I have recently added more RTX cards to my rendering rig. It suddenly appeared to me that the number of cards added linearly to the scene build time. As if the PC is loading data to VRAMs one by one.
The average render time per frame is 60 seconds, 20 of which is required for scene build. As such, the additional cards brought minimal benefit to total render times per frame. (i.e. new card cuts actual rednering time by 10 seconds but adds 5 seconds to scene build.)
I had a workaround when using CUDA only. I would run two instances of blender, one for even numbered frames and the other for odds. This way, when one instance was building the scene, the other would still make use of the GPU. But with Optix, “building Optix accelaration structure” still makes use of the GPU while scene build (ray intersection I guess) and makes my trick useless.
I would appreciate any help. Hopefully, I am missing a setting or another workaround.