GTX 970 flaw

Status
Not open for further replies.
TaintedSquirrel said:
Yeah my HDD's have "expected speeds" as well but that doesn't mean I always get it.
They also advertise capacity in base 10 which means I lose a few hundred GB in formatting for Windows.

We oughta sue those bastards.
Click to expand...

Already happened, and the won they lawsuit from what I remember.

edit: I think it was back in 2001-2002?
 
I have a new 970 and this talk of class action lawsuit is just plain dumb. I remember when the ATI 5870 came out and there was this issue of the card getting a gray screen.It was a driver issue as far as I know that resolved itself. There was a thread here and people were going on and on about class action lawsuit.There never was any lawsuit that I'm aware of. Say what you want but the card does have 4GB of ram. I'm not saying the claimed issue does not exist but I haven't experienced the issue myself even when playing Far Cry 4 which is a pretty demanding game. That said hopefully a bios update when fix the issue somewhat.
 
Last edited:
I"ve always thought lawyers encourage class action suits

They are the only ones who make money anyway
 
Stoly said:
I"ve always thought lawyers encourage class action suits

They are the only ones who make money anyway
Click to expand...

Yeah. I got included in one (without knowing about it). I received a letter that I could claim my award. It was something less than 50 cents. I read the fine print. The law firm which brought the suit got the vast majority of a double-digit millions payout. I refuse to be party to any class action suit in which the lawyers/law firm make more than any other individual.
 
hottubjohn said:
I have a new 970 and this talk of class action lawsuit is just plain dumb. I remember when the ATI 5870 came out and there was this issue of the card getting a gray screen.It was a driver issue as far as I know that resolved itself. There was a thread here and people were going on and on about class action lawsuit.There never was any lawsuit that I'm aware of. Say what you want but the card does have 4GB of ram. I'm not saying the claimed issue does not exist but I haven't experienced the issue myself even when playing Far Cry 4 which is a pretty demanding game. That said hopefully a bios update when fix the issue somewhat.
Click to expand...

Not sure on the bios update. Others mentioned it was likely akin to the disabled ROPs in the 970s so it is a hardware problem. Again just speculation that looks accurate enough.
 
AMN3S1AC said:
The official specs state the bandwidth that is to be expected. But that seems to drop by 75% when using the last 512MB.

Nvidia GTX 970 Official Specs

GTX 970 Memory Specs:
7.0 Gbps Memory Clock
4 GB Standard Memory Config
GDDR5 Memory Interface
256-bit Memory Interface Width
224 Memory Bandwidth (GB/sec)
Click to expand...

Well, if the card can't hold 224gb/sec bandwidth using all 4gb as both are noted on the official product specifications; then this is false advertising and definitely looks like this would have legal grounds for being brought to court.
 
EdZ said:
I don't see any huge backlash. Once the weekend is over and tech sites can run more comprehensive FCAT benchmarks, and the real world performance impact is revealed to be minimal, only the same few with stuttering due to unrelated issues (i.e. the "I turned all the settings up the their highest possible level and it's not performing well, it must be due to this thing I read on a forum!" crowd) will continue to moan about it.
Click to expand...

Yeah i want to see some FCAT data before I consider the issue as important or not.

I got mine 970 only few days so I could have avoided it but 980 was 200 euro more expensive so I'd still get 970 even if it was advertised as 3,5GB card.
 
I rather sue for coil whine issues instead of this 500mb less vram. Big whoop. My ears are more important to me. Glad I sold these junked out 970s for $280 each.
 
seanclayton said:
I rather sue for coil whine issues instead of this 500mb less vram. Big whoop. My ears are more important to me. Glad I sold these junked out 970s for $280 each.
Click to expand...

Why not both?

The VRAM speed issue seems easily reproducible enough, so that's easy. The coil whine issue is very real also, though probably needs some expert analysis to have this problem reproduced with 100% accuracy every time.
 
ocellaris said:
What a tool, who the hell values their time so little they go hit up Nvidia chat support and quote the Nai Benchmark to troll them asking for a refund? Get out of the house kids.
Click to expand...

This, and not only that but he then asks for an upgrade to a GTX 980 when Nvidia's rep just told him there is no official statement or resolution yet determined. Toolbag, indeed.
 
GoldenTiger said:
This, and not only that but he then asks for an upgrade to a GTX 980 when Nvidia's rep just told him there is no official statement or resolution yet determined. Toolbag, indeed.
Click to expand...

You mean the chat rep is not authorized to give us all 980's?!? ;)
 
I am an owner of an 970, and have been watching this issue develop. I would like to ask why this site is not reporting on this issue, when many other sites are???

