Sure, but if you're really confident some task is going to be very common, you might want to parse it out and put into its own optimised silicon.
I guess confidence increases or decreases depending on how standardised a certain type of task has become. For example people don't blink so much at fixed function audio or decompression hardware...it's almost boilerplate.
Something like tessellation might be considered boilerplate too, and is fixed function inside GCN GPUs anyway. Maybe they wanted higher performance for that specific task than their GCN GPU target allowed?
I agree that some other candidates might be riskier though. But we're kind of poking around in the dark here