thanks man. it's a little sloppy as i was doing it mostly to appease my own curiosity but i figured i'd throw em up here for others to peruse. i'm hoping to find something to run that can max out my VRAM while still being playable and test that as well.
-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
GeForce GTX 970s seem to have an issue using all 4GB of VRAM, Nvidia looking into it
- Thread starter Naked Snake
- Start date
- Status
thanks man. it's a little sloppy as i was doing it mostly to appease my own curiosity but i figured i'd throw em up here for others to peruse. i'm hoping to find something to run that can max out my VRAM while still being playable and test that as well.
Ok, now we're talking.. This I don't like to see at all. It means that someone knows more about this than we thought..
You're running the tests in full screen mode and then logging the results in the bg right? Just making sure the 3500 barrier isn't artificially allocated by the OS in windowed mode.
Great job!
You should post this all around reddit and the Geforce etc. forums.
It's not about that thing being an issue, it's about selling a card as 4GB while effectively only using 3,5GB... :-/
That's the sad thing.
It doesn't specify, but I'm assuming the game is running the exact same settings on both cards. The 970 not breaking 3.5GB, with the 980 regularly maxing out the 4GB. Very good visual.
Incorrect. I absolutely assume a certain minimum amount of memory in the games I write and I absolutely optimize for that min spec.
Mengy
wishes it were bannable to say mean things about Marvel
I have to wonder if the Nvidia drivers are purposely limiting the 970's to 3.5GB as much as they possibly can, but the 980 drivers don't. That would sort of prove that Nvidia knew full well about this beforehand.
Naked Snake
Member
Haven't they admitted precisely that already?
Nvidia said:
Edit: or it seems they're saying reported memory use by 3rd party applications (e.g. MSI Afterburner) isn't correct?
i'm working on this. i snagged a cheap copy of shadow of mordor but its a 44gb install. i'll test it the same way that i did with the others. my guess though, it's not gonna be anything crazy different. while it's obvious that the last .5GiB is only used when pushed, nvidia admitted to setting up the partition separately, so it makes sense it would happen. having a huge bandwidth dropoff would not be explained in the same way and there's been no evidence to support it happening other than that super flawed nai benchmark tool.
i don't know that they're saying that afterburner's reporting isn't correct, and it doesn't seem to be. they seem to be weaseling around the issue by saying that it may not in some situations. the fact that afterburner DOES report >3.5GiB usage would suggest that it is in fact correctly reading the VRAM usage and there is a plateau at the 3.5GiB mark. however, to be sure, i'd checked in process explorer, with the same results (showing both <>3.5GiB in use at different points). unless nvidia wants to come out and explain why the reporting at the plateau would be different, i'd assume it's them dodging the real issue, which is the limiting of the mem usage to the 3.5GiB plateau.
Well yeah, I should have figured out that by your sound testing methodology/results presentation. Sorry for doubting you.yup, everything's being done in fullscreen. lol, didn't mean that sloppy
chromatic9
Member
The game seemed no different when at 3.8GB at around 30fps. No way was it running on some last resort slow ram.
Obviously the 970 doesn't behave like the 980 in the COD example posted but I'm getting usage above 3.6 that's playable. I wonder if the vram reading is incorrect sometimes.. Say you start a game 3.6GB or under it will stay that way on the reading and maybe use more in the background when needed. If the game uses more than 3.6 on first loading then you get the reading above 3.6GB.
Obviously the 970 doesn't behave like the 980 in the COD example posted but I'm getting usage above 3.6 that's playable. I wonder if the vram reading is incorrect sometimes.. Say you start a game 3.6GB or under it will stay that way on the reading and maybe use more in the background when needed. If the game uses more than 3.6 on first loading then you get the reading above 3.6GB.
Objectively Bad Opinion
Member
The game seemed no different when at 3.8GB at around 30fps. No way was it running on some last resort slow ram.
