[color="#1C2837"]Is the only difference the amount of memory needed?[/quote]Them being 32, 64 and 128bpp means storage sizes are quite different, but besides taking up more memory, it's also slower to read/write the larger formats (because more data has to be sent around). Obviously you get something for that price, namely higher precision with the larger formats
Generally you want to use the smallest formats that give you the desired precision (e.g. 16-bit float is probably more than enough precision, meaning you shouldn't have to use 32-bit float).

- I was reading the HDRFormats10 sample and when it creates the render targets instead of using R9G9B9E5_SHAREDEXP format, it uses the DXGI_FORMAT_R16G16B16A16_FLOAT format... Why cant it use R9G9B9E5_SHAREDEXP format?

- Which render target formats convert from gamma 2.2 to gamma 1.0 automatically for free?

- Do I have to apply any gamma correction when I sample diffuse maps?

-I'm guessing it's because that sample supports D3D9, where that format isn't available. You should be able to use it yourself with no problems, as long as it provides adequate precision for your needs.

-Any format that end with the _SRGB suffix will automatically be converted from sRGB to linear when sampled, and vice versa when written to by a pixel shader.

-Usually you do...it depends on how the texture was authored. In most cases a diffuse map will be authored in an image editing program like Photoshop or GIMP, in which case they are authored in sRGB space and saved that way. In that case you want to load those images using an sRGB format, so that they get gamma corrected when sampled. There can be a few exceptions to this case, so it's always nice to be able to easily select sRGB or linear on a per-texture basis.

So, since HDRFormats10 use DXGI_FORMAT_R16G16B16A16_FLOAT it should convert gamma when it samples and writes to render targets right? But I cannot find any gamma correction code... So I guess I'm missing something...

No, you don't want to convert to sRGB when rendering HDR values to a floating point surface. This is because

A) sRGB only works for [0, 1] values

and

Floating point formats have sufficient precision for accurately storing linear color values