970's having performance issues using 4GB Vram - Nvidia investigating

Mei · 25 Jan 2015 at 17:31

yeh i read that, i thought it was a very specific test tho
and not sure how much it applies to

Kaapstad · 25 Jan 2015 at 17:34

CAT-THE-FIFTH said:
Its quite simple - the GTX970 does not have the fully enabled memory controller resources the GTX980 has it appears.

TR were the first site to notice something was bit off last year:

http://techreport.com/blog/27143/here-another-reason-the-geforce-gtx-970-is-slower-than-the-gtx-980

TBH,its not like the Nvidia second tier top cards have not had slightly less RAM or bandwidth than the top end ones before. I don't think it will make much bearing on whether you would buy a GTX970 or not TBH.

I would say TR have missed the obvious that the 980 has got a lot more silicon driving the fill rate and all that graph shows is that difference.

The other thing that stands out a mile is the 780ti on there, the 970 outscored it easy which is not bad for a faulty card.

Silent_Scone · 25 Jan 2015 at 17:36

CAT-THE-FIFTH said:
Its quite simple - the GTX970 does not have the fully enabled memory controller resources the GTX980 has it appears.

TR were the first site to notice something was bit off last year:

http://techreport.com/blog/27143/here-another-reason-the-geforce-gtx-970-is-slower-than-the-gtx-980

TBH,its not like the Nvidia second tier top cards have not had slightly less RAM or bandwidth than the top end ones before. I don't think it will make much bearing on whether you would buy a GTX970 or not TBH.

Did you read the explaination he received at all?

Mei · 25 Jan 2015 at 17:38

you need a tool/benchmark for testing just the last part of the memory
if it is a lot slower its only going to show up in a game as random dips & stutters, and they could happen anytime randomly not all the time ...because most the time it will be reading from the 3.5gb part

i think you know this tho right?

Kaapstad · 25 Jan 2015 at 17:41

Mei said:
you need a tool/benchmark for testing just the last part of the memory
if it is a lot slower its only going to show up in a game as random dips & stutters, and they could happen anytime randomly not all the time ...because most the time it will be reading from the 3.5gb part

i think you know this tho right?

I suspect that it runs at the same speed as the rest of the VRAM as we are talking about a small part of the memory accessing a small part of the bus.

If there is a problem it could be with shifting large chucks of data around as this could be slower.

LoadsaMoney · 25 Jan 2015 at 17:43

Is the fix just to turn your settings down a bit then, so you don't go over the 3.5GB, and have to access the other 0.5, which is tanking performance ?

CAT-THE-FIFTH · 25 Jan 2015 at 17:48

Kaapstad said:
I would say TR have missed the obvious that the 980 has got a lot more silicon driving the fill rate and all that graph shows is that difference.

The other thing that stands out a mile is the 780ti on there, the 970 outscored it easy which is not bad for a faulty card.

Its primarily a memory bandwidth driven benchmark. Its why the newer cards like the R9 285 and GTX960(he says with very baited breadth) actually do well in that benchmark as they incorporate memory compression techniques which older cards like the GTX780TI don't have.

Its not the first time a 70 series card has less memory bandwidth or useable RAM than say the 80 series card. Accept the GTX970 as a 3.5GB card with an addtional 512MB Turbomemory built in(yes.I know its not the same).

On top of this the GTX660 has something which ends up having a similar effect- the first 1.5GB is accessed faster than the last 512MB,and its one reason the HD7870/R9 270 was faster with more AA piled on. However,I already knew about before I bought it though.The GTX660TI did the same thing too.Like I said if you want a GTX970 its not going to suddenly make you stop getting one.

bloodkil · 25 Jan 2015 at 17:48

This makes me glad to have 290X's at this point. Very annoying for anyone with the new Nvidia cards, a bit naughty of them to mis-sell them like that, if that's how to say it.

Mei · 25 Jan 2015 at 17:49

LoadsaMoney said:
Is the fix just to turn your settings down a bit then, so you don't go over the 3.5GB, and have to access the other 0.5, which is tanking performance ?

maybe but its pretty hard to prove there is even a problem without a better tool ><

Gregster · 25 Jan 2015 at 17:49

SoM tested on a pair of Titans and a 3930K.

1440P fully maxed out (including the HD pack) VRAM being used/cached/whatever was 5984MB

I am uploading a video to show what stutter I see and hopefully will help some out.

Legend · 25 Jan 2015 at 18:04

Gregster said:
The GTX 970 has 1664 Cuda cores
The GTX 980 has 2048 Cuda cores

That is where it has been hobbled and from what I can see, nothing more. The memory reading is done differently on the 970 over the 980 because of the differing SMs and fewer crossbar resources on the memory system. It is all explained in ManuelG's post but I guess you don't want to see that answer and want something else instead.

