vorko

misiak
Also I don't know why everybody is talking about 3,5GB when Nai's benchmark clearly states that first  significant drop (3x slower) occurs from chunk 25 what is from >3.2GB. So not 3,5GB but problem starts with 3,2GB!

Pretty sure that's because no one is running it headless and the ~300 MB is taken by OS

Thanks for explanation, this make sense. It is possible to run that software in headless mode ?
 
Edit: But why there is no drop when testing for 980?!

bcavnaugh

JFoul
 
 
Taken from overclock.net. 

 
What tool or software is this?

It's Nai's benchmark.

You shouldn't. It doesn't show what most people think it shows. 

vorko
You shouldn't. It doesn't show what most people think it shows. 

Shouldn't what ? What does it show ?

IMO it shows that the last 0.5GB memory is accessed very slow, this is the point, isn't it ? I don't know what to think, maybe time will show. I'm curios how this end, because we bought something which was clearly not advertised. And it does not matter it was misunderstanding between devs and PR. I don't believe it personally.

From what I've read, it looks like they are going to need to figure out a better way to utilize that .5GB of partitioned RAM and dedicate it to another process so video games can't utilize it. They might be coming up with some kind of driver or BIOS fix that will basically make the card a 3.5GB card. If they do that, it shouldn't have issues because the games won't try to utilize the slower partition. So far, that's just speculation on what certain sites think need to be done.
 
As for the whole Nai's benchmark ordeal... It's not designed properly to actually test that partitioned RAM, where the test is usually failing. However, that benchmark is what revealed that there was a disconnect between some of the VRAM and promted Nvidia to reveal that they partitioned the .5GB of the RAM. Once further tests were done, it was revealed that the partitioned RAM is slower than the bulk of the 3.5GB VRAM on the card. Which is why games start to stutter once the partitioned RAM is accessed. 

Can we take a second and imagine what it would do without the .5gb of cache/vram/partition that does run slower?

I overload with vram with Firestrike Ultra. I drop down to 8-10fps. Now, every real world case that pushes past 3.5gb has shown the Sutter and lag.. At 25-40 fps compared to 80-90fps.

That is a huge difference when going from a Flagship to a midrange gpu. For everyone that is in this ordeal, I hope they don't disable the .5gb and you'll regret it, instead of getting a stutter you will get unplayable.

RainStryke
From what I've read, it looks like they are going to need to figure out a better way to utilize that .5GB of partitioned RAM and dedicate it to another process so video games can't utilize it. They might be coming up with some kind of driver or BIOS fix that will basically make the card a 3.5GB card. If they do that, it shouldn't have issues because the games won't try to utilize the slower partition. So far, that's just speculation on what certain sites think need to be done.
 
As for the whole Nai's benchmark ordeal... It's not designed properly to actually test that partitioned RAM, where the test is usually failing. However, that benchmark is what revealed that there was a disconnect between some of the VRAM and promted Nvidia to reveal that they partitioned the .5GB of the RAM. Once further tests were done, it was revealed that the partitioned RAM is slower than the bulk of the 3.5GB VRAM on the card. Which is why games start to stutter once the partitioned RAM is accessed. 

The question is if vram is accessed randomly or continuously... If it is random access then slow part can be filled with data even if application uses only 1GB vram and then it could be disaster. But I believe the 0.5GB part is only accessed if 3.5GB part is full.... Otherwise the slow partition has to be disabled.

this is the same thing as what they did with the gtx660ti it is handled by drivers. 
 
below are two articles that fully explain how this works
 
http://www.pcper.com/news/Graphics-Cards/NVIDIA-Responds-GTX-970-35GB-Memory-Issue
 
http://www.pcper.com/reviews/Graphics-Cards/NVIDIA-Discloses-Full-Memory-Structure-and-Limitations-GTX-970
 
yes you can access all 4 gigs at once (they work as two separate units at 3.5 and 0.5)
 
NVIDIA divided the memory into two pools, a 3.5GB pool which maps to seven of the DRAMs and a 0.5GB pool which maps to the eighth DRAM.  The larger, primary pool is given priority and is then accessed in the expected 1-2-3-4-5-6-7-1-2-3-4-5-6-7 pattern, with equal request rates on each crossbar port, so bandwidth is balanced and can be maximized.
 
For the other0.5g things like compressed textures that are not as time sensitive as other material require much less bandwidth and can be moved around to other memory locations with less performance penalty
 
the above is copy/pasted from the above articles but they seem to actually explain how this works and why the performance hit is only in the 3% range while gaming.

i do not think that any application would be in charge of where to write on the ram,  this would be handled by your OS and then the drivers.
 
when I open the chrome browser this application writes to the RAM but it does decide where on the RAM to write.  this would be done by the OS and whatever hardware driver that is associated.
 
this would seem to be the same for games.  you load up a game it talks to the OS who sends to the appropriate hardware and the hardware drivers assign instructions to the correct area.  this also makes sense when you read the articles since the card is prioritized to use the first 3.5GDDR first then to open up the last 0.5GDDR if needed.  at that point since the drivers or OS would know the call order and how often used the memory is it should be able to control traffic of RAM calls pretty efficiently.  This is evidenced by only a 3% loss in performance.
 
The easiest way for me to understand or how I would solve this issue would be to make sure that really compressed data was sent to that last 0.5 GDDR.  since compressed data will be more dependent on CPU/GPU strength than GDDR bandwidth that should resolve mmmost if not all of the issue

so now all think they will fix it ??
rly rly i should buy amd card like always now im unhappy first grafic from nvidia and fail amd 0 fail if any will back graphic and they give back money pm here thank you or pm prvi

I figured I should add this test done today by guru3d showing that this is really not an issue that renders this card unplayable
 
http://www.guru3d.com/news-story/middle-earth-shadow-of-mordor-geforce-gtx-970-vram-stress-test.html
 
the truth is unless you are playing at 4k you will be hard pressed to use 4gb of memory and the gtx970 even in sli is not really what you want at 4k.  a pair of r9 290x's would be better at 4k.
 
I agree that misstating a product is terrible and I have lost some faith in nvidia but the truth is this is still a great card and it never did and still does not have performance issues.
 
if we see some kind of fcat situation that effects me at 1440p then I will be upset but I just cannot find any place where that is occurring right now.  I cannot put the settings any higher than I already have them and have playable framerates.
 

I just ran that firestrike extreme 4k test and I do not know what this is going to prove other than you get stuttering at less than 20 fps.
 
my score was a 2920 btw.

Who wants to play 4K buy R9 290x, who wants to play 2K buy 980, who wants to play 1080p buy 970. It's simple like that :)

970 plays 2k very well, except for terrible ports like Shadows of Mordor that requests 6gb of vram for ultra.

980 can't even provide that.

970 does, and has been shown on this forum, use all 4gb of vram. Like I said on other threads, hypochondriac syndrome.

NVidia is working on a driver update now, so give it a few days and see what happens.