NVIDIA Responds to GTX 970 3.5GB Memory Issue

[SteveGrabowski]

Reading through the same thread on anandtech, it seems like this is a flawed benchmark just accessing system memory instead of the other 500GB pool for the last 500GB, and hence the slowdown. Lots of people there are testing with 4GB of VRAM usage and their 970's are running fine. 22GB/s is right there in line with DDR3 1600 bandwidth.

[Majestic]

@SteveGrabowski Well I don't really have a game that hits 4GB. How can I otherwise test it?

[Lewellyn]

How people can play devil's advocate right now is beyond me. It's not getting the advertised bandwidth nor is it able to use the advertised memory (effectively). And it's by design, you are totally ok to rip on them and defending them on this is pretty irredeemable. This is not buyer's remorse talking, I still have 14 days to return it without costs. 

 

How LTT members can jump on a fear mongering bandwagon without full conclusive evidence doesn't surprise me.

[Fate]

@SteveGrabowski Well I don't really have a game that hits 4GB. How can I otherwise test it?

Maybe you could use DSR on a relatively intensive game and crank the settings up as high as possible (still might not even hit nearly 4GB though). 

[KeltonDSMer]

You are assuming that the Pcper article is in fact, correct on that non-cited "Nvidia Response". Why would Nvidia respond only to Pcper, and no other site? It is fishy at best.

 

 

I thought this was posted earlier, but Scott Wasson released the same NV statement 5 hours ago: http://techreport.com/news/27721/nvidia-admits-explains-geforce-gtx-970-memory-allocation-issue

[MageTank]

I thought this was posted earlier, but Scott Wasson released the same NV statement 5 hours ago: http://techreport.com/news/27721/nvidia-admits-explains-geforce-gtx-970-memory-allocation-issue

Yeah, i just spent my time digging through 50 pages of absolute garbage to find a response from Nvidia. I retract my previous statements regarding the authenticity of the information provided in this thread. However, looking at the information Nvidia is also providing, is there any performance loss involved? Their test is showing little to no loss in performance when using that 0.5gb partition when compared to that of the 980, which seems to be immune from these complaints.

 

Source of the Nvidia post made by ManuelG: https://forums.geforce.com/default/topic/803518/geforce-900-series/gtx-970-3-5gb-vram-issue/post/4432672/#4432672

 

It is recent, but it does confirm that the statements made in both articles were in fact made by Nvidia, as he is considered Nvidia staff.

[deathjester]

It just baffles the mind how the fuck they thought noone would notice. I just can't even right now.

I'm not even mad at myself, noone could've known this.

 

I am more concerned how this was not noticed by benchmarkers/reviwers in the first place.

 

We saw certain games that used a lot of VRAM like Shadows of Mordor where the AMD cards were killing the benchmarks, and those were REFERENCE AMD cards that throttled LOL.

 

The review industry is a joke for the most part. The GTX 970 was sold as the greatest card in the history of mankind, with no negatives. The bias shown by some sites is downright disgusting, to the point they will often use a outdated driver to skew results on tests in favor of one company. 

 

We also saw benchmarks like this where downsampling/supersampling in game settings DESTROYED Nvidia performance and no one stopped to wonder why, when that is their job as a reviewer. Hell this is the ONLY site I see that even bothered to benchmark some of these graphics settings, and they actually used AMD cards people buy.

 

Look at the higher supersampling (click 2x2 supersampling) benchmarks and a 280x is above a GTX 970 in Ryse. Ryse uses the Crysis Engine and I expected to see extensive benchmarks everywhere for this game, especially when the "holy grail" of PC Gaming, Star Citizen is going to use the Crysis engine. What did I see? Nothing. Complete silence from most of the big outlets on it and or failing to even test IN GAME SETTINGS that were present in the game.

 

http://www.pcgameshardware.de/Ryse-PC-259308/Specials/Test-Technik-1138543/

[ViruZ_HUN]

So it is 3.5 GB usable now? I thought it was only 3.

 

Either way, I played with a GTX 550 ti for almost 4 years, which is a 1 GB card. SO I guess 3.5 GB is enough for me:D

[kpreg]

So it is 3.5 GB usable now? I thought it was only 3.

 

Either way, I played with a GTX 550 ti for almost 4 years, which is a 1 GB card. SO I guess 3.5 GB is enough for me:D

The thing is once you get above 3.5GB (and you can't stop it from doing so) performance tanks. I'd be fine with 3.5 GB normally, but since we can't cut it off it's an issue.

[BiG StroOnZ]

 

Nai is a poster in that thread just scroll down to post #4 & 7.

