GTX 970 GPU series wide memory allocation issue

http://www.neogaf.com/forum/showthread.php?t=977645

just starting to heat up. Nvidia forum rep says the company is looking in to the issue.

apparently it is starting to come out that GTX 970s suffer from a significant decrease in throughput on the last 500-700mb of memory, effectively limiting the card to 3.5gb of ram. GPU usage also significantly drops when the cards starts accessing this last bank of RAM, causing stuttering and performance drops. this could be a hardware issue with all GTX 970s featuring Hynix RAM (the vast majority of the product line).... as some are reporting that they do not suffer from this on cards without Hynix RAM.

also of note, the GTX 980 apparently does not suffer from this issue, and apparently does not feature Hynix RAM.

considering i just bought an SLI 970 setup... im extremely frustrated by this. Nvidia has a week to release an official statement, or im contacting Amazon and demanding refunds.

Its not limited to 970's some 980's also experience the issue.

there's hardly going to be any difference between 3.5 gb and 4 gb, they'll just release a driver that disables the faulty bank

asinine comment, considering there is a huge difference in the benchmarks that are being done to show that this is legitimately an issue. not only is the memory shut off.... but the GPU usage takes a nose dive as well when you hit that bank of memory, causing performance degradation and stuttering.

not to mention that people bought a 4gb card.... not a 3.5gb card. its false advertising, and a legitimate issue that deserves a recall. every single person that bought a 970 should get their money back if they want it.... i dont care what it costs NVIDIA. they were sold a faulty product.

either they did a limited run with Hynix RAM... or newer revisions are now using Hynix RAM... because the 980s that were tested did not suffer from the same problem.

He's well within his right to expect a refund for a defective card.

I don't agree with him that it's false advertising or that they did it on purpose, but that's besides the point.

I have had 970 SLI since launch and I have not yet noticed any major drops in performance or stuttering (aside from the shit Ubisoft titles). I am only running 1440p though. I am sure 4K users would run into the problem easier.

Here is a bit of info

"

ike every other 970 user today, I ran the Nai's memory benchmark.

Results yup, broken.

After pressing F to pay respects, I've decided to read more about the issue.

One of the things I've noticed is that people with different GPUs (Titan, 670, 980) were reporting similar results.

While reading what different people were reporting about the issue, I've noticed that one of the users said that the benchmark is meant to be ran in headless mode (with the screen being connected to iGPU) due to VRAM usage being anything other than 0 negatively affecting the results - something that isn't exactly mentioned a lot, and I'm willing to bet that a lot of you (and me) ran it "as is" and got incorrect scores.

Being lazy and noticing that other people were getting strange results (hundrets of TB/s, problems with other GPUs etc.) I've decided to try to replicate the issue with anything BUT Nai's benchmark.

First thing tested: EVGA OC Scanner X.

Although the highest setting of it's GPU memory burner is 3072MB, it filled up ~3900MB (which according to Nai's bench belongs in the "slow" area).

This screenshot shows the test in progress with gpu-z used for monitoring. Important to notice is that GPU Load is at 99% - i.e. not bottlenecked by memory speeds being 10x worse than what they're supposed to be.

Next test: DA:I

I've ran Inquisition at 4k using nVidia DSR with everything set to maximum, including 4x MSAA to try to fill up as much VRAM as possible.

Starting VRAM usage - 3037MB

A bit later - 3305MB

Towards the end - 3611MB

Important thing to note is that there was no sudden performance drop - something you would expect when a part of your VRAM is 10x slower.

Next game: watch_dogs, something I've seen people bring up a lot when talking about this issue.

Ran in 4k (dsr) with 4x msaa and 1440p (dsr) with 8x msaa on highest settings and max pre-rendered frames set to 1.

Switching the game to 4k with 4x msaa filled the VRAM completely really fast

4067MB

4062MB

Surprisingly, the game didn't run that bad considering the settings. There was some stuttering when rapidly turning around, but that is probably more due to VRAM being full rather than slow and it is to be expected in a 4k+4xMSAA scenario.

Running the game in 1440p combined with 8xMSAA resulted in no stutter, while keeping the VRAM usage at around 3,8GB

3832MB

3838MB

In conclusion: it really seems like that everything is fine - there was no performance drop in DA:I as the VRAM usage gradually rised from 3GB to 3,6 (according to my Nai's benchmark results, the "slow" area starts at 3200MiB, which is equal to 3355MB), Watch_Dogs ran at 4k with 4xmsaa at ~20-22 FPS, which is much better than I expected, and at 1440p with 8xmsaa at cinematic (~30-32) fps, with no stuttering that was present in 4k+4xmsaa due to VRAM being simply full.

And the most important part: GPU Usage was pretty much locked at 99% (except during sutters in w_d at 4k+4x), indicating that there was no bottlenecking in terms of memory bandwith suddenly becoming worse etc."

Sure, I may make asinine comments but I'm not the one stuck with two faulty 400$ cards lmao

Well I seen 2 defective GTX 970's one was mine and one was a friend... Out of the 6 I installed for people. I got my GTX 980 Matrix this past week and that too was defective with blinking green light with no way to get out of safe mode.

