a few nvidia cards:
GF 9500 GT GF GT 220 GF GT 240 EVGA GT 240 SC GF 9600 GT GF 9800 GT
Shader units 32 48 96 96 64 112
ROPs 8 8 8 8 16 16
GPU G96 GT216 GT215 GT215 G94 G92
Transistors 314M 486M 727M 727M 505M 754M
Mem Size 256/512 MB 512/1024 MB 512/1024MB 512 MB 512/1024MB 512/1024MB
Memory Width 128-bit 128-bit 128-bit 128-bit 256-bit 256-bit
Memory bandw 25.6 GB/s 25.3 GB/s 54.4 GB/s 57.4 GB/s 57.6 GB/s 57.6GB/s
Core Clock 550 MHz 625 MHz 550 MHz 550 MHz 650 MHz 600 MHz
Memory Clock 800 MHz 790 MHz 1700MHz 1800 MHz 900 MHz 900 MHz
Shader Clock 1400MHz 1360MHz 1340MHz 1340 MHz 1625MHz 1500MHz
The fact that the 9600 uses a different GPU entirely then the GTX 240 kind of precludes the latter from being simply a faster version of the former.
ATI 5970:
transistors: 4308
interface: PCIe 2.1 x16
max memory: 2048
Core Mhz: 725
Memory Mhz: 1000
Texture fillrate: 116
Pixel fillrate: 46.4
bandwidth GB/s: 256
mem type: GDDR5
memory Bus width: 2x256
DirectX: 11
OpenCL: 1.0
OpenGL: 3.2
GFLOPS: 4640
TDP Idle: 51W
TDP max: 294W
Features: Dual GPU solution on single PCB, Angle independent anisotropic filtering, Eyefinity
certainly a excellent card by any means.
However, is it the best? that certainly depends. Without a perfect understanding of how the GPU and various memory pipelines work with various cards, it's impossible to come to a conclusion that has no chance of being superceded when new information becomes available.
A prime example is the fact that many manufacturers "cheat" on benchmark tests, just as they cheat on WHQL certification. How do they do that? quite simple; the driver simply looks at the program that is running, if it's a well known benchmark app, it "enables dubious optimizations" to increase it's score. cheating at WHQL is even easier; they simply package the driver in an installer that sets a key to not enable any form of optimization; this prevents any sort of issue, and WHQL doesn't care about pixel fillrates or anything like that, but only wether the driver works, they get a passing score. then they package that exact same signed driver into another installer that sets a registry key to enable dubious optimizations, optimizations not tested by the WHQL, of course, so they get the best of both worlds.
Now, this practice is not something you see today, but on the other hand, it's impossible to tell. perhaps they simply made these "dubious optimizations" less obvious? perhaps when one card simply edges out another, we should consider that maybe the card that seems to perform better is really taking shortcuts.
Yet another example; back in teh early days of windows 95, when DirectDraw was just taking shape; for the most part, everything worked. but for some cards, DirectDraw would completely bork when using certain features that the video card claimed to support. The cause was in fact that the driver was ALWAYS saying it supported the feature. Basically, a certain driver function was called, essentially, "doessupportfeature" and the driver was simply returning true no matter what.
MS had to workaround the laziness of these shortcut takers. One of the DDraw dev team members took a seldom used PC on campus, generated a GUID with it, smashed the network card (thus guaranteeing that that GUID will never be generated again) and made a function that called the "doessupportfeature" function of the driver with that GUID. basically, if it returned true, then DirectDraw would go, "AHA! caught you! now I'll never believe what you say" and took a careful route that avoided any non-standard features the driver otherwise claimed to support. This had far-reaching consequences for the driver manufacturer, who, originally probably made the function in the interest of "optimization" and now finds that all the stuff they actually do support is no longer being used because directdraw has confirmed that it cannot trust the driver. So they had to rewrite the driver to actually flesh out the function to work properly.
What does this have to do with the topic? well, not a whole lot, but it goes to show that the performance of a card weighs heavily on the quality of the driver itself; a newer card can be outperformed by an older one simply because they are using different drivers.
Of course, in this case I don't think anything could outperform a ATI 5970, but given that you have to consider wether the manufacturers are *really* not cheating at benchmarks or wether they simply are able to hide it a lot better? nobody can really tell...
Also, I better mention that some of the above re: directdraw stuff was from Raymond Chen's excellent book "the old new thing".