Makes sense... We're also seeing energy efficiency and model size and latency becoming significant constraints these days, and the more unique constraints an application has, perhaps the more beneficial it is to have many different implementations with different tradeoffs.