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

[iiiankiii]

He may be saying you need to disable the visual effects as well..

everything disabled. There was no change. I'm only using 1080p on this card. I haven't pushed above 3.5gb in gaming. In that sense, I haven't felt the performance hit (if any). I'll push the games toward that limit to see if there is a major hit.

 

[Gloomy]

Hm, has somebody a GTX980m?!
This card has the full 256bit interface but only 12SMM. If there is a hardware problem the GTX980 would only use 75% of the memory with full speed - so around 3GB/6GB.

This is an interesting idea, good thinking.

 

[cmdrdredd]

Thanks to the OP, there seems to be something to this.
ManuelG

I don't find any of those links reputable. As was said why didn't this come up when reviews tested 4k? Why do I not see a problem? The claim is it affects all 970 cards. This cannot be the case at all.

 

[sontin]

So, i found someone with a GTX980m at geforce.com:

https://forums.geforce.com/default/...king-with-347-09-347-25/post/4430922/#4430922

So there doesnt exist a problem with the active SMM. His GTX980m can use all of the free memory minus the shared one.

 

[Vesku]

So there may actually be something to this?

https://www.reddit.com/r/pcgaming/comments/2s2968/gtx970_memoryvram_allocation_bug/

https://forums.geforce.com/default/...tx-970-3-5gb-vram-issue/post/4430032/#4430032

 

[ShintaiDK]

So there may actually be something to this?

Not from the tests we have done with GTX970 and GTX980.

However I am sure it takes some time to debunk with proper explanation of how the OS and other apps influence VRAM usage.

Just look at this thread. Certain people refuse to accept any edvidence against them. And the goalpost keeps changing.

A good example is these 2 screenshots.

in the first one you can claim my GTX980 is "broken". And in the second one it works. Yet its the exact same card. (Just with different VRAM allocation by apps.)

And here it goes completely wrong. obviously a 768MB card.

 

[Vesku]

Is there a program that does something similar to Nai's test for GCN Radeons? Curious if I will see a slowdown in the last few 100MBs of allocation. I'm hesitant to simply discount the experience of the Skyrim high res texture folk, some of them spend a lot of time getting their installs just right.

In the meantime I'll just wait to see the results of Nvidia looking into it.

Edit: If MSI Kombustor Memory test 2GB vs 3GB is similar to that Nai test, it may indeed be people misinterpreting behavior caused by VRAM being allocated for other things like OS. When Kombustor tries to use absolutely all the RAM on my 7950 it takes a big performance hit.

 

[NomanA]

A good example is these 2 screenshots.

in the first one you can claim my GTX980 is "broken". And in the second one it works. Yet its the exact same card. (Just with different VRAM allocation by apps.)

And here it goes completely wrong. obviously a 768MB card.

Of course VRAM usage on the GPU under test will influence the result. That's why most of the tests were done with the DGPU as headless, and an IGPU connected to the display.

At this point, there is no confusion about this issue. With GTX 980 and 980m, you don't see any hit to bandwidth. With GTX 970, there's a severe hit and it happens after 3250 MB. What's not clear is why it's happening (especially because GTX 980m is unaffected). Hopefully a driver fix will be one possible solution.

 

[ShintaiDK]

For anyone wanting to try:

You need to download this. Choose custom install and only choose the toolkit part.
https://developer.nvidia.com/cuda-downloads-geforce-gtx9xx

rec.exe can be downloaded here:
http://shintai.ambition.cz/rec.zip

 

[ShintaiDK]

Of course VRAM usage on the GPU under test will influence the result. That's why most of the tests were done with the DGPU as headless, and an IGPU connected to the display.

At this point, there is no confusion about this issue. With GTX 980 and 980m, you don't see any hit to bandwidth. With GTX 970, there's a severe hit and it happens after 3250 MB. What's not clear is why it's happening (especially because GTX 980m is unaffected). Hopefully a driver fix will be one possible solution.

I havent seen any clean GTX970 test showing this. And I dont have any issue showing both cases with a GTX980 without being headless or using IGP.

Here is a few more cases on why people just run things without any knowledge and right away assumes the worst.

And a GTX970:

 

[mikk]

Most people don't understand that they can't get solid results by using their dedicated as primary in this window application. This is the problem. GTX 980 can use all of its 4 GB in full speed while GTX 970 can use only 3.5GB in full speed in this CUDA bench. This is fact. For consistent and best possible results without WDDM influence boot up with your iGPU so that your dedicated Nvidia is loaded with 0MB VRAM.

