I was reading through the IBM 970 initial presentation (the is a link for it somewhere in the posts previous). One thing that popped out at me was on this page... (i think page 13 of the presentation PDF).
It states that normal floating point operations (non-vectorized) can be as fast as 7.2 GFlops/sec while vectorized (SIMD) floating point operations will be able to be as fast as 14.4 GFlops/sec. If this is true, then applications will not be required to get significant improvements in speed by coding with the AltiVec only extensions. Possibly this will mean that a single DOUBLE floating point calculation would take the same time as two Altivec FLOAT floating point calcuations would take.
This would be a significant improvement in speed over the G4 processor.
It states that normal floating point operations (non-vectorized) can be as fast as 7.2 GFlops/sec while vectorized (SIMD) floating point operations will be able to be as fast as 14.4 GFlops/sec. If this is true, then applications will not be required to get significant improvements in speed by coding with the AltiVec only extensions. Possibly this will mean that a single DOUBLE floating point calculation would take the same time as two Altivec FLOAT floating point calcuations would take.
This would be a significant improvement in speed over the G4 processor.