I'm not familiar with the algorithms, but this might be true. MMX deals with integers and if iDCT traditionally required floating-point numbers, then using integers would involve some drastic approximations resulting in quality loss. That being said, if a totally different algorithm for iDCT has been made then MMX - even using integers - might produce results comparable and/or equivalent to the old iDCT algorithms involving floating point numbers.
I'd say that, at least with MPEG, using MMX and consequently, using integers which yield very gross approximations may not be such a big deal since the codec is a lossy one and the "lossy" part of the quality might be more attributable to another aspect of MPEG than the loss of precision attributed to using integer approximations in doing iDCT.
Anyway, this is pure speculation on my part but given the same algorithms, floating-point approximations should produce better results (at least in theory) because they more precisely approximate real numbers than do integers.
-GL