It seemed to me that the trade-off is somewhere around 64K blocks. Caching is and reading access can be optimized and varies on OS factors, so it may increase beyond the actual reading capabilities of the HD, but you can still improve time, so I'd definitely argue they do count.