NVIDIA Responds to GTX 970 3.5GB Memory Issue

[ACatWithThumbs]

 

"From the Nai's Benchmark, assuming if the allocation is caused by disabled of SMM units, and different bandwidth for each different gpus once Nai's Benchmark memory allocation reaches 2816MiBytes to 3500MiBytes range, I can only assume this is caused by the way SMM units being disabled.

Allow me to elaborate my assumption. As  we know, there are four raster engines for GTX 970 and GTX 980.

Each raster engine has four SMM units. GTX 980 has full SMM units for each raster engine, so there are 16 SMM units.

GTX970 is made by disabling 3 of SMM units. What nvidia refused to told us is which one of the raster engine has its SMM unit being disabled.

I found most reviewers simply modified the high level architecture overview of GTX 980 diagram by removing one SMM unit for each three raster engine with one raster engine has four SMM unit intact.

First scenario

What if the first (or the second, third, fourth) raster engine has its 3 SMM units disabled instead of evenly spread across four raster engine?

Second scenario

Or, first raster engine has two SMM units disabled and second raster engine has one SMM unit disabled?

Oh, please do notice the memory controller diagram for each of the raster engine too. >.< If we follow the first scenario, definitely, the raster engine will not be able to make fully use of the memory controller bandwidth

64bit memory controller, total 4 memory controllers = 256 bit memory controller.

Assuming if there are 3 raster engines with each three has one SMM disabled leaving 1 raster engine with 4 SMM intact.

Mathematically ;

16 SMM = 256 bit = 4096 Mb

13 SMM = 208 bit = 3328 Mb

208 bit = effective width after disabling SMM with 256 bit being actual memory controller width

IT is hardware problem=GTX970 is 208bit card."

 

Shouldn't this problem also be on Kepler and GCN since most cards have disabled units?

 

[mr moose]

970 was sold as a 4k solution in SLI. It was marketed as such everywhere. It was lauded as such. The benchmarks Nvidia linked mean nothing without frame latency. We don't know wtf they are benchmarking. It could be staring at the damn ground...

 

Like I said the only game I can think of at 1440p that has high ram usage atm (besides Supersampling AA/downsampling which I also have a link for) is shadows of mordor maxed without the ultra texture pack. Benchmark it with frame latency, The game has a damn Nvidia logo when you load it. I don't think a more fair test could be had. 

 

So in a single card configuration? Yup wouldn't mean much. Also as far as someone saying you would run into a horsepower problem before a VRAM problem? That is 100 percent BS. the 6 gig VRAM textures on Shadows of Mordor prove that. My R9 290 runs into a vram problem and the game runs like butter at 1440p without them.

 

That is the case on like a 760 with 4GB of VRAM (stupid card) but that card could still benefit from 3 (if there was a 3 gig card). That is why you see a R9 280/x blowing 2 gig cards away in Ryse with SSAA on, which is similar to Nvidia DSR/AMD VSR. Star Citizen is going to use Crysis engine like Ryse. I bet a ton of people bought dual GTX 970's for the game. THAT is the people who are probably upset and feel like they were sold snakeoil, and higher texture packs than the consoles which can use mid 2's are not going to be limited to one game and higher AA is here to stay. If the GTX 970 was marketed as a single card solution you would not see near the complaints from people.

 

 

I get what your saying, and I still agree, but only to a point.  If people bought sli 970s for star citizen,  then on what marketing did they make their choice?

Do we even know what the minimum requirement for half decent graphics on SC is yet?

 

I just googled and got this:

http://www.guru3d.com/articles_pages/geforce_gtx_970_sli_review,21.html

 

Now if I was buying a set of cards, I would use this (along with other reviews) as a guide,  So unless Nvidia pulled a bait and swap,  then it is capable of 4k.  And this is what people who own a pair should expect regardless of them having a memory flaw or not.

 

I won't comment on the horsepower v vram debate because  1. I don't know enough and 2. it could still be a driver/software issue making it very difficult to know what hardware will become the limiting factor first. 

 

@Lays did a quick test for me with his sli970s in SoM,  @4k when the ram usage goes over 3.5 the frames drop of and the game stutters.  I think he said It didn't matter where he was looking.  So there is no doubt it's there and this is whats holding the card back, however the question is, is it a flaw or is it the reason it's cheaper?

 

After reviewing this post I notice a lot of question marks,  these are genuine questions, although I don't think there are answers to some of them just yet.

[Sammael]

 

 

Looks like Ryan talked to an nvidia gpu architect and got some questions answered, he is supposed to have an update up by 10 AM Pacific Time going over the details.

