WCCftech: Memory allocation problem with GTX 970 [UPDATE] PCPer: NVidia response

[Silverforce11]

I originally dismissed this thread, so just to be clear, this is the situation put forth.

1) 980 can max out 4 GB with no performance penalty.

2) 980 can be shown to take a performance hit before 4 GB if Windows/other programs are taking some VRAM.

3) 970 can max out ~3.5 GB with no performance penalty.

4) 970 can max out 4 GB but incurs a severe performance penalty above ~3.5 GB.

Is this correct?

4) Depends on the game engine, if that allocation is caching as seen in Frostbite games, it just auto-enhance LOD for distant objects, without performance degradation. Arma 3 or Skyrim ultra 4k textures is a different story, its either enough vram to load and run fine, or performance tanks.

 

[Carfax83]

4) 970 can max out 4 GB but incurs a severe performance penalty above ~3.5 GB.

This is false. The GTX 970 can easily sustain playable frame rates when VRAM goes above 3.5 GB. I posted a screenshot a few pages back of SoM using 3.6 GB+ of VRAM and getting more than 40 FPS.

Also, when I play AC Unity, VRAM usage can top out at slightly above 3.7 GB and the frame rate remains stable at 60..

So if a bandwidth decrease was truly there, then that wouldn't be the case.

 

[cmdrdredd]

I originally dismissed this thread, so just to be clear, this is the situation put forth.



1) 980 can max out 4 GB with no performance penalty.



2) 980 can be shown to take a performance hit before 4 GB if Windows/other programs are taking some VRAM.



3) 970 can max out ~3.5 GB with no performance penalty.



4) 970 can max out 4 GB but incurs a severe performance penalty above ~3.5 GB.



Is this correct?

Problem with #4 is that when I am using 3.7GB the game doesn't suddenly turn to a stutter fest and there are no noticeable performance issues. Playing far cry 4 at 1440p and smaa uses about that much and no problem. Still 60fps averages if not better.

I am willing to bet the drops in bandwidth are a specific cuda issue and unrelated to gaming.

 

[Hitman928]

4) Depends on the game engine, if that allocation is caching as seen in Frostbite games, it just auto-enhance LOD for distant objects, without performance degradation. Arma 3 or Skyrim ultra 4k textures is a different story, its either enough vram to load and run fine, or performance tanks.

Ok, gotcha, makes sense. What about the whole reported VRAM usage in games versus actually used VRAM usage? Doesn't afterburner et. al. report how much RAM is allocated versus actually being used? E.g. in a game, even if you are "seeing" 4 GB being used, it could actually only be using 3 GB?

This is false. The GTX 970 can easily sustain playable frame rates when VRAM goes above 3.5 GB. I posted a screenshot a few pages back of SoM using 3.6 GB+ of VRAM and getting more than 40 FPS.

Also, when I play AC Unity, VRAM usage can top out at slightly above 3.7 GB and the frame rate remains stable at 60..

So if a bandwidth decrease was truly there, then that wouldn't be the case.

I imagine that even if you are using above the "limit" in a game, that it probably wouldn't tank your fps in general, but would probably cause some stuttering or other issues as stuff loads in. I don't have a 970 to play with otherwise I'd be on this. Seems like nVidia will have to respond at this point, I'll just look forward to what they have to say.

 

[NomanA]

I originally dismissed this thread, so just to be clear, this is the situation put forth.

1) 980 can max out 4 GB with no performance penalty.

2) 980 can be shown to take a performance hit before 4 GB if Windows/other programs are taking some VRAM.

3) 970 can max out ~3.5 GB with no performance penalty.

4) 970 can max out 4 GB but incurs a severe performance penalty above ~3.5 GB.

Is this correct?

Yes. And regarding (2), that's true for any graphic card. The 3.5 GB is for 3500000000 bytes, which is around 3300 MiB (1024x1024). The test tool reports in the regular MB (or MiB) format.

 

[NomanA]

This is false. The GTX 970 can easily sustain playable frame rates when VRAM goes above 3.5 GB. I posted a screenshot a few pages back of SoM using 3.6 GB+ of VRAM and getting more than 40 FPS.

