The real disadvantage appears to be that a fast implementation is likely to be more expensive, i.e. it requires more complex circuits, thus a greater chip area. Most designers of recent CPUs or GPUs appear to be much more concerned with obtaining a greater speed and a minimum chip area than with obtaining adequate numeric accuracy, because most buyers are influenced by speed benchmark results and know little or nothing about the accuracy required by their computations.
Moreover, the subnormals are an optional feature of the floating-point standard, the alternative to using subnormals is to enable the underflow exception. However, an overwhelming majority of the programmers are too lazy to enable and handle overflow exceptions, much less to enable and handle underflow exceptions.
Unlike using either underflow exceptions or subnormals, the use of flush-to-zero and denormals-are-zero is acceptable only in the programs where computation errors do not matter at all, such as games.
That is not true. FTZ and DAZ are perfectly reasonable in a lot of scientific computing scenario's, where computation errors in general are closely scrutinized.
FTZ and DAZ and any other methods that are certain to introduce errors can be accepted only inside a computation for which there is an independent way to verify the results and that way is always used, without exceptions.
For example, it is indeed OK to use FTZ and DAZ when solving a system of equations, if, and only if, after obtaining a solution it is inserted in the original equations and a residual error is computed, this time without using FTZ and DAZ, and an excessive residual error triggers an appropriate means of notifying the user about the failure.
For high-precision numbers, e.g. for 64 bits, the standard double-precision floating-point numbers are the best.
Any number format using a given number of bits has the same number of values that can be represented. The only difference between various formats, e.g. integers, fixed-point numbers, floating-point numbers or posits, is how the same number of points is distributed over the line of the real numbers, being more dense in some parts and more rare in other parts.
For scientific computation, the optimal distribution of the representable numbers is when the relative error is constant. The ideal number format would be logarithmic, but it is too difficult to add and subtract logarithmic numbers, so the second best format is used, i.e. floating-point numbers, which allow fast implementations for all important arithmetic operations.
Posits have a non-uniform relative error, low close too 1 and high for large and small numbers, so they are unsuitable for complex physics modelling, which needs more computations with large numbers and small numbers than with numbers close to 1.
On the other hand, for the applications where low-precision numbers are suitable, e.g. machine learning, graphics or some kinds of DSP, posits could be better than floating-point numbers.
Logarithmic distribution is optimal for complex physics models where there are relationships between many different kinds of physical quantities, so it is not possible or convenient to choose a scale factor that would make all quantities have values close to 1. In such applications many numbers are very large or very small and any other distribution except logarithmic leads to increased errors in the results. Such models also typically need a 64-bit precision. Moreover, having a much greater exponent range than provided by the 32-bit floating-point format, in order to avoid overflows and underflows, is frequently even more important for them than having a greater precision.
If for an application the range of the 32-bit floating-point numbers is good enough, then there are indeed chances that posits might be better for it.