[quote name='Chris_F' timestamp='1307246947' post='4819629']
You really think float16 is precise enough for your normals?

You could just store your normal buffer in spherical coordinates, which cuts it down to 2 channels instead of 3. This is a common approach and shouldn't effect your precision.

Sorry, I wasn't completely clear. I would be storing each component into a byte. So if I'm storing emission, for example, in the buffer:


| 16 bits | 16 bits | 16 bits | 16 bits |
| buff_r | buff_g | buff_b | buff_a |
| n.x | n.y | n.z | pad | e.r | e.g | e.b | pad |


A conversion to fixed point bytes is probably necessary for the attributes.
[/quote]
Frist off, you can't divide a floating point number simple into two parts, for this you should use an integer format (not greatly supported by older hardware).

For my deferred rendering engine I always use 16bit floats for all rendering targets. Here are some tips to compress low and high detail data:
1. normals
As already mentionend, try to use normal compression (3 component => 2 component) to save some space. Take a look at aras normal compression page over here.

2. depth
16bit for depth is too low, use 2 16bit floats (precison ~24 bit):

encode
float depth = ...
float d_high = floor(depth*1024);
float d_low = fract(depth*1024);

decode
float depth = (d_high+d_low)/1024;


3. flags
For flags use simply the sign.

4. int + float[0..1)
Sometimes an index plus an alpha value [0..0.99999] are useful and these values can simply be compressed into a single float by:

encode
float result = (float)index + alpha

decode
int index = (int)floor(input)
float alpha = fract(alpha)