Also, when I play AC Unity, VRAM usage can top out at slightly above 3.7 GB and the frame rate remains stable at 60..

So if a bandwidth decrease was truly there, then that wouldn't be the case.

The bandwidth decrease is real. However the impact on a game is hard to determine. How that chunk of memory is used can dictate performance.

 

[cmdrdredd]

Ok, gotcha, makes sense. What about the whole reported VRAM usage in games versus actually used VRAM usage? Doesn't afterburner et. al. report how much RAM is allocated versus actually being used? E.g. in a game, even if you are "seeing" 4 GB being used, it could actually only be using 3 GB?



I imagine that even if you are using above the "limit" in a game, that it probably wouldn't tank your fps in general, but would probably cause some stuttering or other issues as stuff loads in. I don't have a 970 to play with otherwise I'd be on this. Seems like nVidia will have to respond at this point, I'll just look forward to what they have to say.

I have used process explorer and it tells you the actual memory in use by the hardware. Sometimes it varies from what afterburner tells me. As for the second comment, I haven't had any stuttering whatsoever with my 970s. There are a few times when I am close to 3.8GB in use. Still chugging along with 60fps averages or better in far cry 4 at 1440p and SMAA.

The only quantifiable test that shows any issues with bandwidth is a cuda test. I am thinking that the issue is cuda and not the card not the game drivers.

 

[Silverforce11]

Ok, gotcha, makes sense. What about the whole reported VRAM usage in games versus actually used VRAM usage? Doesn't afterburner et. al. report how much RAM is allocated versus actually being used? E.g. in a game, even if you are "seeing" 4 GB being used, it could actually only be using 3 GB?


I imagine that even if you are using above the "limit" in a game, that it probably wouldn't tank your fps in general, but would probably cause some stuttering or other issues as stuff loads in. I don't have a 970 to play with otherwise I'd be on this. Seems like nVidia will have to respond at this point, I'll just look forward to what they have to say.

Yup, it would definitely depend on how that game is using the memory. It could be streaming assets so that "pop-in" textures (which occur at very short distances for AC Unity!) happen further away or less frequently, but you won't get performance reductions.

If you guys recall the tech demo in BF3, it was recommended to have 2gb cards, but we found results with 1.5gb cards at 1080p with 4x MSAA to be just fine. It turns out, 2gb cards have enhanced visuals, particular further objects are more detailed. It does it dynamically, so if you have less vram, it reduces the scene complexity automatically.

Then you get game engines that are less dynamic, the problem will instantly show itself as soon as you go above the vram barrier.

 

[Hitman928]

I have used process explorer and it tells you the actual memory in use by the hardware. Sometimes it varies from what afterburner tells me. As for the second comment, I haven't had any stuttering whatsoever with my 970s. There are a few times when I am close to 3.8GB in use. Still chugging along with 60fps averages or better in far cry 4 at 1440p and SMAA.

Ok, cool. It doesn't matter so much to me as I won't be shopping for a new card until the fall at the earliest, but it's interesting. Glad it's not effecting your games.

 

[tential]

Does anyone thing the GTX 970 used market will take a hit from this?
Might be able to pick up SLI GTX 970 for cheap.

 

[RampantAndroid]

As I said, all headless results have shown without exception that GTX 980 is fine, but 970 isn't. And since the test can't give you false negatives (unless a truly pathological behavior by the OS), *all* 970 results (headless or otherwise) show that it can't go past the 3200-3328MiByte brick wall without getting hit. If the test isn't run correctly then the drop happens even sooner on a 970.

I've seen a 980 WITH the problem, a 980m WITHOUT the problem...and haven't seen the 970 headless runs. If I missed, could you link me to them?

 

[cmdrdredd]

Yup, it would definitely depend on how that game is using the memory. It could be streaming assets so that "pop-in" textures (which occur at very short distances for AC Unity!) happen further away or less frequently, but you won't get performance reductions.

If you guys recall the tech demo in BF3, it was recommended to have 2gb cards, but we found results with 1.5gb cards at 1080p with 4x MSAA to be just fine. It turns out, 2gb cards have enhanced visuals, particular further objects are more detailed. It does it dynamically, so if you have less vram, it reduces the scene complexity automatically.

