Those numbers were just the minimum required precision that GPU's shaders had to operate at in order to be considered SM2.0 or SM3.0-capable. Any SM3.0-capable GPU will have 32-bit FP ALU's, and will executate a shader compiled at any profile in 32-bit precision. The only GPU's to use 24-bit precision were ATI's 9-series and X-series which were SM2.0 GPU's. NVidia's DX9 GPU's (FX-series, 6-series, 7-series) operated at 32-bit precision when using float's, and 16-bit when using the half datatype in shader programs. Any DX10+ capable GPU will just use 32-bit for both float and half.

FP16 blending refers to the ability to blend 16-bit floating point surface formats such as D3DFMT_A16B16G16R16F. Earlier GPU's couldn't do it, which made it difficult to implement HDR due to the use of blending for multipass lighting or transparents.

SM3.0 supports certain shader assembly instructions that SM2.0 doesn't. For instance, dynamic branching and dynamic looping. So if you compile as ps_3_0 and you use a loop or branch, you may get different code than if you compiled wih ps_2_0.