. For lighting I'm using studio HDRI from hdri haven.
With Blending set to 1 in triplanar map, there is a significant drop in thread performance. In my case core usage goes down to 65%. With blending set to 0 there is no drop. With 0.5 there is some drop. So it seems like that the perfromance drop is directly correlated with the blending amount.
Let me know how it goes.