GTX 970 VRAM (update: 900M GPUs not affected by 'ramgate')

Anyone still got that benchmark tool?

 

You can still download it from Guru3D. This issue affects only desktop 970 cards, so no need to be concerned about the 980M or 970M.

What baffles me the most about this RAMgate is that this isn't the first time nVidia has used segmented memory. Both the GTX 660 and GTX 660 Ti had segmented memory (1.5GB @ 192 bit, 500MB @ 64 bit). The difference being in those cases nVidia disclosed this fact at launch, so there was no fuss. The fact that nVidia set their own precedent makes this 970 vram issue all the more aggrevating.

 

Not only that, but as far as I remember, those cards' final 512MB was faster than the 22GB/s of the 970, sitting at 48GB/s and being affected by memory overclocks and quite likely the SLI benefit... whereas since the 970 can't have its 512MB read by conventional programs, and I've determined from the way cakefish's 980M had higher bandwidth in the CUDA program test than my 780M, that that particular CUDA program entirely bypasses the SLI bandwidth doubling feature and works flatly on a single card.

Also, as I newly updated in my vRAM guide, the 970 having that final 512MB of vRAM is more of a detriment than a benefit. Instead of pushing caching data out of vRAM and compressing what's actually in there as much as possible at the 3.5GB vRAM limit, it forces itself to access the final 512MB at that ungodly slow speed and will attempt to fill that before compressing data or using virtual memory. In other words, for anyone who is going to run windowed/borderless windowed games with 2+ monitors plugged in and using Windows 8/8.1/(possibly) 10, they'll be almost instantaneously limited to 3GB of vRAM before things start to slow down/get stuttery. In that case, it appears to be better to buy a GTX 780 (especially the 6GB version), GTX 980, R9 290 or R9 290X instead of a 970. If someone is going to play lower vRAM-costing games like Battlefield/cod/Street Fighter/pre-2013 titles/etc, then the 970 would be a fine addition. But for anyone who even has a chance of hitting that vRAM limit? It's a miss.

What nVidia COULD do to alleviate the problem is they could write drivers or issue a vBIOS update to users which either:
1 - Segments off the final 512MB chip, actually making it a 3.5GB vRAM card, or
2 - Forces Windows' AERO and other driver vRAM requirements into the 512MB buffer FIRST, and allow non-windows app memory requirements access to the fast memory block, and prevent any non-native-windows driver/program requests to access that final 512MB block of vRAM, so that as far as 3D gaming works, that card acts as a 3.5GB card, and when at that vRAM limit, acts like the 780 does when you max out its 3GB pool. Because believe me when I tell you Windows can use a LOT OF vRAM without you actually doing anything. For example, here's a screenshot below of my monitors now. No games running, but look at the vRAM count being used in GPU-Z.

Spoiler

 

The last 500mb of my 8Gb GTX 980m are affected by this bug

 

Nope, it's because of Windows.

 

Are you sure?

 

There is like 250 posts of us explaining this earlier in the thread...

 

I thought so too at first, but if it was that bug it either should have affected the 8GB at both the 7.5GB point and the 3.5GB point or at least taking both 0.5GB together ...at the 7GB point.

 

Yes.

 

:/

Nvidia had this type of problem on the 8800, 500 and 600 generations. I'm not sure if returning the 970 and buying 980 really communicates their message.

I'd recommend people who bought a 970 to try and verify if it's an actual problem for them. Most pre-2015 games should be fine.

If it's a problem then get a refund. If you cannot then try legal proceedings. Nvidia has lost the last legal cases and this one is vastly easier to win with the amount of data presented. People in the EU and Aussies have already started legal proceedings for consumer protection so you can use that in your case as well.

 

They have never had this specific problem before, but I know the problem you are alluding to. This problem is gimping available memory on a memory bus that has a correct amount of vRAM attached to it. The previous problems were attaching TOO MUCH memory onto a memory bus (e.g. GTX 660 and GTX 660Ti had 192-bit mem buses, so they should have had 1.5GB or 3GB of vRAM; 2GB is above that limit).

 

That makes no sense. What does Windows have to do with slower memory? If an app overwrites the memory segment it'll overwrite it at full speed or none at all. It shouldn't slow down. Also, HD graphics will generally house the Windows components if integrated/discreet solutions.

