R700 has too many shaders and not enough ROPs

[MrMilli]

4830 doesn't have 40 texture units but 32. (shader units and texture units are coupled)
Furthermore, with every generation ATI chips everybody always claims that 16 rops is the bottleneck. Since R520 the performance has increased >6x while ATI still uses 16 rops. Why? Because a rop isn't something as simplistic as most people think. Their performance and efficiency goes up too. What's more important is that ATI's and nVidia's rops aren't comparable. So their amount isn't either comparable.
There's also difference in efficiency (and latency) in GDDR3 and GDDR5.
And performance doesn't scale linear because there are many interconnects and cache structures inside these chips (with limitations ofcourse). So basically all your calculations are void.

 

[daehxxiD]

A few days ago I read an article that stated ATI shaders are less complex than NVIDIA ones, therefore ATI simply adds more shaders to make up for this, and it worked fine in the 4000 series.

Chris

Afaik it's the other way around actually. Nvidia uses 1-D shaders, whilst ATI uses 5-D shaders. The 800 Shaderprocessors Ati lists on it's HD4850/70 are actually "only" 160 Shaderprocessors... The HD3850/70 had 64 (5-D) Shaderprocessors respectively.

 

[lemonadesoda]

So basically all your calculations are void.

LOL. Never mind the explanation, the empirical evidence is clear.

 

[W1zzard]

LOL. Never mind the explanation, the empirical evidence is clear.

empirical would be if a rv770 card with 16 rops and 24 rops (+50%) were compared and there was a substantial (50%) increase in performance. anyone seriously believe that is possible?

 

[lemonadesoda]

No. We are talking scaling with shaders. We dont need to empirically test inreased TMU or ROP, just the shaders.

You have the bitmask W1z. So why not take a 4850 on 800, 640 and 480 shaders and run COD4 and GRID at 1920x1200 single and crossfire always same clocks, and i'll redo the analysis.

The only criticism of the OP analysis is the scaling of the benchmarks due to different clocks. But i dont have a 4xxx card to test, or i'd do it. Anyone want to lend me a 4850 crossfire setup for a week?