I thought that the [H] staff would be all over NVidia, on behalf of their readers... Hopefully their silence means that they are working on an in-depth article and analysis of this bug/feature. Us little people have no power over NVidia, but maybe the [H] does.
 
Stimpy88 said:
I am an owner of an 970, and have been watching this issue develop. I would like to ask why this site is not reporting on this issue, when many other sites are???

I thought that the [H] staff would be all over NVidia, on behalf of their readers... Hopefully their silence means that they are working on an in-depth article and analysis of this bug/feature. Us little people have no power over NVidia, but maybe the [H] does.
Click to expand...

I think the weekend is causing the slow response from the hard crew.
 
EdZ said:
That is exactly what is occurring at the moment. Then only time that last 500MB segment will be used in active operation is if a specific command will need to use more than 3.5GB to execute (e.g. you have two 1.2GB megatextures you need to blend into a third 1.2GB megatexture for a total operation usage of 3.6GB). Needless to say, this is a very uncommon occurrence. Otherwise, the 3.5GB is the portion of VRAM that is being used, and that 500MB is used to hold data that either isn't being used in an operation (e.g. cached textures that can be moved into the 3.5GB portion for active use) or data that doesn't require very high access rates (e.g. physics calculation results).

We also don't actually know the real drop access rate to this 500MB portion. The CUDA benchmark everyone is running does not actually access that portion of vRAM, and instead is paging to system memory (hence the bandwidth to that portion looking suspiciously close the the PCI-E bus bandwidth), because it was never designed to test this scenario.
Click to expand...

Reposting this from the other thread over in the nV subforum. Some food for thought perhaps?
 
So when the memory shows that lower bandwidth does it mean just that portion or the whole memory starts to exhibit the lower bandwidth once >3.5Gb is achieved?
 
trick0502 said:
And, why doesn't the 980 do the same?
Click to expand...

The question that remains unresolved is how to fill all 4 GB on a 970 with a CUDA program. NVIDIA claims that third party programs that query available memory on the 970 will receive the value 3.5 GB. NVIDIA says that this returns 4 GB on the 980 because on that card they're not managing a segmented memory system behind the scenes like the 970 is.

On older architectures and older versions of CUDA Nai's program would probably just bomb. But with CUDA 5.0ish (?) they introduced a programming feature called "unified virtual memory" (unified memory between GPU and CPU) which allows a developer to allocate memory in a way that is transparent to the developer. (In older versions, you would have to explicitly allocate memory on the GPU, and then explicitly copy your data from system memory to the GPU (and vice-versa when you're done processing). This made some of the coding a little "verbose". With UVM this process is condensed into fewer steps, but provides no performance benefit.) It appears that he allocated 4 GB and 3.5 of that went onto the GPU and 0.5 went onto system ram. This could be a bug in his code or a true CUDA limitation that needs to be fixed.
 
trick0502 said:
And, why doesn't the 980 do the same?
Click to expand...

That wasn't the point of the quoted post. I was trying to illustrate that it may not be possible to test the true bandwidth of the last 500MB of the vram on the 970 due to inherent segmented memory as well as driver tweaks for the 970. It is entirely plausible that the driver only presents 3.5GB to Nai's benchmark, and anything after that it simply starts to swap between system memory instead of actually accessing the last 500MB, hence the highlighted part where EdZ says the bandwidth "[is] suspiciously close the the PCI-E bus bandwidth".
 
After reading the 2 threads on this it's obvious that we don't know exactly what the deal with this is in terms of affecting gameplay but it does appear that Nvidia lied about the specs on the 970. Nvidia's admission that the memory is segmented and citing average framerates to prove that there's no stuttering issue after the whole FCAT thing is fairly damning IMO though.

I am getting a kick out of all the Nvidia fans doing their best Leslie Nielsen impersonation but I'm hoping this will make some people rethink the whole brand loyalty thing.
1gjdAX7.gif
 
Hmmm, It's looking more and more like the 970 is actually some kind of defective die harvest of the 980. I just can't see why the card cannot see or use the last 512MB of memory.

It makes me wonder what us 970 owners are going to be left with when the new console ports start coming over to the PC, as they will have higher VRAM requirements than most of the current games. You have to admit 512MB (13% Approx) of memory is a fair chunk to loose from your video card, especially if it results in games that stutter and generally just fuck up while playing, then the 970 is going to be total garbage in 6 months time, so maybe it's no wonder nVidia was selling them cheap!!!

