Unless we get some kind of very different architecture I’m not sure we’re going to get a 10-100x improvement any time soon, api models will always preference throughput over individual task latency, so any improvement that sacrifices serial decode for better parallelism is going to be more economically viable.