dollas@uiucdcsb.Uiuc.ARPA (07/12/85)
To the best of my knowledge there is no such thing as a floating point
A/D converter. The job of an A/D converter is to resolve an
analog signal within a range to a binary number. Usually
16 or 18 bits are about as good as you can get (and on top of that I
claim that even if you had more bits of accuracy there would be little
use for them since the signals you are most likely to digitize are not
that accurate, there is noise, etc). With these 16 or 18 bits you can
convince yourself that the IEEE floating point format is an overkill,
even if you decide to only use the mantissa part of it.
Since for the computational part (or for any other reason) you want to
use the IEEE format, if all the chips you use are of the same family (ie
compatible) you can set all the bits of the exponent to the value that
is effectively zero (that would be 127 due to the exponent bias I think)
and use the bits of your A/D as the most significant digits of the
mantissa (and hardwire the rest to logical 'zero'). Here you see that
if you want to change the offset of the range you can do it hardware-wise,
or in the software.
As a last remark, there may be out there a format converter chip that
translates a weighted binary number to IEEE standard format; I do not
know of the existence of such a chip, maybe other netters do however.
Good Luck!
Apostolos Dollas
U of Illinois
Dept. of Computer Science
USENET: ...!{pur-ee,ihnp4}!uiucdcs!dollas
CSNET: dollas%uiuc@csnet-relay.arpa
ARPA: dollas@uiuc.arpamball@noscvax.UUCP (Michael S. Ball) (07/16/85)
In article <5600033@uiucdcsb> dollas@uiucdcsb.Uiuc.ARPA writes: > >To the best of my knowledge there is no such thing as a floating point >A/D converter. The job of an A/D converter is to resolve an >analog signal within a range to a binary number. Usually >16 or 18 bits are about as good as you can get (and on top of that I >claim that even if you had more bits of accuracy there would be little >use for them since the signals you are most likely to digitize are not >that accurate, there is noise, etc). Actually, there are floating point A/D converters. They are used in specialized applications like active sonar systems, where the dynamic range may be up to 120 dB. In such cases the exponent is provided by some external device, usually driven from a clock. The result is still unnormalized, and must be further processed. For most applications, including almost all audio applications, 16 bits of raw data is quite adequate. This is not to imply that 16 bits is adequate for the signal processing, as the number of significant bits can rise rapidly as the signal is filtered. -Mike Ball- TauMetric Corporation. San Diego, CA.