AMD heeft een reaktie gepost op AnandTech's artikelen over de K7 en 3DNow vs Katmai New Instructions:
Ga'ash: "The second major improvement is that KNI can perform single instructions on multiple DOUBLE PRECISION floating point values. While double precision is not used that much in 3D games, the ability to do double precision SIMD style will virtually eliminate uses of the x86 FPU. (It is almost obsolete, since KNI can do almost everything the regular FPU can do, faster)."
AMD: No public disclosure of KNI has ever described such a thing. For example, KNI has not been reported to implementing a 2x64 bit SIMD mode. So I believe this is incorrect. Besides, if this were true it would probably be the most important improvement since it would totally obsolete the x87 FPU in all cases.
I contend that this is not true.
AMD: "This allows game developers to use the 8 64bit FP registers entirely for MMX, to optimize all their integer coding. This is impossible with 3DNow."
Ga'ash: MMX can always be used with 3DNow! The fact that the registers are shared doesn't seem to have affected any loops that we have come across. One thing that you fail to point out is that MMX instructions can be used to hack the single precision floating point format (for tricky BSR simulations, changing sign, exp/log simulations, masking etc.) Its not clear that KNI will have the same opportunities available to it