 

Yeah that's because the EU and Australia (or hell, even most 3rd world countries) have much better consumer protection laws compared to the US. Why do you think nVidia is silently giving the green light to retailers accepting returns in the EU? Because they know they're in the wrong and if anyone were to take them to court there, they'd be in a world of pain.

 

Right. And with Cakefish's Optimus setup, it used all his vRAM without fault. It's literally on the first page. And there's no such thing as "overwrite" what's there. Windows needs vRAM. You don't kick windows' required vRAM into oblivion and say "HEY I ARE CUDA BENCHMARK" and expect things to work perfectly. Again, the last 250 or so posts explain all this. My entire vRAM guide explains this in more detail than 98.75% of the users annoyed about this even understand. Pure dGPU running + windows? Accessing last blocks of memory = slow. iGPU + dGPU in laptops OR headless mode on desktops? No problem. 970 is the only problem card with its specific issue, and its specific issue is worse than if it ONLY had 3.5GB vRAM as I explained on the last page in my post quoting n=1.

 

I'm thinking to create a thread about possibly lameness of Anandtech describing this issue but not sure in which forum.

GTX 970 is a GPU with 224bit Memory Bus while Anandtech claims it's 256bit which looks like total BS to me.
Thoughts?

 

http://linustechtips.com/main/topic/198568-the-video-ram-information-guide/
Read the "mismatched memory" section.

 

Thanks, a lot of useful info but I already know all that. I am telling that Nvidia continue lying about techspecs calling 970 a 256bit memory bus one.
P.S. I strongly disagree about "remove last 512MB" part.

 

Well, that 512MB is causing problems for people that wouldn't exist if it were a simple 3.5GB, 224-bit card. If you read the section, you know that when it forces itself to use that final 512MB for any reason, games begin to slow down. That's what the REAL problem is. If it were simply a 224-bit, 3.5GB card, it wouldn't have nearly as many problems as it does now.

That being said, nVidia DOES need to change the way they market the card... but technically stating "256-bit, 4GB" is not incorrect. The card DOES have eight 32-bit memory controllers and 512MB memory chips attached to each one. Seven of them are connected and run synchronously, and one does not. 32/8*7000/1000 = 28GB/s for that final section, so it does check out mathematically.

Games need to stop trying to access that buffer. At least on the 660Ti cards, it used a full 64-bit controller (2 x 32-bit) for 48GB/s, which could benefit with SLI for 96GB/s (overclockable).

 

I believe that without those 512MB it would stutter the same. It tries not to use it as long as it can meaning it will only if it would have to use system RAM otherwise.

7x32 = 224 bit. It is 224 bit card.

 

You didn't read my guide very well then. When a card runs out of vRAM, it starts to dump as many cache files and unnecessary information and compress as much information as it can. When it can do this no longer is when it begins to attempt using virtual memory of some kind. The final 512MB on the 970 gets used * BEFORE* the video card begins compressing/dumping/virtual-memory-hunting. If that final 512MB were blocked from games, it would begin compressing data at the 3.5GB mark, and it would never attempt to make use of that slow memory section, and it would likely take a game requiring well over 4GB of vRAM, not counting caching, to actually work (of which the only one I can think of offhand at 1080p would be Middle Earth: Shadows of Mordor, at the absolute maximum graphical settings, not counting SSAA, that it can take) before virtual memory would be accessed. Maybe AC:Unity above 1080p might trigger it too.

But that being said, no matter how I look at it, blocking that final 512MB from being accessed is going to help the GPU perform better. It accessing that 512MB portion before compressing what's in the fast portion is what is causing stuttering and slowdowns in games where even a 3GB card would not. Just like in this video HERE I linked as proof in the very same section. You can clearly see that even though the GTX 780 3GB card is losing out in FPS over the 970, its video feed is much smoother than the 970, which is using enough vRAM to tap into its slow pool of memory.

 

Yeah, this is proved by dozens comparisons of monitoring data between 980 and 970 where at the very same settings 980 takes all 4GB of RAM where 970 almost always only 3.5GB. Oh wait... doesn't it contradict?

 

It is technically 4GB, but not 256b. Only 224b can be accessed at the same time.

