I may not understand much of OpenCL or graphics at all but as network engeneer I still think that hardware limits cannot be exceeded for the purpuse mentioned above.
There is no queue mechanism or alike that can be used to queue solution rate AFAIK. Protocol specifics doesn't allow workarounds.
Yes we can not exceed hardware limits, but for now miner use 63% of mcu and 80% of power, so it not utilize full hardware capacity, why? I suspect that miner computation have 2 phases, one when fetching from memory - cores wait (are blocked), and second cores compute (not use memory).
in simple way:
now threads do: [[external memory read to registers][comp][external memory write result]][[external memory read to registers][comp][external memory write result]]
so one part cores waiting, second part memory not used...
i want (all at the same time):
thread1: [external memory read to cache][external memory read to cache][external memory read to cache]
thread2: [[read cache][comp][write cache]][[read cache][comp][write cache]][[read cache][comp][write cache]]
thread3: [external memory write result][external memory write result][external memory write result]
yes it have flaws in logic, but why i can not discuss that?
. 
