question about SPs

its been bugging me, so here goes. how come a part like the 9600GSO, which only has 96 shader processors (SPs) trade blows with a part like the Radeon 4670 which has 320 SPs? Granted the 9600GSO has more memory bandwidth, (192-bit vs. 128-bit), but the Radeon is clocked higher. If the nVdia part can do that much on just 96 SPs, why dont they just stick on more, like what ATI is doing, and kill off the competition performance-wise?

It's comparing apples to oranges. Different architecture. ATI's SPs are each specialized to do a type of calculation, while Nvidia's SPs are multipurpose.

ATIs procesors are slightly more efficient than nvidias. But they are MUCH different in design. To guestimate the relative power, you have to divide ati's number by 5 first. I don't know the exact science. This is a *LOOSE* formula.
ATI/5 = X(1.25)

ATI 320/5=64 64x1.25= 80~ Nvidia shaders

Then you have to take into account clockspeed and the differences in the memory archetecture. This round, ati seems to to handle memory much better than NV. This explains the 128 bit bus on the 4670.

This is not exact science, just a thumbs rule kind of thing.

While that's true in theory, Nvidia architecture end up being physically larger and more complex, and larger die means higher manufacturing cost. It's a matter of how much you can fit in the same die. ATI's approach gives them more profits.

As for memory bandwidth, consider this:

Use this formula for memory access bandwidth/capability:
(ddr# - 1) x single channel clock = effective clock
(Bus bandwidth / 8) x effective clock in ghz = overall capability

Let's do a bit of calculation:

4670 = 2 x 1000mhz = 2000mhz -> 2ghz
9600gso = 2 x 900mhz = 1800mhz -> 1.8ghz
4850 = 2 x 993mhz = 1986mhz -> 1.986ghz
4870 = 4 x 900mhz = 3600mhz -> 3.6ghz
gtx280 = 2 x 1107 = 2214mhz -> 2.214ghz

4670 = (128/8) x 2 = 32 gb/s
9600gso = (192/8) x 1.8 = 43.2 gb/s
4850 = (256/8) x 1.986 = 62 gb/s
4870 = (256/8) x 3.600 = 115.2 gb/s
gtx280 = (512/8) x 2.214 = 141.7 gb/s

It's a combination of bus width and effective memory speed. Imagine it as 2 rivers, one is narrower, but flows faster, the other is wider, but flows slower.

^^9600GSO uses a 192 bit bus.

(192/8)x1.8= 43.2 gb/s
my GS had a 2000mhz clock.
(192/8)x2= 48 gb/s

Isn't that exactly what a unified design was supposed to get rid of?

Each SP was supposed to be able to do anything, but ati just said they will put a butload of all kinds of things on the chip to process everything where Nvidia's design said here is my shader, it can do ANYTHING.