 

Because the final 512MB can't be read by conventional software. This has been talked about before. You don't see when it's using slow vRAM. As far as GPU-Z/MSI Afterburner/Playclaw 5/etc is concerned, your card stops at 3.5GB. But the stuttering the game shows is indicative that the slow vRAM pool is being used, and the frametimes are inconsistent. And the reason why we know is because using a similarly powerful card (like the 780 3GB) we see that the frametimes are consistent when its vRAM pool is maxed... because the card is properly tossing out cache data etc. The DIFFERENCE would be that random texture pop-ins would be a bit more likely with the lower vRAM card, due to kicking out as much data as it can. But I'm fairly sure if you told everybody they had a choice: lock off the slow vRAM portion and keep the card as a 3.5GB card and fix the stuttering issues, OR keep the card the same and only change the marketing data available for it, the majority will probably ask for their stuttering issues to be fixed.

What I am saying is, nVidia is throwing a loophole.
Does the card have 256-bits total of memory bus controllers? YES.
Does the card have 4096MB total video RAM installed? YES.
Does the fast portion of memory only have access to 224 bits of the memory bus, effectively lowering overall memory bandwidth? Yes, but we don't need to say all that BY LAW.
Does accessing the slow portion of vRAM cause issues in frametimes for users? Yes, but framerates do not necessarily drop, and framerate data is enough to support our tests and most hardware website tests, so that's all we'll offer.

Get it? They're walking on a technicality here, and they know it, however they're not lying. They're simply "not telling the whole story". Think of it like this: if I come home hurt one day and say I was in a fight because I was protecting a classmate, would I sound like a hero to my family? Probably. Did I mention that I was only in the area because I was a part of the gang who was hurting the classmate and I only fought them when they went too far? No. Did I lie? No, I did not. Did I tell the whole story? No. But it's enough for people to be satisfied.

Of course there'll be some people who pester, and look, and try to ask more. And that's good. That's what we need. But it's not "incorrect" to say that the card has 256 bits of bus width on the memory controller. Doesn't matter if it can only access 7/8 of it. It HAS 256 bits. It's not "incorrect" to say it has 4GB vRAM on the card. In fact, it's a LIE to say it has 3.5GB of vRAM and not mention the final 512MB. Which is why they're not being sued by a billion people.

What it is, however, is deceitful marketing, and that's likely why people in Europe etc have been getting refunds easily, as people have reported. The US however has no such laws to protect against deceitful marketing. In fact, if it did, I guarantee you many gaming companies would have quite a few lawsuits against them in the US.

 

I knew you gonna write that But I also read that MSI is going to make a new vBIOS which lets GPU-z see all memory taken and only then we will really know if it takes 512MB when doesn't need to or not.
As I said, I don't believe it does take because there were screen 970 takes more tan 3.5 but they are rare.

 

I disagree.

1-Nvidia may not need to say anything about bandwidth "by law", but having listed the bandwidth of the GTX 970, they lied in their advertising none the less.

2-Having lied about the bandwidth, that lie misrepresents the true Memory Interface Width. You never get a 256bit bus access for any of the available video memory. I don't think it matters that the 970 has a total of 256bits in its memory controller because their own website presents it as unified across all of the available video memory by the way of the listed bandwidth.

So that's at least TWO LIES in their own advertising about the GTX 970. It isn't just a case of Nvidia "not telling the whole story", Nvidia's lying on the bandwidth essentially makes up a whole new work of fiction about what the GTX 970 is.

 

There might be 256 bits in chip bus but memory width is surely just 224bit. I call Nvidia a lier. And I call Anandtech either lame workers or frauds who took Nvidia's money.

BULL SH!T. It can achieve 196GB/sec and WILL NEVER achieve 224GB/sec.

It has no 256 bit memory bus. It has 224 bit memory bus and just 32 when is over 3.5GB.

Anandtech is trying to legalize BStness of Nvidia so in the future they would write any "technically right information" without sequences. They had to destroy nvidia's fairy-tales like reporters in USA destroy too cheap and filthy politicions.

 

So they just continue dog fighting.

 

What on earth are you guys talking about?

GTX 970 is a 256bit card.
The GPU can access both memory banks simultaneously through the crossbar with the 7th L2 cache.
Its the one on the lower right corner and all memory controllers are in use.

 

I might test mine later today - Lord knows I have enough issues with this card as is. Why not add another to the list... 8GB variant here. Anyone else checked that out?

 

Already confirmed to not have the problem. Only affects desktop 970.

 

Yeah, I'm done explaining things on this thread. I've explained much much more than I should have needed to, and nobody seems to be listening to reason.