 

[NomanA]

Most people don't understand that they can't get solid results by using their dedicated as primary in this window application. This is the problem. GTX 980 can use all of its 4 GB in full speed while GTX 970 can use only 3.5GB in full speed in this CUDA bench. This is fact. For consistent and best possible results without WDDM influence boot up with your iGPU so that your dedicated Nvidia is loaded with 0MB VRAM.

It's not even 3.5GB. The cutoff point is between 3200 and 3328MB. The bandwidth usage for that chunk is not reduced to the lowest because some of access in that 128MB is still at full speed.

I suspect, that threshold is at 3264MB.

[Edit: Sorry bad math. I did a 13/16 x 4GB, but that's actually 3328MB. And besides, it's pointless to do that math, when we don't know the reasons. In any case, the point stands that the threshold is somewhere around 3300MB, as shown by the test results]

 

[96Firebird]

Why do these tests always downclock the memory to 6GHz?

 

[ShintaiDK]

Why do these tests always downclock the memory to 6GHz?

Its a CUDA thing I think.

 

[HurleyBird]

@ShintaiDK,

At this point I think it's extremely unlikely that everyone running the test on GTX 970's is just failing to set it up correctly, especially given that the instructions for proper usage are there for everyone to see. Until someone posts results from a GTX 970 that is able to use its entire pool of VRAM the safe assumption is that there is indeed something wrong. Of course it's still healthy to take things with a pinch of salt, but the time for flat out denial is now long past.

 

[ShintaiDK]

@ShintaiDK,

At this point I think it's extremely unlikely that everyone running the test on GTX 970's is just failing to set it up correctly, especially given that the instructions for proper usage are there for everyone to see. Until someone posts results from a GTX 970 that is able to use its entire pool of VRAM the safe assumption is that there is indeed something wrong. Of course it's still healthy to take things with a pinch of salt, but the time for flat out denial is now long past.

Im just saying if I dont take special care. My test "fails" as well on the GTX980 in this benchmark. The 6GB Titan user did as well for that matter.

I know this benchmark is obvious the new hot thing. But it doesnt show anything in terms of VRAM usage. Because it doesnt account for whats already used.

 

[NomanA]

Once again, the test didn't fail, rather you didn't run it properly. You can get false positives from any test.

*All* GTX970 tests when using this tool properly are positive for this behavior. None of the GTX 980 tests show it. One or two GTX 980M tests don't show it either. There is no doubt right now that 970 VRAM bandwidth is getting hit very very hard around 3200-3300MiBytes.

 

[PPB]

Once again, the test didn't fail, rather you didn't run it properly. You can get false positives from any test.

*All* GTX970 tests when using this tool properly are positive for this behavior. None of the GTX 980 tests show it. One or two GTX 980M tests don't show it either. There is no doubt right now that 970 VRAM bandwidth is getting hit very very hard around 3200-3300MiBytes.

So in practical terms, the 970 has got the "x86 OS RAM limit" treatment, with little over 3.2GB as really performant memory, after that bandwith goes down the drain.

Dat $330 killer GPU was just too good to be true.

 

[cmdrdredd]

This still doesn't show anything the OP said to be true. The GTX 970 CAN use all 4GB of it's memory and the claim that it isn't a 256bit memory controller.

 

[ShintaiDK]

Once again, the test didn't fail, rather you didn't run it properly. You can get false positives from any test.

*All* GTX970 tests when using this tool properly are positive for this behavior. None of the GTX 980 tests show it. One or two GTX 980M tests don't show it either. There is no doubt right now that 970 VRAM bandwidth is getting hit very very hard around 3200-3300MiBytes.

I showed you a GTX980 test that did both. In this thread a person also showed perfectly fine 4GB usage in Far Cry 4.

But its all dismissed isnt it?

 

[RampantAndroid]

Has anyone tried this headless (did you Shintai?) What if it's run on 970s in SLI - the second 970 should not have any used VRAM on the desktop, right?

 

[Silverforce11]

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

This is ridiculous..
Their staff also have the problem.

"While the card is still a performance champ, it makes it hard to recommend right now, as it’ll never reach its full potential. If you have money burning a hole in your pocket and you’re aching to buy a new card, you may just prefer to opt for the 980, which isn’t affected by this. Or you could, loathe as I am to say it, wait to see what AMD has up its sleeve."

You know something is messed up when typical NV shill sites say such blasphemous things.

 

[Hitman928]

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?

 

[NomanA]

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.

 

[tential]

So how many GTX 970 owners will be willing to pay to upgrade to a GTX 970 Ti?
$360 MSRP