The link to the Guru3D forum thread is in the top left right corner. (also here: http://forums.guru3d.com/showthread.php?p=4998341#post4998341 )

 

Just read that, noticed he said he created it because he was bored. Also, he says it does the same thing on his Titan for the last 250MB,  which is interesting. As other people have been pointing out it does this on more cards than just the 970. 

 

Here is my summary as far as i have read the german posts (i am german )

 

Someone posted the appearance of the VRAM issue on the german "computerbase"-forum (Thread: http://www.computerbase.de/forum/showthread.php?t=1435408 )

Until then nobody had any "usefull" (the benchmark isn't the most usefull either) way to take a look at this problem so everybody was demanding for further investigation.

Nai, a user of this forum, offered his help by writing a benchmark in CUDA . In the beginning only GTX Titan (his own graphicscard) and upwards were able to produce reasonable results because of the code he used. 

 

 

The code he used can be checked here: http://www.computerbase.de/forum/showthread.php?t=1435408&p=16868213#post16868213

Apparently some people helped him fixing minor issues and allowing other Nvidia graphicscards, but until last time i checked nobody found any big mistakes he could have made which would explain the bad performance of the 970, so it seems to be the truth.

 

Sry for bad english

 

Ok, well I'm reading a bit of his post on the translated website. According to him, even his Titan had the same problem for the last 250MB. 

 

They might have not found any big mistakes, but it doesn't make sense that other people are benching cards different than the 970 and are getting similar results.

 

 

I thought this was posted earlier, but Scott Wasson released the same NV statement 5 hours ago: http://techreport.com/news/27721/nvidia-admits-explains-geforce-gtx-970-memory-allocation-issue

 

And if you read the bottom of the finished benchmarking they did it says the following:

 

On GTX 980, Shadows of Mordor drops about 24% on GTX 980 and 25% on GTX 970, a 1% difference.  On Battlefield 4, the drop is 47% on GTX 980 and 50% on GTX 970, a 3% difference.  On CoD: AW, the drop is 41% on GTX 980 and 44% on GTX 970, a 3% difference.  As you can see, there is very little change in the performance of the GTX 970 relative to GTX 980 on these games when it is using the 0.5GB segment.

 

So the conclusion is, it's fine. Just like the benchmarks I posted of other people going over 3.5GB of VRAM. 

[MageTank]

The thing is once you get above 3.5GB (and you can't stop it from doing so) performance tanks. I'd be fine with 3.5 GB normally, but since we can't cut it off it's an issue.

Have you personally noticed a performance loss while gaming? I am now interested in this, because Nvidia is specifically stating that the memory performance from a 970 compared to a 980 is at worst, 3% difference.

[GoodBytes]

I am seeing all this, and all I see is that is looks like to be it is a driver bug.

I think Nvidia claim could be correct, but a driver bugs makes is behave incorrectly.

I am sure Nvidia is looking into it, and we will shortly know. What I wonder, is why no one noticed it until NOW.

It's been a while the 970 is out, and it is a popular card. No one wondered why their game goes crap performance at random, despite owning this card with massive performance?

My guess is that it is a bug in the latest drivers or firmware distributed to manufactures.

All I am saying is that better investigation needs to be done. You guys should compile a list of 970 models. See if they are all affected or some only. How about the reference design of the 970? How about older drivers?

[Majestic]

Just tested at Crysis2, DSR at 4K, got to about 3.4GB, fps was at 24, frametime at 80. Noticed some jitter when walking around. As if it was going 40ms, 80ms, 40ms 80ms.

Same with Star Citizen, going over 3.5 gives me infitely high frametime spikes. Note that i hit 3.5GB at 1080p. So who says you will never use that amount is clearly wrong!

 

Either that 0.5GB isn't being addressed and it's grabbing system memory, or it is and the bandwidth just tanks. From the way it looked though, since it was so clearly edging the 3.5, that it cannot access the 0.5GB.

Maybe there is still something that can be fixed with drivers.

[KeltonDSMer]

 

And if you read the bottom of the finished benchmarking they did it says the following:

 

On GTX 980, Shadows of Mordor drops about 24% on GTX 980 and 25% on GTX 970, a 1% difference.  On Battlefield 4, the drop is 47% on GTX 980 and 50% on GTX 970, a 3% difference.  On CoD: AW, the drop is 41% on GTX 980 and 44% on GTX 970, a 3% difference.  As you can see, there is very little change in the performance of the GTX 970 relative to GTX 980 on these games when it is using the 0.5GB segment.

 

So the conclusion is, it's fine. Just like the benchmarks I posted of other people going over 3.5GB of VRAM. 

