So basically serial binary operations with a set bit size as the data input is not ideal if you want full dynamic precision.
could we not return to cisc and have a pipe.
it should usually take one pass worth to do the main calculation only this time it takes a part of a pass dependent on its dynamic size and say 8 steps on the pipe.
at each stage of the pipe its doing the same thing in parallel however this same operation can produce the full turing machine with parallel bitwise operator cells which do various different comparative and flip operations using comparasion's between bijective and standard binary form along with other operations like to memorize or take a result from the pipe.
effective the circuitry logic would have to work part base 3 part base 2.
technically an algorithm in the system could run ahead of itself and you may get speed ups unrealised before with partial differential equations and lining up using fractal triangular curve logic deduced dynamic exponentials to predict the behaviour of such equations.