Um, you've probably noticed this in benchmarks already.. but the 9400GT probably is a step down from a 7900GS in performance, although it's got a larger featureset.

@Loosely Coupled

Perhaps AMD have a fused multiply-add instruction?

They don't release a lot of data to work from, but it's feasible.

Well, if we measure the core count of Larrabee in the same fashion AMD count the shaders on the radeon line R600 and newer.. it has

32*16 shaders (each 'core' has a 16 operand vector unit, AMD has 160*5 shaders, Nvidia have 240 single operand shaders)

So it's 512 "shaders"/Operands @ 2000Mhz? (that's a number I saw floated for Larrabee at least)
vs 800 Operands at 1010Mhz (newest batch of 4890's)
vs 240 @ 1476MHz (GTX285)

So the peak theoretical throughput is

Larrabee = 1024000
Radeon 4890 = 808000
GTX 285 = 354240

Obviously there are other things going on than the raw theoretical pixel shader throughput, for one thing, it's much easier to fill the pipes on the GTX285 because you don't have to keep things in groups of 5 or 16 to keep the pipes saturated.

the closest thing to the 24" HS-IPS panels that Apple are using is the only other thing I'm aware of that uses them, which is the NEC 2490WUXI

and THOSE are like $1100

Ruben, CUDA exists on OSX

http://www.nvidia.com/object/cuda_get.html

Needs 10.5.2 or later, and there it is. (hit the dropdown box)