 

 

Of course I read the whole thing

 

Like I said in mt first post in this thread, I'd like to see a reviewer use NV's FCAT tool in the scenario in question. I don't form strong opinions on something like this until we (average consumers) know more information. All we have now are a bunch of people riled up and a vague chart from NV rush-released on a weekend.

[BiG StroOnZ]

Of course I read the whole thing

 

Like I said in mt first post in this thread, I'd like to see a reviewer use NV's FCAT tool in the scenario in question. I don't form strong opinions on something like this until we (average consumers) know more information. All we have now are a bunch of people riled up and a vague chart from NV rush-released on a weekend.

 

I agree. We have one benchmark tool that everyone is going by, that also happens to do the same thing with the 970 with other cards. Then we have people who are benchmarking ACTUAL games themselves going over 3.5GB and showing no issues, then there are people going over 3.5GB who are claiming they are finding some sort of issue. And then there is the chart from NVIDIA which says there's no difference at all.

[PerfectTemplar]

It's not fishy, it was already freaking proven that the benchmarks are not accurate.

[PatrikStar24]

It's not fishy, it was already freaking proven that the benchmarks are not accurate.

[Citation Needed]

[Majestic]

Huh. I cannot get my card to run over 3584MB. Which is exactly 512MB less (one memory chip). It simply pools the last 512MB with system memory, even in games.

This cannot be by design.

 

It can, nevermind. Hit 4GB with some fiddling. Absolute nightmare on the frametimes.

[spartaman64]

It's like on a mobile phone, they advertise it as a 16gb model and you only have 13/14 because that's occupied by system files, in this case they are saying 4GB but really the memory available to the user is 3.5GB

but for amd gpus they include 4.5 so you get the full 4GB

[Majestic]

It's like on a mobile phone, they advertise it as a 16gb model and you only have 13/14 because that's occupied by system files, in this case they are saying 4GB but really the memory available to the user is 3.5GB

 

No not really. Because the system thinks it's a 4GB card and some DX11 titles pool the full memory. I cannot play Star Citizen because it tries to fill the 4GB and it pools system memory.

I literally cannot play Star Citizen because the frametimes are all over the place.

 

Card is fundementally broken for that game (if you also use multiple monitors that eat up memory). A phone with less memory isnt.

[BiG StroOnZ]

 

Huh. I cannot get my card to run over 3584MB. Which is exactly 512MB less (one memory chip). It simply pools the last 512MB with system memory, even in games.

This cannot be by design.

 

It can, nevermind. Hit 4GB with some fiddling. Absolute nightmare on the frametimes.

 

Yes but increasing the graphical options and fiddling with settings to achieve 4GB of VRAM usage is going to hurt performance anyway. You would need neutral way of increasing memory usage, without hurting performance to test the theory properly. Along-side an FCAT analysis of the high memory usage situation, because that would reveal any stuttering. Also, Nai's program that people have been using is clearly bugged because I doubt every single card other than the 970 that people are testing with also have this same problem. So far these are the cards testing with Nai's benchmark that also show the same problem as the 970; 580, 680, 670, 760, 770, 780 Ti, Titan. I'm sure if I searched around I could find more included in that list, but these are just some from off the top of my head that I saw pictures of the benchmark being done. Which means, Nai's benchmark is not accurate.

[SkulD]

but technically as it is mentioned it still has 4gigs or ram and performance dip isnt that much,yes nvidia did kinda cheat but still it does have 4g's technically so stop acting like kids and calm the fuck down .... yes my comment contradicts my previous statement but after thinking about this i think we should really calm down

If this where AMDs doing i can bet a whole month salary that the flames and hate threads will be endless .

[Cainn]

Yep... Look at Nvidia put "4GB GDDR5" on the 970 when they included the .5GB that HAS to be dedicated to the system... Yeah, if AMD said that on their 290 and 290x, they'd really have 4.5GB... Also, I now know why I stutter a lot on Assetto Corsa....

Note: I still call BS on 

Running a 12 car grid at Nuremberg on max settings at 1920x1080 my Gtx 980 never goes over 2.4 gigs of vram usage. Unless your running some insane grid size or resolution, I doubt that's what's causing your stuttering.. 

[Darkman]

Running a 12 car grid at Nuremberg on max settings at 1920x1080 my Gtx 980 never goes over 2.4 gigs of vram usage. Unless your running some insane grid size or resolution, I doubt that's what's causing your stuttering.. 

Well, the more I think about it... I'm pretty sure it was a REALLY poorly made map I went on.. I don't have that issue on any other maps I have installed.

[kpreg]

Can't they release a temporary driver to limit the VRAM usage until they figure this shit out?