Obviously the 970 doesn't behave like the 980 in the COD example posted but I'm getting usage above 3.6 that's playable. I wonder if the vram reading is incorrect sometimes..
when you change settings, restart the game so it loads with those settings and check what the vram number says
also, off-topic but that is a decidedly last-gen torch that guy in your screenshot is holding
Mad Season
Banned
Awesome bootski. Send that to nvidia, post on their Twitter and reddit
chromatic9
Member
I didn't change settings mid game
in closing
there's an issue with the cards certainly. they are most definitely under utilizing the full 4GiB of memory. however, outside synthetics, to break the 3.5GiB mark i need to crank games up to ludicrous settings i'd never use when normally playing. all in all, the issue exists but it's certainly been blown out of proportion from what i've seen. i did these tests to see if i needed to return the card and just go right to the 980 series. i still might, i have until friday to decide, but this issue would certainly not have me trading the card in for a 290x or certainly not down to a 960 eww.
if anyone knows of any games i can test, outside shadow of mordor, that runs well while utilizing tons of VRAM, i'd love to hear about it. shadow of mordor has been tested to death post 3.5GiB and it certainly doesn't seem to have any major stuttering issues from any of the videos that have been posted.
also, in testing i used downsampling. to do it properly, i had to disconnect my 2nd monitor
I both agree and disagree that this has been blown out of proportion. I agree because we don't need conspiracy theorist saying that that last 512 MB is some super slow ram. I disagree because your analysis lends evidence to the fact that someone running a game where they would need that extra half gig would have to 'fight' the GPU for it with supersampling, downsampling, copious amounts of AA or extra effects.
Other than that, your analysis was phenomenal, thanks
.Dictator93
Member
Dan Baker (Oxide games)
How so? Wouldn't all other video cards have the same issue if that were true?
SocksForWoks
Banned
Would it be fallacious to say this is what happens when you buy the cheaper option? (970 instead of the 980)
This is a hugely overblown issue to say the least, can pretty much be summed up as "Cut down version of the gtx 980 performs like a cut down version of the gtx 980", blown up into some huge 'controvesy' by a bunch of people who have no idea what they are talking about posting benchmark results from a random unreliable cuda benchmark.
The 970 is just a gtx 980 with 3 of the smm units severed.
each smm unit uses crossbars to access the memory channels, with 3 units severed you end up with increased load on fewer crossbars, which results in not being able to access all 4gb of vram at full efficiency.
As a performance optimization, nvidia segmented the vram into logical chunks of 3.5gb and 0.5 gb. This way the first 3.5gb of vram on the 970 maintains the same ratio of vram to smm units as the gtx 980, and can be accessed with full efficiency, this is why the 970 prefers to allocate a max of 3.5 gb when possible, its an optimization.
There really isn't any 'memory bug', its the way the card is designed.
This is a ~300 dollar video card, of course it isn't going to perform as well at 4k as a gtx 980, if it did there would be no reason for the gtx 980 to exist.
The 970 is just a gtx 980 with 3 of the smm units severed.
each smm unit uses crossbars to access the memory channels, with 3 units severed you end up with increased load on fewer crossbars, which results in not being able to access all 4gb of vram at full efficiency.
As a performance optimization, nvidia segmented the vram into logical chunks of 3.5gb and 0.5 gb. This way the first 3.5gb of vram on the 970 maintains the same ratio of vram to smm units as the gtx 980, and can be accessed with full efficiency, this is why the 970 prefers to allocate a max of 3.5 gb when possible, its an optimization.
There really isn't any 'memory bug', its the way the card is designed.
This is a ~300 dollar video card, of course it isn't going to perform as well at 4k as a gtx 980, if it did there would be no reason for the gtx 980 to exist.
Xdrive05
Member
A guy on the geforce forums just posted a YT video demonstrating the throttling behavior, and the memory bus juggling vram back down to 3.5. Interesting.
http://youtu.be/wgRir5JwKyU
http://youtu.be/wgRir5JwKyU
So not freaking out or acting like a chicken little is considered "corporate ball washing"?
SMH indeed.
i can confirm that afterburner reports erractic bus usage when allocating vram into the 2nd partition. shown below are the same memburner tests (1366x768, no aa) run to 2gb and 4gb.
edit: i've also done some preliminary testing on shadow of mordor running @ 1080p with the ultra hd texture pack, ultra settings. it starts off below the 3.5GiB mark and ends ~3.7GiB. the results seem to be showing the same 3.5GiB limiting as before but there were some troubling issues that i started encountering when i went above that mark (lots of tearing, stuttering, etc). i will try to reproduce tomorrow afternoon and report back.
CyanideFuse
Member
I have run the MSI Kombuster GPU memory burner tests. Starting VRAM usage was 263MB; I had closed browser and switched to best performance visual mode in Windows 7.