My feeling is that the likely fix, if a fix is at all possible, would be a BIOS that re-enables something that nVidia disabled (if it is something that can still be enabled, and is not just a defective part of the chip!), which may result in a card that could butcher the sales of the 980, or they have to admit that the 970 is only equipped with 3.5GB of usable memory, and take the bad feelings this would cause towards them on the chin. I feel that nVidia does not really care about non-corporate customers, so are more likely to do nothing.

This 970 is my first nVidia card since the TNT days, and I now feel like I should have waited to give AMD a chance, instead of falling for all the positive reviews on this card. I have to say that I was already thinking of ditching AMD, because I really don't think they have much time left, let alone the fact that they don't seem to have any money for R&D anymore, so after reading the [H] review in particular, I decided to pull the trigger... Now I'm feeling like I might have been cheated, and really hope that this can be solved in a positive way, and not with bullshit from nVidia.
 
Last edited:
but it does appear that Nvidia lied about the specs on the 970
Click to expand...

well its obvious nvidia is goining to ignore the issue, people tend to whine much but at the end of the day thy do very little.

The only thing at this point to influence nvidia is a lawsuit

As for me owning two 970 cards, let me just say this: Next person who claims how much superior nvdia is to amd (and vice versa) i say to you: "Mister,you are an idiot."
 
Stimpy88 said:
It makes me wonder what us 970 owners are going to be left with when the new console ports start coming over to the PC, as they will have higher VRAM requirements than most of the current games.
Click to expand...

No, they won't. Graphics performance of the PS4 is about equivalent to a GTX 750 Ti. The only case in which those games use more than 4GB of VRAM on the PC is when you have them running at settings current consoles can't even dream of.

You have to admit 512MB (13% Approx) of memory is a fair chunk to loose from your video card
Click to expand...

Who's saying it's lost? You think it doesn't work at all?

"When a game needs less than 3.5GB of video memory per draw command then it will only access the first partition, and 3rd party applications that measure memory usage will report 3.5GB of memory in use on GTX 970, but may report more for GTX 980 if there is more memory used by other commands. When a game requires more than 3.5GB of memory then we use both segments."

The reason a lot of people in the witch hunt think their GTX 970 will only use 3.5GB of VRAM is that they have no idea how high the game settings have to be to use that much VRAM.

especially if it results in games that stutter and generally just fuck up while playing
Click to expand...

It doesn't cause stuttering or glitching. Anyone who thinks that has failed to understand the problem.

It also doesn't cause the extreme slowdown you see in NAI's benchmark. See EdZ's post below.

Now I'm feeling like I might have been cheated, and really hope that this can be solved in a positive way, and not with bullshit from nVidia.
Click to expand...

It's easy to solve. Attempt to reproduce the problem on your card. See what settings you need to run between 3.5GB and 4GB of VRAM on the games you're playing now. You'll figure out who is bullshitting who then.
 
Stimpy88 said:
Hmmm, It's looking more and more like the 970 is actually some kind of defective die harvest of the 980.
Click to expand...
We know this from the start, even before release. It's the same GM204 core with an extra 3 SMs disabled.
I just can't see why the card cannot see or use the last 512MB of memory.
Click to expand...
It can see it. It can use it. It just preferentially allocates from the first 3.5GB whenever possible.

The reason the Nai benchmark is seeing such a dramatic effect is because of the way it allocates memory. It will claim a chuck of vRAM that it assumes corresponds to a certain GPC, then test access to it. this works fine, up until that last 500BM.
The problem is the benchmark never relinquishes the vRAM it has allocated itself. The GPU does not want to allocate that last 500MB without clearing out the first 3.5GB, and because the benchmark reports itself as still using that vRAM it cannot shuffle the contents into the 500BM portion to make room in the 3.5GB portion, so it starts paging to system RAM.
The only time this would be a problem with a game would be if the programmers forgot to release vRAM no longer in immediate use. However, if a game did have this flaw, it would quickly fill up vRAM in ANY scenario because it would continue to load textures but never unload them. This is known as a 'memory leak'.


The worse case scenario to come out of this would be for Nvidia to remove the current segmentation to appease irate internet people. This would make the Nai benchmark work fine, but would make allocating ram for everything else a massive pain for developers. At the moment, Nvidia handles keeping active use memory in the 3.5GB segment and shuffling data in and out when needed. If the segmentation were top be removed, developers would need to handle this themselves. Existing applications would not do this, so would need to be patched. It would overall be a monumentally stupid idea.
 
