On the buses

[Kaapstad]

I got asked this in another thread so lets find out.

Kaap, have you tried comparing say your GTX690 0xAA vs SuperSampling and Titan 0xAA vs SuperSampling and see the the performance hit/reduction in % term? Pretty sure the 690 would take a far heaviler hit (assuming if it can handle SuperSampling that is) than the Titan.

Software used - Heaven 4 @1600p
Maxed settings with 8XAA
Maxed settings with 0XAA

GTX 690 running on a single core (GTX 680) @stock
3960X @4.5

With 8XAA

With 0XAA

Single Titan @stock
3930k @4.5

With 8XAA

With 0XAA

GTX 690 increase 8XAA to 0XAA = 39.75%
Titan increase 8XAA to 0XAA = 46.24%

Well that is not supposed to be in the script, the above figures mean that the 690 performed better than the Titan when turning off the AA. The bigger the % difference is between on and off the worse the card is performing and the bigger the hit when using 8XAA.

Having said that these figures are very close to each other and only one bench has been used.




Now just for interest as it is a totally different architecture and behaves totally different on the Heaven 4 bench here is a stock 290X doing the same tests.

Single 290X @stock
4930k @4.5

With 8XAA

With 0XAA

290X increase 8XAA to 0XAA = 28.7%

 

[Marine-RX179]

Kaap, thanks for doing this. However it look a little confusing, and I'm not sure which FPS figures you used for the calculation. Could you may be put the results like this for example:

GTX680
0xAA: ?fps
8xAA: ?fps

Performance hit going from 0xAA to 8xAA: -?%

and so on?

Judging from your posts, you seem to be comparing the 8xAA to 0xAA, but what I had in mind was that 0xAA is the "original performance", and by applying AA how much of that original performance is lost (performance hit).

Also when I stated that I was expecting the GK104 to get bigger performance hit than the GK110, I wasn't so much thinking about the bus-size "specifically", but more about the huge memory bandwidth difference

 

[Kaapstad]

Kaap, thanks for doing this. However it look a little confusing, and I'm not sure which FPS figures you used for the calculation. Could you may be put the results like this for example:

GTX680
0xAA: ?fps
8xAA: ?fps

Performance hit going from 0xAA to 8xAA: -?%

and so on?

Judging from your posts, you seem to be comparing the 8xAA to 0xAA, but what I had in mind was that 0xAA is the "original performance", and by applying AA how much of that original performance is lost (performance hit).

Also when I stated that I was expecting the GK104 to get bigger performance hit than the GK110, I wasn't so much thinking about the bus-size "specifically", but more about the huge memory bandwidth difference

I am using the points scored in Heaven for the Calculations

Using the GTX 690 results 675 is a 39.75% increase over 483.

The bigger the percentage increase is, the more effect using AA has.

To do it your way is also very easy to work out using the same numbers

GTX 690
0XAA = 675 = 100.00%
8XAA = 483 = 71.55%


Single Titan
0XAA = 1205 = 100.00%
8XAA = 824 = 68.38%


Single 290X
0XAA = 991 = 100.00%
8XAA = 770 = 77.7%

 

[TheRealDeal]

I would go for a different benchmark as i think tesselation performance will show badly for the 690. A more game type bench would show the gtx690 in a better light fps wise. After all, in just about every bench it is faster but in this it's slower. To me that's already a false result even though the percentage show's in the gtx690's favour against the Titan.

 

[Yaayuh!]

I would go for a different benchmark as i think tesselation performance will show badly for the 690. A more game type bench would show the gtx690 in a better light fps wise. After all, in just about every bench it is faster but in this it's slower. To me that's already a false result even though the percentage show's in the gtx690's favour against the Titan.

It's slower because it's actually a GTX680.

 

[Kaapstad]

I would go for a different benchmark as i think tesselation performance will show badly for the 690. A more game type bench would show the gtx690 in a better light fps wise. After all, in just about every bench it is faster but in this it's slower. To me that's already a false result even though the percentage show's in the gtx690's favour against the Titan.

I disabled a core on the GTX 690 to keep it to single GPU comparisons and also avoid SLI influencing the results.

It's slower because it's actually a GTX680.

+1

And it actually performed better than the Titan when turning on the AA.

 

[Marine-RX179]

And it actually performed better than the Titan when turning on the AA.

It's strange, cause for the SleepingDogs bench with Extreme AA (SuperSampling), the GK104 based card performance tanks hugely and do noticably work than the 7950/7970, while the GTX780s don't have too much problem holding their own against the 290...

 

[TheRealDeal]

It's slower because it's actually a GTX680.

Yea missed that part. Reading failure on my behalf.

 

[Yaayuh!]

GTX690 4 EVA <3

 

[oweneades]

Realitically you need far more than a single benchmark test for this.

To start with an actual game would be useful (even if that meant using a built in benchmark).

 

[Kaapstad]

It's strange, cause for the SleepingDogs bench with Extreme AA (SuperSampling), the GK104 based card performance tanks hugely and do noticably work than the 7950/7970, while the GTX780s don't have too much problem holding their own against the 290...

I don't think my GTX 690s do too bad on the Sleeping Dogs bench using max settings.

4 GPU setups

1. Score 216.2, GPU nvTitan @981/1789, CPU 3930k @5.0 Kaapstad Link 337.5 drivers
2. Score 187.8, GPU 290X @1220/1500, CPU 4930k @4.8 Kaapstad Link 14.6 drivers
3. Score 186.7, GPU 690 @1160/1800, CPU 3960x @4.9 Kaapstad Link 337.5 drivers
4. Score 98.9, GPU 590 @613/855, CPU i7 980x @4.2 Kaapstad Link 320.49 drivers
5. Score 85.2, GPU 5970 @725/1200, CPU i7 975 @4.2 Kaapstad Link 13.8 drivers.

Those 4 x 290Xs are sweating a bit though, they are only 1fps in front.

As has been said above though when I get time I will do more benches as one is not enough.

The other thing to be careful of is making the distinction between testing bus efficiency and running out of VRAM.

What I mean by that is a 290X @4K on BF4 maxed with Mantle will run out of VRAM and things stutter badly with poor performance but none of this means there is anything wrong with the 512bit bus, it just means that there is not enough memory.