It was read of initialized memory around holdout_weight in cases when holdout material is used. Seems that it should be assigned to result of shader_holdout_eval here. If Brecht could double check this it'll be great. This could potentially fix #32224: Holdout Error with CUDA Cycles Render