What is the point of an add-in physics card?

what is the point of an math coprocessor? After all, the integer processor is able to do floating point math.

the idea of the add-in physics card is to create an architecture specialised to do JUST the calculations needed for physics. If you say physics use matrix (three by three) , then your architecture is made to make matrix multiplication in one step. The architecture could be so specialized that you would need the same step to multiply two integers.

This is the reason T&L calculation moved from processor to graphic card - the chip in the graphic card was optimised to do T&L, and it soon overpowered the main processor in this task (thanks to speed, memory bandwidth and so on) that now even a quad core with three cores dedicated to T&L could not keep up.

So, the idea of the physics cards is to have a specialised unit that could make specialised operations very fast - as compared to a microprocessor.

By example, a 3x3 matrix multiplication is comprised of 9 multiplications and an addition of the 9 results. If only for physics, you could use 32-bits floats (or maybe even 24-bits floats). The math coprocessor will do at most one floating point multiplication per clock - this could do one matrix multiplication per clock.

Multiplying a matrix by a value is a 9-step operation - and would take 9 steps for the coprocessor. This could be done in a single step on a physics engine.

Now, it might be possible to move the physics calculation to the T&L unit on a graphic card - if the calculations needed fit what the T&L unit can do well, you no longer need a physics. If not, you would be better downloading it on a second core.

Now, would a second core prove to be faster than a dedicated physics card? If so, use a dual core (which will help you not only in games). If not, no. Everybody keeps to its side of the story, and it is difficult to get impartial results