My setting: 1920x1080 8xAA. I ran two memory tests of:
The two segments on the graphs cover the 2048 test (purple box outline) and 3072 test (blue box outline).
My setting: 1920x1080 8xAA. I ran two memory tests of:
- 2048MB = Score of 3588 of 59FPS
- 3072MB = Score of 2693 of 44FPS
The two segments on the graphs cover the 2048 test (purple box outline) and 3072 test (blue box outline).
- 2048 test produced frame times of 16-18ms; framerate of 60 without fluctuation
- 3072 (edit to clarify: usage during test up at 3964MB according to Afterburner) test produced frame times of 24-26ms with two large spikes of 78ms; framerate of 40-45 with fluctuations; GPU usage didn't waiver
Naked Snake
Member
But 263MB + 3072MB shouldn't be enough to even run into the slower memory chunk. So I'm not sure what your test is meant to reveal.
I think if Nvidia advertised the 970 as a 3.5GB graphics card this would all be a non-issue.
I think it's a typo. Pretty sure that image has a max VRAM usage of 3900ish MB
CyanideFuse
Member
The VRAM usage was at 3964MB according to Afterburner (although Kombustor does say 3072 test; I am not sure where the additional VRAM usage is coming from).
Edit: Test I saw from nVidia forum with similar VRAM usage under this test: https://forums.geforce.com/default/topic/803518/gtx-970-3-5gb-vram-issue/?offset=978
Naked Snake
Member
Ah I see it now.
Horsemama1956
Member
Run a game with higher resolution textures/Uses for VRAM on higher settings so you don't have to pump SSAA to make it unplayable?
Mengy
wishes it were bannable to say mean things about Marvel
And that right there is the issue at hand here. I paid for a 4GB card, not a 3.5GB card. I'm okay with it performing less than a 980, but my 970 shouldn't be gimping itself just to make the 980 more attractive. I don't like false advertising, and this seems to be a pretty solid case of it.
FrontalMonk
Banned
As someone who bought the 970 on release to replace his 560ti, I'm basically taking this tack. It sucks that the card isn't using all 4gb the way it's supposed to, and I think it's fine for nvidia to get burned in the press for not making all 4gb run as well as it should, but I'm not going to pretend it isn't a MAJOR upgrade over what I had, and I'm nor going to pretend that the 970 should perform just as well as the 980 because otherwise...what is the damn point?
Graphics card manufacturers have been selling budget models of their top of the line by cutting corners since the dawn of time. I'm not pretending that my old Voodoo 3 2000 was somehow just as good or better than a Voodoo 3 3500.
bootski, I like your comments about needing non-synthetic benchmarks. Maybe there aren't games out there that use >3.5GB simply with textures and models etc - you are needing to use supersampling to get there which is impacting performance. But I would expect in the next year or two that there will be games that will benefit from >3.5GB ram in 'normal' non supersampled use. In that situation, with your comments about the GPU throttling down to 85% occasionally, would that be consistent with the memory not being able to feed it enough data to keep it 100% active? So there may be lower framerates in games still to come?
Ok I tried the same with Crysis 3 and got massive stuttering even with 30fps. I suppose there must be frame times of over 200ms to cause this.
But when I set the settings for a framerate I would actually play (don't forget that I even use SLI) I get a VRAM usage of 2100MB.
So will this even be a problem in the future? Can't think of a game that would use more than 3.5GB VRAM and still perform well with my cards.
But when I set the settings for a framerate I would actually play (don't forget that I even use SLI) I get a VRAM usage of 2100MB.
So will this even be a problem in the future? Can't think of a game that would use more than 3.5GB VRAM and still perform well with my cards.
Lalalandia
Member
Excellent analysis Bootski if Nvidia are allowed to get away with this weaselly nonsense we'll never be able to trust their cards again. What I don't get is why they even did it in the first place, all the Maxwell GPUs perform very well and beat the competition, is any one willing to plonk down the cash for a 970 going to be turned off by 3.5GB vs 4GB?
To everyone saying 'it's as fast today as it was last week' you're forgetting that people buy these cards for next year,not next month. There will without a doubt be games in the next 6 months that will easily burn through the 3.5GB ceiling leading to inconsistent or very poor framerates. Also the 'it's not a 980' thing is a complete canard, while the performance drop is related to the changes they made to cut the chip back the problem is that they deceived people into thinking they had 4GB cards when the really have only 3.5GB. The removal of memory controllers is a consequence of how Nvidia designed the chip, the placing of an extra 0.5GB to make the card look good on shelves despite shitty performance is a consequence of they marketed the chip.
