Anywhere from 1x to 16x performance. It really depends.
More things can be accelerated for AltiVec than other 'SIMD' units since it has a more complete instruction set. Also it has the permute operator which is useful in various non-SIMD operations.
Yes, I remember Altivec was once described as MMX done right! It's not appropiate for some things but very useful for others, any DSP (read FFT or wavelet) algorithum is going to get a massive boost from Altivec. So If you are into audio or video processing you are going to be very happy.
Unfortunately the G4 has a now aging bus structure which means the Altivec is going to be starved for data, the PPC970 should fix this and Altivec will get to show it's true colours.
The most important factor is, of course, the ability of the programmer.
Intels' Floating point numbers are pretty respectable these days and this is largly due to them using the SSE2 unit to do floating point calculations. There is a clever trick called auto-vectorisation which can find simple segments of code that can be converted to SSE instructions and since the SSE unit is much better than the standard x86 FPU it gets much better numbers.
Apple are doing a lot of work on GCC so I imagine they are putting something similar in. That means any PPC OS can benefit from it. Anything even vaguly complex however is best done by hand.
BTW anyone know who did Altivec?
Most seem to assume it's a Motorola thing but thats not the case. It was a joint project between IBM Motorola and Apple. Apple did the Architecture.
