Ten also mentioned that the KNI "will unarguably be better than 3D-Now." I think that any claims of superiority or inferiority should be measured against the range of applications the particular technology is targeting. Even more, from economical standpoint, some extra functionality may be not a plus whatsoever. The hypothetical ability of KNI to operate with double precision numbers is of no use for the 3-D transform engines. Therefore, the KNI "superiority" is very "arguable" under given economical constraints.
I agree. I have to wonder at the usefulness of double-precision SIMD in 3-D apps and games. I think Intel has suggested that KNI will be useful for more than just "mutant-blasting" 3D games, but I'm not sure what that means. But like I said before, at the very least, KNI will probably match the functionality and usefulness of 3D-Now.
On the other hand, all our talk about KNI vs. 3D-Now means very little, since those who really know what the new KNI instructions will be aren't talking. My hope is that KNI will not only be able to support double-precision SIMD, but also that KNI will be able to perform floating-point operations on four single-precision numbers in one register. This is crucial for 3D, since everything in 3D is based on the 4x4 matrix. In comparison, the K6-2 can currently achieve this by issuing two 3D-Now instructions concurrently.
It was also suggested that Katmai could have an expanded register file. That would be quite a feat to implement, unlike the relative simplicity of 3D-Now which doesn't do anything to the size of the register file.
Like I said before, all this is mere speculation. Of course, those who know the details about KNI are chosen under the "need-to-know" basis, so ... |