I wouldn't doubt there would be other issues with these cards. I read about this about 3 weeks ago. I never seen about Hynix though... Till now.

yet another asinine comment from you.... since A: they arent 400 dollars and B: consumers dont have access to this information prior to buying the product.

the cards have been out for months. this is just now being discovered. its not like myself, and THOUSANDS of others that recently bought one are "early adopters"

Well, I feel bad for those users, for you not so much, since you're projecting your frustation on me while I'm trying to help you. What I said was pretty much worst case scenario and in the end that ain't too bad.

But it probably won't come to that, since you have warranty.

If you haven't noticed it until now, no need to immediately panic. It will get figured out. I certainly wouldn't mind if EVGA was willing to replace my 970s with 980s for all my troubles. ;)

Well that's not pretty. I guess I need to check to see if I can use all my VRAM.

oh sorry... i didnt realize "there's hardly going to be any difference between 3.5 gb and 4 gb" is a helpful comment *rolls eyes*

also "Well i'm surely not as asinine as you"... yeah... soooooo helpful.

you know... that is only like 12.5% of the total RAM...pssshhh, you wont even miss it right? especially not at 1440p or 4K where VRAM really starts to be important....

Massive grain of salt with this. This tool used by whoever this no name person is who found this (and this late into the card's life? Cmon...) is loaded with bugs. Also, I cannot replicate this issue with anything other than this faulty tool of his. I regularly fill my 4gb of vram and no stutter or massive drops in framerate.

My 970 Strix has Samsung RAM and this program is showing that the last 1GB is unusable, is the program bugged?

Hmm I've got samsung, and I run three monitors. Which I already lose out on some vram due to that. I'll test this when I get a chance.

very, it cant even read the ram on other models correctly. Some random dude makes his own program and finds this months after release, nvidia says nothing and neither do reviews......its very off and sketchy

Did nobody else read what this guy had to say?

No need to panic? Think again buddy. No one is gonna replace 970's with 980's. Nvidia got away with this before when they released 2GB GTX 660's and only 1.5GB was actually usable. Looks like, they're on the same track.
If It was me, I'd be worried. Newer titles @ 1440P are starting to require well over 3.5GB. I already noticed that Advanced Warfare alone eats away 3.8GB on my 290's.
I saw this heat up in the morning (bunch of threads on overclock.net). Doesn't seem to matter what card, what brand, what BIOS. All 970's are looking like, they're faulty. I hope, Nvidia doesn't get away with this again.

I'm waiting for @kitty to post his results, just want to know if its BS or not, is every 970 affected or just Hynix RAM ones...

Well I just gave you an option that can happen if you can't get a new or other card through warranty because that was your concern. If you don't want answers then don't ask questions. I could have told you the obvious thing that you'll probably get refunded through warranty but that's a useless answer, anyone knows that.

And yes 3.5 gb ram is still better than 4 gb when the faulty ram bank is actually used and indeed you won't even miss it, not even at 4k. The 780ti 3gb is slightly better than gtx 970 (even in sli) and ithas only 3 gb. Memory size doesn't do much once you have enough and at 4k you have enough with 3gb.

The gtx 970 will be bottlenecked by the 256 bit bus first , then the gpu clock not that 500 mb vram size. Not to mention that the gtx 970 in sli is not really enough for current gen games at 4k. It's enough for previous gen games though but again 3.5 gb or 4 gb won't change a thing.

You better think before you call someone assinine because my comment is more than usefull, you simply won't notice the difference. And I only called you assinine because you did it first, while i was trying to help you and was even correct on top of that.

YEAH TOTALLY... that 500meg wont be missed at all.... i mean people arent already hitting 4gb VRAM at 1440p or nothing... its TOTALLY not happening.

It's not because the ram bank is accessed that it is needed. This doesn't work like harddrive space.

SLI 960s or Cross 290s? XD

Hello guyz.

I am monitoring this as well and both the programme and the users that try this are doing something wrong.

A few people run the test properly. I even saw people with GTX Titan having last 500MB with low GB/s. Even GTX780 and GTX780ti. (at Guru3D)

Also this programme runs on a CUDA enviroment and the bios automatically sets the memory to run at lower speeds. This justifies the 150GB/s instead of 224GB/s, but not the 10x times lower bandwidth drops.

The real problem would be what @04dcarraher said. If performance in games is affected when GTX970 uses more than 3500MB VRAM, then we have a problem.

The good news is that Nvidia is aware of this issue and they are investigating this.

I hope if there is a problem it could be fixed by driver/bios update and not by RMAing, cause this would take a looooooooooooong time to RMA all the affected GPUs...

....

Did I mention that I ordered a Gigabyte GTX970 Gaming a few hours before this problem was surfaced (after I saw the 60% diff with the GTX960) for 400 euro? :P

@cyloninside: I completely understand how you feel :P

edit: This is from a guy at guru3D:

From Nvidia Chat Support this afternoon - (names changed)

[10:11:39 PM] NV Chat: We have our entire team working on this issue with a high priority. This will soon be fixed for sure.

[10:11:54 PM] Me: So, what is the issue?

[10:12:07 PM] Me: What needs to be fixed?

