Is there any point in going to double precision floats, single precision
floats have a dynamic range of over 910 dB and a signal to error ratio of a
least 138 dB below the peak signal level, the threshold of pain is 132dB
above the threshold of hearing. Though most processor now can handle these
size numbers themselves quiet well, my PowerPC processor has now problem
with them, double precision requires twice the memory as single precision
which means our caches fill up quicker causing more cache misses and
therefore slowing down of calculations. I am very interested to know if
there is any signifcant errors, or any,  in the 16 most signifcant bits in
numerous single precision calculation, as in DSP, as I plan to do a thesis
on it for my Audio Engineering class, perhaps it has some effect through
filter co-effectiants.

Nathan Day wrote:
> Is there any point in going to double precision floats, single precision
> floats have a dynamic range of over 910 dB and a signal to error ratio of a
> least 138 dB below the peak signal level, the threshold of pain is 132dB
> above the threshold of hearing.

The size of the mantissa is the limiting factor -- remember the
pitch-stepping problem on large tables that was discussed recently.
Filters can have problems too if you're not careful.

I agree that it's not worth the memory-bandwidth hit to move to
doubles throughout.  If you're having problems getting good 16-bit
output with floats, you're probably using a numerically unstable
algorithm... throwing precision at it may only postpone the problem.
Though it would be nice to be able to specify doubles for particular
purposes.

-- 
   Eli Brandt
   eli+@cs.cmu.edu