970 memory allocation issue revisited

Welp, if you guys aren't happy with your card then nows the time to return it lol.

 

That was a great laugh, LOL.

It's funny seeing how the 290x matches and even outdo the 970 in frametimes :flip2:

 

Bad Nvidia, bad! If i knew that 970 had only 3,5GB of RAM, well... i would buy it anyway.

 

Seems to me that some of the Ancient Gurus here owe an apology to Nai.

About 40 pages back, there are repeated and insistent posts that Nai's benchmark was flawed and had no value whatsoever. So much so, that they even convinced Nai himself that it was flawed.

Now in retrospect we can see that his tool was right. Maybe not flawless, but it showed a genuine flaw, and has been proven to be correct.

He found a genuine thing. It may or may not be an actual problem, but his goal was just to explore an interesting slowdown.

I'd like to thank Nai for bringing this to fore, because it has made it clear that I need to wait on my upgrade.

 

Test shows, that Watchdogs is unplayable in UHD on the 970 and 980 (anyone expected anything else?)

Frametimes chart also suggests that even in full HD using about 3,5 GB the game is also not very playable (about 30 FPS on AVERAGE?)

And the frametimes for the 290X look worse than the 970 because it has way more spikes even if the spikes are slightly lower. Also the 970 did not use over 3536 MB but just 3474 MB and thus should not have the problem right ? How did they measure which parts of VRAM where occupied in the tests ?



Really would like to see some real tests that states how much VRAM is really used, what else uses VRAM at the same time, how much got swapped to RAM at the same time, in which areas the used VRAM is allocated and with what tool this was controlled.
GPU-Z is not a tool for something like that because it does not show the processes and it does not show the usage accurately (hell, it just reads most of the stuff from a database )
The game engine outputs are out of the question because they are not accurate enough (I do not even think they are really aware of how much VRAM is even still available)
Process explorer from MS does a better job there, I only have the feeling it does not really show absolutely everything that does occupy VRAM, because I am pretty sure that the drivers already occupy quite some part of VRAM just after booting but so far have not found anything that shows it.

Bottomline....all tests are not accurate enough and the tester only finds what he was expecting to find. A more scientific approach should be done.

 

I tried to tell them that Nai's convinced nvidia to reveal something wrong about 970 and yet they were not (like they had more IQ than nvidia)

 

Seems like Dying Light is a very heavy vram using game:



Yet, the "3.5gb" 970 ahead of the 6gb Titan at 4k :



:banana: :nerd: :banana:

source: http://gamegpu.ru/action-/-fps-/-tps/dying-light-test-gpu.html

 

And the 2gb 690 beats the titan at 1600p.

 

I've always thought GPU power trumps vram in all but the most extreme or isolated situations.

 

Because Titan(X,Y,Z) is more for professional graphics,and after that for some kind of games.
Titan(X,Y,Z) is for Cuda,a little brother of Quadro,if you know what I mean.

And of course if you wanna 4K get 2 -3 cards in SLI/CF.
Old cards (780TI-290X) still performs well in SLI/CF.

 

I should of been more clear (didn't quote the link), I was referring to the GeForce forums, not here.

 

For me things are simple right now, the 970 it´s an amazing card with great performance at a decent price.
And although i consider that many people are exaggerating their "performance issues" by expecting 2 970 in SLI to run perfectly fine at 4K with everything maxed out, in the end Nvidia ****ed up bigtime by lying like this and they are the ones to blame. There´s no excuse to the fact that it took 4 months and lots of complaints for Nvidia to come forward and recognize that they made a "mistake"... Not to mention that still they haven´t issued an official apology to everyone you bought a 970.

So for me they should compensate everyone that bought a 970 with some sort of freebie or offer an set up program to trade the 970 for a 980, for those who desired it.

And make sure this doesn´t happen again...

 

soo its the same song a few month later besides nv bs'in about the spec's

jajajaja you said clusterscrew

 

Thanks. So no real dramatic performance drops, the drop scales nicely with that of a 980. Hope we can all move on now.

 

For what reason exactly?
Tests show there is nothing wrong with the card.
And even if they lied about the ROPS, they did not promise a 980 performance or did they?

Nvidia did a quite few mistakes in the past and allways very reluctantly admitted them, they do not have a history to "compensate" for wrong hopes, they do not even compensate when they really screw up (see the Nvidia solder problem).

 

Great post. I noticed that it was mostly 970 owners that tried to bad mouth Nai.

 

Well, technically, it is thanks to Nai's benchmark and VultureX modification that nvidia finally admit GTX970 hardware 'issue'. Even though the benchmark is technically 'flawed' since the final 0.5gb chunk isn't being read properly.

What important is that nvidia finally acknowledge the memory segmentation has resulted in some stuttering issue and they are currently investigating it.

I read at techpowerup Nvidia will release new re-tuned driver to address some of problem caused by memory segmentation soon.