I keep on hearing this "Nvidia's L2 cache is a novel concept" line. I'm pretty sure it's out of a press release because it's there verbatim on every Nvidia plug I see. Just for the record, it isn't novel. ATI's 4800 series has 128 KB of L2 and their 5800 series has 256 KB. Furthermore cache isn't very important in a GPU, as GPU's work mainly on large data sets. That's why a GPU's speed is, in part, so dependent on the memory speed.
PS I thought you guys were experts?

The nvidias fermi L2 cache is unified for all shader clusters while ati-s L2 cache is divided to 4 for all shader clusters (and its much smaler) for your expertise.
And without cache with several hundred GB/s agregated bandwith the shaders would stall hundreds of clock cycles with just memory acces. Those large datasets are computed paralel trough each pixel many times.

I'm aware Zool. My point is that a) Nvidia's cache hierarchy may be unified, but having L2 isn't unique to their designs, and b) although cache is important, increased cache sizes won't necessarily lead to increased performance... You still need a high-speed memory link.
Also, I think you mean "for your information"

Actualy the unified L2 cache will be very efective in GPUGP for tesla cards where on 768KB data all shader cores can work equaly.
But it seems it wont be even near the agregated bandwith of the divided L2 cache in ATI cards becouse of this. And for graphics paralel nature of work trough each pixel the unified or divided cache should give litle advantage. Mainly bandwith counts there.
Thats the reason why all gpus for graphic usualy worked quite well without unified cache. (gpu-s used caches all the time)
Also 512 shaders accesing one 768KB L2 cache is quite more complicated than ati-s or previous nvidias designs.

So, basically you just proved that it's not novel concept after all.