EdZ said:
The reason the Nai benchmark is seeing such a dramatic effect is because of the way it allocates memory. It will claim a chuck of vRAM that it assumes corresponds to a certain GPC, then test access to it. this works fine, up until that last 500BM.
The problem is the benchmark never relinquishes the vRAM it has allocated itself. The GPU does not want to allocate that last 500MB without clearing out the first 3.5GB, and because the benchmark reports itself as still using that vRAM it cannot shuffle the contents into the 500BM portion to make room in the 3.5GB portion, so it starts paging to system RAM.
The only time this would be a problem with a game would be if the programmers forgot to release vRAM no longer in immediate use. However, if a game did have this flaw, it would quickly fill up vRAM in ANY scenario because it would continue to load textures but never unload them. This is known as a 'memory leak'.


The worse case scenario to come out of this would be for Nvidia to remove the current segmentation to appease irate internet people. This would make the Nai benchmark work fine, but would make allocating ram for everything else a massive pain for developers. At the moment, Nvidia handles keeping active use memory in the 3.5GB segment and shuffling data in and out when needed. If the segmentation were top be removed, developers would need to handle this themselves. Existing applications would not do this, so would need to be patched. It would overall be a monumentally stupid idea.
Click to expand...

This is ridiculous. "The card can only use 4GB if it is actually not using 500MB of that 4GB."

If it has to free up 500MB to use 500MB, it might as well not have 4GB at all.

It should be able to use all the RAM in one single context. Do we need to be playing 2 games at once to allow it to use all 4GB?
 
Last edited:
dandragonrage said:
This is ridiculous. "The card can only use 4GB if it is actually not using 500MB of that 4GB."

If it has to free up 500MB to use 500MB, it might as well not have 4GB at all.
Click to expand...

Please stop embarrassing yourself, you have no idea what you are talking about.

Segmented memory and data caching/preload - learn what it means.

This has nothing to do with brand loyalty and everything to do with clueless forum mobs raising a shitstorm over an issue which they do not understand at all.
I am glad there are at least some intelligent people on this forum, unlike others where there are 50+ pages of useless speculation only to find out that Nai's benchmark (as communicated by Nai himself!) is not suitable to adding any value to this discussion at all.

I would like to see Nvidia add FCAt testing to their official response, although I suspect the 970/980 performance will be similar again (resulting in probably roughly the same perf loss).
 
The question is not whether you can explain what is happening by telling me it's segmented memory. The question is whether or not that is an acceptable design in the first place. And most of us seem to think it is not.
 
dandragonrage said:
The question is not whether you can explain what is happening by telling me it's segmented memory. The question is whether or not that is an acceptable design in the first place. And most of us seem to think it is not.
Click to expand...

I think the majority of people are more inclined to think Nvidia somehow tried to hide this and are making a big deal about that. Wake up, engineering = compromises, even more so when consumer pricing/markets are taken into account.

Why did it take this long for anyone to notice? Why did none of the reviews of the 970 find this? Does this really affect performance? - Honestly without corresponding frametime data it is impossible to say.

Is this simply a case of some users having unrealistic expectations and running 4k downsampled, 8x aa, ultra high, and the GPU being too weak? That is not a memory system design problem, that is the case of a workload where the GPU is just too slow, no matter if it had 12GB of low latency vRAM.
 
dandragonrage said:
This is ridiculous. "The card can only use 4GB if it is actually not using 500MB of that 4GB."

If it has to free up 500MB to use 500MB, it might as well not have 4GB at all.
Click to expand...

This is signature worthy.
 
This has been an interesting topic lately. Some of the comments are pretty damn amusing. I don't own any 970s myself.
I think as long as my GPU(s) were playing the games I play fine, I wouldn't sweat it.
 
schoolslave said:
I think the majority of people are more inclined to think Nvidia somehow tried to hide this and are making a big deal about that. Wake up, engineering = compromises, even more so when consumer pricing/markets are taken into account.

Why did it take this long for anyone to notice? Why did none of the reviews of the 970 find this? Does this really affect performance? - Honestly without corresponding frametime data it is impossible to say.

Is this simply a case of some users having unrealistic expectations and running 4k downsampled, 8x aa, ultra high, and the GPU being too weak? That is not a memory system design problem, that is the case of a workload where the GPU is just too slow, no matter if it had 12GB of low latency vRAM.
Click to expand...

When I play Shadow of Mordor at 1080p (no 4k dsr etc), I run fine for 20 minutes with 70-90 fps and the vram hovering around 3.5.

Then, the vram creeps up a little as I play and goes into the 3.6 - 3.8 range. I begin the notice lag, stutters and hitching. I look at the on screen fps and see that around 80 fps is still being shown.

When I am 3.5GB or lower, smooth gameplay. More than 3.5GB, same fps but stutters.

I have the same vram/fps/stutters relationship in other games too.
 
Status
Not open for further replies.
Back
Top