The discussion about cpu caching challenges etc seems like it's missing the point. Wouldn't it be more likely to DMA the results to a GPU anyway?