floating point, need confirmation

Hi,

I 'd like a confirmation about some information that become a bit old for me

In floating point technology used by modern graphic cards do you confirm
that any number is well defined by several bytes using the next relation:
x=m2e, where x =number, m=mantisse (english word ?), e=exponent
These bytes includes (not sure ) one or more bits for the signal, several
bits for the exponant and the remain for the mantiss (m).
(or is it obsolete now and replaced by a new one )?

For example, do you confirm that in a performing 128 bits graphic card
working in floating point, each pixel can have value ranging between 2^ -218
and 2^127, or 2^345 +1 shades (thus tens of billions of billions of billions
of billions combinations instead of 256) ?

Thnaks in advance

Thierry

"David Phillip Oster" <oster.TakeThisOut@ieee.org> wrote in message
news:oster-616800.10084511032006@newsclstr02.news.prodigy.com...
> In article <4412e148.TakeThisOut@news.vo.lu>, "Thierry" <: -> wrote:
>
>...
>> For example, do you confirm that in a performing 128 bits graphic card
>> working in floating point, each pixel can have value ranging between
>> 2^ -218
>> and 2^127, or 2^345 +1 shades (thus tens of billions of billions of
>> billions
>> of billions combinations instead of 256) ?
>
>...
> Now, assuming 3 color components, and 16-bits per component, that size
> bit pattern can only hold 2^48 distinct states. The number of distinct
> states is independent of whether the bits are interpreted as integers or
> floats. The number of useful states is actually smaller for floats,
> since certain bit patterns mean negative infinity, positive infinity,
> negative underflow, positive underflow (too close to zero), and
> not-a-number. More like hundreds of thousands of billions, rather than
> the number you list.
>
OK. I read the HDR link. Very interesting but ot doesn't help me to
understand my problem.

If I understand well, in a floating point method, say on 8 bits/color, these
floats do no only represents real colors ?
In this case what represent in the facts negative values or NaN ? A special
representation in a special color space ?

if I well undetsand, this huge range of value that exceed the 16.7 millions
colors in floating point is only used when the user zoom-in on a small
portion of a drawing, a way for the system to display a gradient instead of
steps of colors ? And this prevent the appearing of pixelisation. Is this
right ? (curious becuase all system show pixelisation at say, a 1000 time
zoom)

Thierry

>...