Thanks for the info (and the condescension), but I missed his post... every time I leave this thread for any length of time and come back there's another 3 pages to read through!

Dirk Diggler · 25 Jan 2015 at 18:11

ICEx21 said:
Hi Gregster, could you please test your titans in SoM in the area of Nurn please? Specifically in the first part where there's slightly more grass and a you can fight about 3-4 orcs, the second area I noticed it in is in the Orc village in the North I believe.
Both times I noticed when vram went over 3.5gb I noticed a lot of stuttering.

In Dragon Age Inquisition I noticed it in the Hinterlands specifically the woods where you fight both mages and templars.
Other games I have noticed this in are Ass Creed Unity but like someone else has said this isn't the best example hence me testing it out on the other games.

I'd be happy to be proven wrong as I feel a little gutted in getting £600 worth of cards and would much rather prefer that I'm doing something stupid that's causing these issues than have to return the cards as otherwise they are awesome and the software feels much better to use than AMD's.

System specs if they help are the following:

4790K at stock
16gb G.Skill 2400Mhz I believe
OcUK Reference 970's SLI
Superflower 1000w Platinum psu

Thanks

Try disabling SLI and check those areas again. It could be an SLI issue?

Gregster · 25 Jan 2015 at 18:13

Just for reference on stutter and might confirm what others are seeing (or not seeing).

Some games can have stutter (pretty much all UBI games of late) and this could make it look like VRAM issues but is just a problem on the game engine.

Gregster · 25 Jan 2015 at 18:17

Legend said:
Thanks for the info (and the condescension), but I missed his post... every time I leave this thread for any length of time and come back there's another 3 pages to read through!

Sorry for coming across as blunt and apologies. This thread just seems to have a few people who are missing what is what or reading things completely differently to how it actually is and trying to find problems where there aren't any.

Mei · 25 Jan 2015 at 18:22

Kaapstad said:
I suspect that it runs at the same speed as the rest of the VRAM as we are talking about a small part of the memory accessing a small part of the bus.

If there is a problem it could be with shifting large chucks of data around as this could be slower.

dont know you lost me there

dont know what the size matters, its the amount it is accessing it?

Mei · 25 Jan 2015 at 18:30

Gregster said:
Just for reference on stutter and might confirm what others are seeing (or not seeing).

Some games can have stutter (pretty much all UBI games of late) and this could make it look like VRAM issues but is just a problem on the game engine.

lol yeh which is probably why so many are complaining, it doesnt mean there isnt a problem with the memory but i bet very few have the same problem huh

Final8y · 25 Jan 2015 at 18:43

Abwx said:
To the contrary, that s quite possible when looking at the GPU diagram, the memory controllers that are connected to the clusters have 16 bits of addresses, on a total of 64, that are indeed not used since they would feed disabled units, and the simplified MCs cant swap the data from the 0.5GB partitions RAM to feed a nearby unit, actually it s quite possible that those 0.5GB are not functional at all even if they can be addressed, that is you can fill it with datas but you have no mean to send those datas to functional units because the design lacks the relevant crossbars.

Abwx said:
Yes you can load 4GB but only 3.5 can be sent to the SMMs, the data in the remaining 0.5GB cant be sent for execution to the SMMs used for executing the datas that are within the 3.5GB, it s only a theory but the GPU design and the presence of a separated partition point to this implementation, what transpire is that each SMM has a fixed adress space in the RAM and that data meant to be executed by a given SMM must be retired in the relevant adress space, if this given SMM is disabled there s no mean to send the data to another SMM, so as said the whole 4GB is adressable but only 3.5GB can be executed by the GPU computing units...

http://forums.anandtech.com/showthread.php?t=2416150&page=18

murah · 25 Jan 2015 at 18:50

Why are people testing on UBI games? I thought it was common knowledge around here that they make some of the most buggy unoptimised mess of code in the industry

Mei · 25 Jan 2015 at 18:53

Final8y said:
stuff

interesting i bet!, wish i understood half of it lol

randal · 25 Jan 2015 at 18:58

So the general rule here is that:

If NVidia use a cut down version of their GPU resulting in fewer shader modules, and thusly fewer crossbar points into the memory system then they have to divide the memory into directly accessed (high speed) and indirectly (SLOW speed) chunks?

Meaning that unless NV have a major rearchitecture then this is going to be the case on anything that isn't a flagship GPU?

If so, that's a hardcore own goal.