This is to figure out the bottlenecks here.

Since Louis prototyped it, there has been some development. Specifically, we created prefetching, so it's quite possible that we're now bottlenecking on CPU.

We'll need to understand if we are and by how much before proceeding with a parallelisation strategy.