Author_Institution :
Computer Centre, Australian National University, Canberra, Australia.
Abstract :
Summary form only give, as follows. Bauer is quite correct in pointing out that the use of unnormalized floating point format improves the overall accuracy of the representation. It is therefore intriguing to ask why more manufacturers have not adopted it. Presumably most designers prefer to keep floating point number representations unique. In some computers the leading bit of each mantissa is dropped when the floating point number is stored in core, but added back when the number IS sent to the arithmetic u~it or an output unit. To do this one must assume that the leading bit of each mantissa is 1, i.e., normalized form. Whether one places the radix point to the left or the right of the mantissa does not affect the advantages or disadvantages of the integrated representation: we are merely altering all the numbers and all the rounding errors by a factor of 2m.
Keywords :
Accuracy; Australia; Computers; Data mining; Probability density function;
Journal_Title :
Computers, IEEE Transactions on
DOI :
10.1109/TC.1977.5009302