There's an important distinction between the decoding speed of documents and overall throughput.

Wonder what can be done to realistically generate something approximating a real world benchmark to cover the throughput scenario.