 

i feel ya buddy, had something like that in the 770ZM lounge

 

You have 8 memory controllers. Each memory controller is connected to the memory by 32bit. In total you have 32bit * 8 = 256bit
Each memory runs at the same speed. Nvidia didnt clock down the last 0.5GB module despite what people and that benchmark says. Its running just as fast as the 7 other and is also connected to a 32bit bus.

There was a rumor going on that the GPU can`t access both memory banks at the same time, also spread by Anandtech (shame on you) but Nvidia explained later that the 7th L2 section can access both the 7th and the 8th memory by accessing two memory controllers ( Crosslink picture posted on previous page).

There are also a lot of clueless people saying you have 192GB/s if you use the 3.5GB memory bank and you get a lot lower speed when using the last 0.5GB. Isolated you get 1/8th of the speed with just that 0.5GB DRAM yes, which is what the benchmark measured. But since the 7th L2 cache can access both memory controllers of the 3.5GB bank and the 0.5DRAM, you get a total of 224GB/s. Exactly like advertised by Nvidia.
That the one L2 section have to read two memory controllers instead of 1 may lead to more stress on that, but the GPU is still a 256bit card and it got 224GB/s in total and can use all 4GB VRAM.

There isnt a "faster portion" of the VRAM when the GPU can use all 8 DRAM. All 8 DRAM runs at the exact same speed.

 

Well

Gigabyte
Amazon
eVGA
OC.uk
Newegg

Are all refunding some users.

Waiting on MSI/ASUS

The damage is done.

 

The part that doesn't make sense is the dGPU is being tested. Windoze writes it's graphics data to the iGPU (unless otherwise configured). The CUDA tests are hitting the gGPU hence the 500MB should have nothing on it unless Nvidia caches the iGPU data as well. For sake of argument let's say there is 500MB of data in gGPU cache. Why would everyone's benchmark hit a snag at 500MB (even headless ones which don't write cache to the 500MB block). On clean machines with low amount of eye candy people should be hitting the problem post 3.5GB mark or a 4GB card.

Your explanation makes sense using exclusively the dGPU as a main card. Either way I won't be able to verify this until my 980M comes to my house.

BTW
I think most people don't seem to realize at the end of the day the problem is minor for most gaming.

 

GTX 980M have the exact same setup on the DRAM/MC as GTX 980. The only thing thats disabled there is SMM

 

Man, can you just trust me that You Are Wrong? Or I have to post some video explaining that in details?
Before doing that I will try to write shortly in words and pictures: that L2 cache can't work simultaneously with 2 Memory Controllers It just stripes between them hence why all this problem appeared! There is no "L2 section has more stress on that, but the GPU is still a 256bit card" there is "... hence why GPU is 224 bit card".

Just look at the comparison after all!
Watch that GTX 980 has 178+GB/sec while GTX 970 has 150+GB/sec which is exactly 7/8th of speed . That's because it is BS that 970 can use all 8 MC at once filling GPU vRAM. Because only 980 is 256 bit card where 32bit*8MC with 224GB/sec theoretical bandwidth.


So after you finaly found that in Maxwell you need to count not just MC but MC-L2 lines will you ask forgiveness for "lot of clueless people saying you have 192GB/s if you use the 3.5GB memory bank and you get a lot lower speed when using the last 0.5GB"?

 

From PC Perspective:
" UPDATE 1/27/15 @ 5:36pm ET: I wanted to clarify a point on the GTX 970's ability to access both the 3.5GB and 0.5GB pools of data at the same. Despite some other outlets reporting that the GPU cannot do that, Alben confirmed to me that because the L2 has multiple request busses, the 7th L2 can indeed access both memories that are attached to it at the same time."

Though given Nvidia's recent spotty record, they could be mistaken again. IMO there needs to be more testing done to confirm/deny this.

 

Sure 7-th LC2 can act with 2 MCs at once... but presumably with the same overall bandwidth as other LC2 act with their one MC.

 

In other news, it seems that the 3.5G limit is only effective with DX, even though VRAM is a global resource that can be shared across APIs.

OpenGL Apps Not Impacted by GTX 970 VRAM Limitations?

 

OMG, I give up. People can`t read.
Nvidia is lying. Its a cover up.