Nor is anyone else, what they're complaining about is being sold 4GB when they can only use 3.5GB
To everyone saying 'it's as fast today as it was last week' you're forgetting that people buy these cards for next year,not next month. There will without a doubt be games in the next 6 months that will easily burn through the 3.5GB ceiling leading to inconsistent or very poor framerates. Also the 'it's not a 980' thing is a complete canard, while the performance drop is related to the changes they made to cut the chip back the problem is that they deceived people into thinking they had 4GB cards when the really have only 3.5GB. The removal of memory controllers is a consequence of how Nvidia designed the chip, the placing of an extra 0.5GB to make the card look good on shelves despite shitty performance is a consequence of they marketed the chip.
Nor is anyone else, what they're complaining about is being sold 4GB when they can only use 3.5GB
CyanideFuse
Member
I had a 30 minute play of Mordor. I know this game has been mentioned to the hilt but it is the best one for me to get the VRAM over 3.5GB.
I noticed that system RAM usage increased when the VRAM went above approx 3590MB.
VRAM Usage | System RAM Usage
3184 | 6679
3308 | 6648
3452 | 6645
3520 | 6676
3644 | 7011
3668 | 7112
3688 | 7138
I noticed that system RAM usage increased when the VRAM went above approx 3590MB.
VRAM Usage | System RAM Usage
3184 | 6679
3308 | 6648
3452 | 6645
3520 | 6676
3644 | 7011
3668 | 7112
3688 | 7138
See if your GPU load stays solid, or if it fluctuates slightly as it passes the 3500 threshold.
The card can use the whole 4GB, why people insist they only have 3.5, I test it and even posted a video with gameplay. The only problem I noticed in Shadows of Mordor is that at 1080p the card or the game refuse to go beyond 3.6GB. At 4K in the same game the memory goes to the limit of 4GB with barely any drops.
This isn't the end all until we see how the 980 allocates the VRAM. We may find that like in watchdogs the 980 uses the 4Gb without being pushed leading to better frame times overall compared to the 970 which requires silly settings as you have shown to even go over 3500Mb.
Imo Nvidia is fooling 970 owners with this algorithm to only let the 500mb be used in insane settings where performance is already terrible.
Agree, they need to realese a fix on the drivers to just let game use the whole VRAM and avoid going to extreme settings to use that 0.5 missing, Imo is just a driver fixable issue.
No that's completely wrong.
They are not allowing (or giving very low priority) us to use that last 500mb in normal situations since it's slower than the rest, it would actually worsen performance if they did what you say.
We need a 980 owner to run tests like I did
Heres Shadow of Mordor from my system
i7 3770 4.2Ghz
Asus P8Z77I Deluxe
Samsung 2x4GB 2133Mhz CL10
GTX 970 1500/3880Mhz
The 1080p one was limited to 62fps by RTSS. The 2880p one was playable not stuttery.
Mean: 18.5ms
Variance: 90.16
Standard Deviation: 9.49531
Mean: 33.4ms
Variance: 159.5
Standard Deviation 12.6
More heavy VRAM usage
Last 2 shots show how SOM hangs onto 3500mb for 20seconds until the game demands more and driver gives up and gives it access.
During that 20sec period GPU usage is fluctuating badly along with frame times.
The fluctuation appears in 2 more periods whilst utilising the entire 4Gb.
Honey Bunny
Member
Well if you're looking for a game to try, the Star Citizen demo hits the 3.5gb cap for me at 1440p on the medium or high preset and stays around 35fps (on one 970). I'll also be really interested in how far The Witcher 3 pushes memory usage.
CyanideFuse
Member
The GPU usage fluctuated but not by a great deal. 99% down to 85% but it didn't coincide with the VRAM usage increase/decrease.
I just played again and the VRAM / RAM usage that I referenced before did not replicate itself. RAM usage stayed around 6.9GB when the VRAM fluctuated between 3.2GB and 3.7GB.
I have 4 CSV files of data that I will have to import into a spreadsheet to see if there are any wild and wonderful patterns I can find.
- Status