Metadata for the ZFS filesystem, e.g. the pointers and checksums for each block, then the pointers and checksums for that, redundantly written, all the way up to the tree. The amplification involved in that ends up totally filling the pipeline with IOPs for every kind of underlying media, and then you end up at a fraction of the throughput. (Although in a theoretical edge case where you are way below the max throughout but need to only occasionally write independent small blocks, it could have lower latency.)
Most filesystems use 64k extents in grouping data together anyways (even though it is divided up into 4k iops) so it ends up lining up nicely with the blocks when volblocksize=64k