The benchmark read 7/8th of the 980 speed because it accessed only the 3.5GB portion. That is the total bandwidth they gave. Then it got 1/8th of that on the 0.5GB when it read that.
Nvidia specified that programs and games can access both SIMULTANEOUSLY. Through the 7th L2 cache. Which is why VRAM usage can go up to 4GB and which is why it is a 256bit card (32*8) and have 224GB/s bandwidth (all 8 memory controllers together).

Why is this so hard to understand?
The change from the original specifications from Nvidia is 1.75MB L2 cache instead of 2MB since one is disabled. And it have 56ROPs instead of 64ROPs you find on GTX 980.
So you have less ROPs than 980, more strain on the 7th L2 cache that needs to access 2 DRAM pools including 2 Memory Controllers, which can cause some issues in cases when you compare against GTX 980. ROPs is especially important on memory intensive tasks such as high resolution and VRAM usage over 3.5GB which have been exclusively what this mess have been about.

 

Is there any explanation why one was disabled?

 

Because it is wrong. Simple as that.

Only the first message of the quotation is truth.

When it got? When a magic Fairy brought it? Or when memory allocation accedes >3.5GB? If 2-nd then why does it get stuttered like hell if you say it increases bandwidth and not decreases?
Considering that it contradicts then the only real explanation would be a magic Fairy. Although I already told the real one.

Will make a video definitely. Not personally but for all GTX 970 buyers (inc. potential) so they would know what the buy.

P.S. Do you at least know what that benchmark program does? It does not write to the exact chips one by one. It just sends info and GPU puts it to all chips altogether (7/8) like RAID HDD. If last L2 cache could write to both MC with same speed as 2 LC write to 2 MC then why the hell it doesn't?

 

"Because it is wrong"
Great argumentation buddy

Try to read what Im writing for once:
The benchmark can`t access both pools at the same time. Games and many other software can. Nvidia said it earlier in their official statement. Why I don`t know, but it can def be driver related.
With GTX 980 it only have one VRAM pool, hence it cannot measure it wrong.
Even Nvidia`s own engineer said to PCPer that it can read both pools simultaneously. Guess he believe in magic too and know less about his own products than you?

Why it stutters like crazy? It doesnt stutter like crazy. Now you are blowing this way out of proportions. There is 3 reviews posted earlier in this thread which found nothing wrong with VRAM over 3.5GB. They used FCAT which measured frametimes over the complete test over many games. Some users found stuttering over 3.5GB, some didnt. Comparisons are hard to do between different rigs, different setups and with various programs running in the background. Like I said above, it have less ROPs and a L2 cache which does more job than on the 980.
ROP is very important for GPU performance with high VRAM usage and could lead to stuttering under certain scenarios.

 

http://www.guru3d.com/news-story/both-mantle-and-dx12-can-combine-video-memory.html

GTX Tri 980 SLI = 12GB vRAM?

GTX 980M SLI = 16GB vRAM?

GTX 970 3.5 + 0.5 = 7GB vRAM + 1GB or since access is easier 8GB vRAM?

 

So with DX12 and Mantle they can spread out VRAM usage across both cards?
wow, thats cool. A nice bonus for us SLI users

Not sure what I would do with 12GB VRAM though lol

 

Sharing VRAM across different VRAM controllers requires fast connection to handle the throughput. Now guess how much bandwidth does PCIe have...

Sharing VRAM on different devices is nothing new. Compute programs have been doing it since the very beginning, and technically in SLI (I'm not sure about CF's implementation detail) rendering you can do it as well. The problem isn't it doesn't work with DX11, it's the performance scaling being poor. Using multiple GPUs efficiently requires data level parallelism and with rendering you simply can't have that. This is why SLI/CF games are all alternating frames nowadays even though split frame has much lower latency. With rendering algorithms getting more and more complex it's harder and harder to code split frame multi-GPU rendering effectively.

 

That is a marketing rep. He told about Split Frame Rendering. And even though he told that they will use 8GB of memory I seriously doubt that all that 8GB will be filled with UNIQUE information. I believe that most of information in 2 GPUs will be the same hence they deal with one actual frame.

@Cloudfire, why would 1 cuda benchmark not use all vRAM while all other games and programs do? Because benchmark is black?

 

Games can when *allowed* by the driver, which means application specific.

Anyway the card is still messed up because frame latency/stuttering will never go away.

Lawsuits will probably be semi-successful (and there are 2 class action suits in motion).

RIP GTX 970.