[10:12:46 PM] NV Chat: We are not sure on that. We are still yet to find the cause of this issue.

[10:12:50 PM] NV Chat: Our team is working on it.

...

You and everyone else needs to chill. Why jump to conclusions? Nothing has been confirmed yet. I have the first run 970s in SLI and have never experienced any issues with VRAM on any games. So, why on earth would I be worried all of the sudden because of some buggy program? As for my comment about swapping to 980s... I think it should be pretty obvious that I said that jokingly. Did you not see the wink at the end? Regardless, if there is indeed an issue, I have faith EVGA will take care of me.

People need to relax. It is not like it was discovered that the cards are faulty and will catch on fire. If that was the case, I would be worried. I will continue to enjoy my 970s and await further testing results. Right now, this is a non-issue from my perspective.

yeah indeed, The 780 ti has 3gb and gets better results. That 4 gb looks nice but that's it. The bottleneck with the gtx 970 will be the 256 bit bus not the vram, not even at 3 gb.

I'm pretty relaxed, I'm not the one sitting on 970's. I just don't think Nvidia has a right to do that to people.
You're fine now but what about games that come out later on requiring more VRAM?
I do agree though EVGA is pretty loyal to their customers so I'm sure, they'd do something about it.

It seems Nvidia responded:

NVIDIA Responds to GTX 970 3.5GB Memory Issue

NVIDIA has finally responded to the widespread online complaints about GeForce GTX 970 cards only utilizing 3.5GB of their 4GB frame buffer. From the horse's mouth:

The GeForce GTX 970 is equipped with 4GB of dedicated graphics memory. However the 970 has a different configuration of SMs than the 980, and fewer crossbar resources to the memory system. To optimally manage memory traffic in this configuration, we segment graphics memory into a 3.5GB section and a 0.5GB section. The GPU has higher priority access to the 3.5GB section. 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.

We understand there have been some questions about how the GTX 970 will perform when it accesses the 0.5GB memory segment. The best way to test that is to look at game performance. Compare a GTX 980 to a 970 on a game that uses less than 3.5GB. Then turn up the settings so the game needs more than 3.5GB and compare 980 and 970 performance again.

Here’s an example of some performance data:

On GTX 980, Shadows of Mordor drops about 24% on GTX 980 and 25% on GTX 970, a 1% difference. On Battlefield 4, the drop is 47% on GTX 980 and 50% on GTX 970, a 3% difference. On CoD: AW, the drop is 41% on GTX 980 and 44% on GTX 970, a 3% difference. As you can see, there is very little change in the performance of the GTX 970 relative to GTX 980 on these games when it is using the 0.5GB segment.

Good response from Nvidia. I just installed my second Asus Strix 970. The first one has samsung memory and the new one has the hynex.

turns out the people getting the error also did the test wrong by testing the gpu while it was being used. No wonder I couldnt replicate this "issue." Yay for people grabbing the pitchforks on first sight, well done internet

Pitch forks! unite! lol

See.... all is good. No need for panic.

probably a single 980 is more likely considering most people are reporting that 980s are not affected..... although it is starting to appear *some* 980s are affected as well.

Then 2x 290s it is! :D

Is that crosfire already free of microstuttering?

So pretty much, it's true. It uses 3.5GB optimally and the rest no so much. People were talking about stuttering using more then 3.5 GB and they didn't even say anything about that. Just the AVERAGE FPS. Which means nothing as stuttering can happen at any frame rate.

Maybe pcper will do something on this... Like they did with the Samsung EVO 840 drives.

Why didn't NVIDIA say this at the card's launch?

? they basically gave a non-answer AND admitted to the issue....

people werent complaining about average FPS... they were complaining about stuttering. average FPS isnt going to reflect that....

Now this is becoming too complicated.

There might be an issue with CUDA using all 4GB VRAM.

post #452 at OCN:

Hello,

just made a test, concerning this.

Loaded Tomb Raider and it used around 2760 MB

Than started the "memory test" to use 2GB (if possible). It allocated 1344MB.

Total memory usage was 4GB

Test was flawless without a performance drop!

This is resolved, it's a CUDA problem with "segmented" usage.

Ofc other people gave their explanation on this subject (like Hynix memory chips have problems etc etc), but this is quite strange even if in the end its not the problem...

It's almost none exist today.

For people getting a 970, I'd recommend trying to find one with Samsung memory anyway because it overclocks better. I have the Galax: Infinity Black edition (guaranteed Samsung memory) and the only game I've had stuttering issues with is Far Cry 4, which can make SLI 980s under perform and is seemingly a game issue.

I just got SLI 970s but luckily they don't seem to be affected by this issues. They don't have the Hynix RAM, but Samsungs and performance in games that use over 3,5 GB (AC Unity maxed with TXAA) seems to be normal.

Definitely a game issue. SLI support in FC4 is pretty crap, it almost runs better with one 970. Not to mention AA seems to be broken with SLI and I've had texture ghosting.

I think Ubi disabled SLI in the most recent FC4 patch. It's def hitting only one my 970s. The other just sits idle when I play.

How does one tell what brand of memory their GPU is using?