Is that e.g. batch 16/32 for each operation e.g. 16-row matmuls in a pipeline? Or a pipeline of vector-math ops that has 16/32 stages? Is the pipeline also double digits deep?