Then you get game engines that are less dynamic, the problem will instantly show itself as soon as you go above the vram barrier.

Well so far crysis 3, far cry 4, shadow of Mordor, dragon age inquisition, and battlefield 4 don't show any signs of issues for me. It is just one sample but I don't think gaming performance is really affected by any of this.

 

[iiiankiii]

I've seen a 980 WITH the problem, a 980m WITHOUT the problem...and haven't seen the 970 headless runs. If I missed, could you link me to them?

Look a few pages back. My card was headless (no display attached). Drop off at 3200mb.

 

[96Firebird]

I just played FC4 with DSR (2715x1697 with 4xMSAA) to get above 3.5GB, and while the FPS was slow, there was no hitching when VRAM usage was at 3794MB.

However, I did get some major hitching when it climbed above 4000MB VRAM usage...

Sorry for the cell phone screenshots, DSR doesn't allow normal screenshots. I wonder if they will ever fix that.

 

[tential]

Well so far crysis 3, far cry 4, shadow of Mordor, dragon age inquisition, and battlefield 4 don't show any signs of issues for me. It is just one sample but I don't think gaming performance is really affected by any of this.

Until someone proves that this lowers performance no one will do anything about it on Nvidia's end.

 

[Carfax83]

Does anyone thing the GTX 970 used market will take a hit from this?
Might be able to pick up SLI GTX 970 for cheap.

I seriously doubt it, since it doesn't seem to be affecting gaming. Although I will be happy to sell you my G1 GTX 970s when the 980 Ti becomes available, provided you live in the U.S and don't try to lowball me

 

[KaRLiToS]

So the OP was right and Nvidia admitted their issue.

http://www.lazygamer.net/general-news/nvidias-gtx970-has-a-rather-serious-memory-allocation-bug/

The problems seems to be widespread – and Nvidia’s admitted that problem affects every single 970 (to varying degrees). They’re looking in to the issue – but unless it’s something that can be resolved via driver or firmware update, a recall may be on the cards.

“We are still looking into this and will have an update as soon as possible,” said NVidia community liaison ManuelG on the GeForce forums.

It's funny that everyone accused the OP while what is happenning is really an issue on every GTX 970.

 

[96Firebird]

Wait, did the author come to that conclusion from that quote by ManuelG? Is that the "admission"?

 

[ShintaiDK]

Wait, did the author come to that conclusion from that quote by ManuelG? Is that the "admission"?

You first noticed the costant moving goalpost now?

Not to mention all the newly created users in that thread.

 

[Carfax83]

Wait, did the author come to that conclusion from that quote by ManuelG? Is that the "admission"?

Exactly, how is that an admission? Sounds like the author has an english comprehension issue..

 

[Keysplayr]

It's way good enough for some apparently. And yes, the new regs are freaking hysterical.

 

[Carfax83]

However, I did get some major hitching when it climbed above 4000MB VRAM usage....

Hitching at max VRAM usage is to be expected, as texture swapping is occurring much more frequently.

 

[Enigmoid]

As I said, all headless results have shown without exception that GTX 980 is fine, but 970 isn't. And since the test can't give you false negatives (unless a truly pathological behavior by the OS), *all* 970 results (headless or otherwise) show that it can't go past the 3200-3328MiByte brick wall without getting hit. If the test isn't run correctly then the drop happens even sooner on a 970.

The sample size of 970s is not large enough.

And yes there are false positives on the titan and 980.

Look at post #131 by ShintaiDK

 

[tential]

I seriously doubt it, since it doesn't seem to be affecting gaming. Although I will be happy to sell you my G1 GTX 970s when the 980 Ti becomes available, provided you live in the U.S and don't try to lowball me

lowballing you?
I was hoping you'd give it away heavily discounted to a fellow anandtech member, that way I don't have to lowball you.

 

[Enigmoid]

So the OP was right and Nvidia admitted their issue.

http://www.lazygamer.net/general-news/nvidias-gtx970-has-a-rather-serious-memory-allocation-bug/

It's funny that everyone accused the OP while what is happenning is really an issue on every GTX 970.

"We are looking into it" is by no means an admission.