This is an old revision of the document!
Summary of the discussion in 2004 (without IEEE 754-2008 knowledge of course):
32 bit 48 bit ----------------------------------------------------------------------- "Standard" (whatever that's Better precision. worth). Will meet the ISO C requirements. Needs slightly less memory and is faster. Format can be optimized for 6502. Compiler can use native fp routines. But: Limits number of host platforms. Doesn't meet ISO C Needs slightly more memory and is requirements. slower. Less precision. Compiler needs fp library. But: will work on all host platforms. Binary format is given and we have to live with that. Implementation needs more code, so programs have a larger memory footprint. Is suboptimal. wrt exponent could be made optimal wrt. processing (8 bit covered in exponent processing 2 bytes)
The C standard allows setting errno, raising an exception, or just ignoring the condition for most operations. The standard does not require NaNs and infinite values, so it may be an option not to implement them, because this removes all corresponding checks in the math functions resulting in smaller code size. The TMS320 DSP has this behaviour and it's quite convenient.