Thanks for your reply.
Maybe I had to make myself clearer: I know how to store the FLOAT 2D array in a HDF5 dataset.
What I want is a way to show an image out of it, without loosing precision. So I thought a solution is to make an image out of my 2D array, using the either 8 or 16 bits pixel representation, as you suggest. But I may loose precision on 8bit, maybe less on 16bits, and surely even less in 32 bits.
In other words, can HDF5 produce images from datasets, without precision loss ?