Quote from ruin:
Are vector processors good or something? 143Mh X 13? Aint THAT much...
It depends on what you're trying to do.
A lot of mathematical - ie computable - problems can be expressed as a matrix, or as vectors. Vectors and matrices are closely related mathematical objects - you can deal with them differently, but they are deeply linked at a level that lowly first-year-uni-mathematicians like me never learnt.
Some problems, expressed as a vector or a matrix, are very quick and easy to solve, because of the mathematical properties of matrices and vectors. Other problems are easier to solve in the standard algebraic form. Other kinds of problems are well suited to be expressed as partial differential equations ... and you get the idea. Different horses, different courses.
Except that there are very large areas of the physical world which just sing and dance mathematically when expressed as matrices or vectors. But, alas, your average processor is more or less serial in nature. If you look at a nice clean assembler code like MIPS, you can see that most instructions act on one, two or three pieces of information.
This makes working on matrices very slow, and vectors slow to compute as well. It turns out that this is on the order of O(n^3) complexity, that is, the computing time rises cubically with the size of the computation. The upshot is that big physical experiments run painfully slowly on standard ISAs like the x86 or MIPS, because they more or less run through things element by element, more than once.
Enter vector processors. These little beauties can do all the neat things that a serial processor can do, except that they do them to matrices and vectors. So you use them to add, subtract, multiply, divide, remainder and logically compare vectors and matricies. They do it at an effective O(n) speed, which gives them a linear, rather than cubic increase in delay as the size of the problem increases.
To compare:
Let's have four vectors, of 1 element, 2 elements, 3 elements and 4 elements.
The vector processor takes a linear time-hit for each increase in vector size. Assuming that each increase takes one processor cycle, we can see that the first vector will take 1 cycle, then 2, then 3 and then 4. No problems,
The serial processor's load, however, rises cubically, becoming: 1, 8, 27, 64 ...
You can see why, in the large areas where matricies and vectors are expressive, a vector processor would be attractive. Note that very large chunks of physics are either vector-like or matrix-like in nature, all the way from Newtownian kinematics down to Quantum Chromodynamics.
Traditionally supercomputers were pure vector machines: the Crays were for as long as Seymour Cray was alive. Fast forward to the creation of Beowulf Linux, and massive clusters of cheap serial systems became the rage, mainly because they could make quite dent in the problem for a lot less than the prevailing vector systems.
However vector systems have made a very strong comeback in the Earth Simulator, a supercomputing project in Japan. This massively parallel vector system is approximately five times faster than anything else in the world, mainly because it moved away from the American trend of massive clusters towards a pure vector model.
I note also that the PowerPC chips gained on-die vector coprocessing with the G4, at least from Motorola ("AltiVec" or "VMX"). Now that IBM has created its own implementation of AltiVec, Apple will probably move away from embedded-market-focus Motorola to high-end-breeding IBM for chip supplies. Also, in cluster super-computing, for centres that cannot afford to buy POWER4-based systems, the PowerPCs may form a very attractive middle-ground of affordability, size, heat dissipation and performance on serial and vector tasks.
Anyway, that's the story about vector processing.