General video card performance question

[ioni]

I was just looking at the GPU-Z results between a 560 ti, 560 ti 448, and the new 660. I'm trying to understand why each successive one outperforms the previous based on the results. The 560 ti has a much higher pixel fill rate than both and the ti's texture fill rate is far higher than the ti 448's. So what makes the 448 so much better than the vanilla ti? Does the shader count matter that much? Or is it that even though one show as being able to draw faster, its bandwidth can't keep up with that draw speed?

 

[RussianSensation]

1) Higher memory bandwidth on the 560Ti 448
2) More tessellation performance since GTX560Ti 448 is a cut down GF110 chip similar to GTX570/580 chip and the chip in GTX560Ti is a GF114.

The tessellation / geometry performance for Fermi architecture is tied to the Polymorph Engines:

GF110 has 14 of them
GF114 has 8

This doesn't show up in the GPU-Z, but will show up in games that use Tessellation (Batman AC, Crysis 2, Lost Planet 2, HAWX 2, STALKER: COP, etc.)

3) 1.28GB of VRAM on the 560Ti 448 helps in some games that exceed 1GB

As far as comparing to 660 to 560 in GPU-Z, you have to be very careful doing that (especially you cannot compare AMD and NV cards without running real world tests on their pixel and texture fill-rate to support the GPU-Z differences). You can't really compare texture and pixel fill-rate performance across 2 different architectures without comparing the real world throughput vs. theoretical throughput.

For example, take a look at HD6970 vs. HD7970 Pixel Fill-rate, tied to the ROP throughput:

"Theoretically AMD can do 32 color operations per clock on Tahiti, which at 925MHz for 7970 means the theoretical limit is 29.6Gpix/sec; not that any architecture is ever that efficient. In practice 7970 hits 13.33Gpix/sec, which is still well short of the theoretical maximum, but pay close attention to 7970’s performance relative to 6970. Even with the same number of 32 ROPs and a similar theoretical performance limit (29.6 vs 28.16), 7970 is pushing 51% more pixels than 6970 is.

http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review/26

Therefore when you are comparing GTX660 to GTX560, you are really comparing the efficiency of 2 different architectures. If you simply compare GPU-Z across Kepler (GTX660) and Fermi (GTX560), you are assuming the efficiency is exactly the same for 2 different GPU architectures.

Basically, the units that make up the graphics card (shaders, texture units (TMUs), ROPs, geometry/tessellation units and memory bandwidth) all work together in a particular GPU architecture. They have some theoretical maximum but it doesn't mean the card can ever hit it in the real world. That's why comparing these specs on paper in GPU-Z is often meaningless.

 

[ioni]

Thanks, good info. I guess I should just stick with bench marks when comparing cards and forget about their advertised specs.

 

[KompuKare]

Thanks, good info. I guess I should just stick with bench marks when comparing cards and forget about their advertised specs.

I wouldn’t go that far. Real word performance is #1 of course but the theoreticals can give you a good insight.

This is especially important with new Nvidia kit where their PR people send out press-kits and stuff: for instance in the recent teaktown leaks of GTX660TI performance was fairly impressive until you looked at the AA performance. Looking at the underlining architecture (mainly the 192-bit memory bus and the performance of GK104 Kepler at high resolutions in general), that was no surprise.

But I would be fairly certain that when it comes to official reviews, the press kit which Nvidia will issue with the 660TI will suggest games and settings which will not show the card tanking with AA. Of course proper hardware sites will use their own selection of games and settings but there a few sites out there (no names) which operate more like a newspaper quoting press-releases for their news.

So, yes look at benchmarks but some understanding of the specs and the theoreticals is useful

 

[RussianSensation]

ioni, you can use the GPU-Z as a guidance and as KompuKare stated, it can be handy, but imo, specifically when comparing the same architecture/generation of GPUs. Otherwise, you start getting into complicated assumptions about efficiency of architectures.

For example HD5800 series used VLIW-5, HD6900 used VLIW-4, HD7900 uses Graphics Core Next 1.0, GTX400/500 used Fermi, GTX670/680 use Kepler.

Comparing specs on paper in GPU-Z between all of those cards is akin to comparing which car has faster acceleration in the quarter mile only looking at the horsepower and torque and ignoring the weight and gearbox ratio, etc. It makes it very complicated.

But if you are comparing GPUs from the same architecture / family tree (GTX660Ti vs. 670), it can tell you a lot. Context is key.

Just to give you an idea, go ahead and compare specs of GTX285 vs. GTX560Ti. Looking at their texture and fill-rate theoreticals doesn't tell us which card is faster. Clearly, GPU-Z fails to answer that question.
http://www.gpureview.com/show_cards.php