 

 

http://www.pcper.com/news/Graphics-Cards/NVIDIA-Responds-GTX-970-35GB-Memory-Issue

 

 

If the nvidia engineers are weighing in I just hope there is not some advanced spin machine trying to minimize damages and not get at the truth.  Even if performance hits are negligible, IF it's really true that the 970 cards can't use the full 4 GB of vram like other cards, that still sounds like deceptive advertising as it might have gotten people to choose different options if they had known there were technical differences in the memory accessibility.  Ryan tends to tread very carefully when talking to nvidia because I'm sure he wants to maintain access.  He can't go all Charlie on them like his semiaccurate "wood screws" coverage went.

 

http://semiaccurate.com/2009/10/01/nvidia-fakes-fermi-boards-gtc/

 

Note:  Charlie is a great nvidia hater, but it's a treat to watch someone filled with such bile towards nvidia pile on with glee at their struggles.

[mr moose]

 

 

 

If the nvidia engineers are weighing in I just hope there is not some advanced spin machine trying to minimize damages and not get at the truth.  Even if performance hits are negligible, I

 

A little bit from column A and a little bit from column B.  You can bet your arse the Nvidia PR crew are working overtime on this, however If the truth of the matter is that this is just the way the hardware is and can't be fixed then it doesn't matter about the engineers.

 

Scanning through the Nvidia forum and I think I found the most truthful and reflective post there is:

 

I don't know one min this is a false post now its true .. im just hear to add to the mix, my cards run just fine .. I just felt left out and wanted to jump on the band wagaon while randomly changing sides just for fun .... that's how much this crap means to me lol

 

 

It goes from a driver issue to a game issue to a hardware issue to a memory brand issue back to a bios issue.  I wonder if we'll ever know what the truth of it is?

[Rock1m1]

I really hope this is a sign that we need more kinds of tests when testing out a new GPU. Tests which targets specific components of the GPUs rather than the overall performance: FPS, frame time, power consumption. 

[Khvarrioiren]

Keep being an asshole then, sure. You say it doesn't drop performance, then prove it. You say something does drop performance, then prove it. The guy I was responding to only gave three data points, have you really never taken a basic science class to know that's a completely bullshit way of testing for something? All performance goes down when the load increases, it's what's expected. This is how god damned skewed and bullshit the graph would be to go on if you used three data points:

 

 

If you don't see how that's stupid to use as a model... I have no words.

 

If I'd really want to convince such an irrespective person who straight up resorts to ad hominems I would actually bother to show you some benches on Survarium which pretty much always maxes out my VRAM without dropping more than an average of 2-3 frames. Have fun with yourself from this point forward.

[.spider.]

Also, I now know why I stutter a lot on Assetto Corsa....

You should set the max pre-rendered frames to one in the nvidia control panel it is a known Assetto corsa issue.

[Shahnewaz]

This is the definition of fraud.

 

After recently buying two 970s, if Nvidia does nothing to repay everyone that bought these cards, I will never buy from Nvidia again.

That's exactly how I feel. If the flaw is set in stone by Nvidia (ever so slightly is all I need, cause sure they'll want to minimize damage), then I'm done with them.

[Shahnewaz]

If the nvidia engineers are weighing in I just hope there is not some advanced spin machine trying to minimize damages and not get at the truth.  Even if performance hits are negligible, IF it's really true that the 970 cards can't use the full 4 GB of vram like other cards, that still sounds like deceptive advertising as it might have gotten people to choose different options if they had known there were technical differences in the memory accessibility.  Ryan tends to tread very carefully when talking to nvidia because I'm sure he wants to maintain access.  He can't go all Charlie on them like his semiaccurate "wood screws" coverage went.

 

http://semiaccurate.com/2009/10/01/nvidia-fakes-fermi-boards-gtc/

 

Note: Charlie is a great nvidia hater, but it's a treat to watch someone filled with such bile towards nvidia pile on with glee at their struggles.

I'm sure he has his reasons to hate Nvidia. Many of them are pretty obvious.

 

At least their articles are more believable/truthful/honest because they don't have to bend their articles to manufacturers' wills to maintain free access to review samples.

 

Oh yes it's a treat :lol:. Maybe it's overkill but if he exposes the truth in a humorous way, then it's all good .

[GoodBytes]

Checkout AnandTech new article about the issue.

http://www.anandtech.com/show/8935/geforce-gtx-970-correcting-the-specs-exploring-memory-allocation

First there is a correction on the specs from the marketing material.

And AnandTech explains in details, what is going on, as Nvidia statement was vague as it was aimed at the general public and not tech sites. This is cleared up inside.

All to say, it is by designed due to the cutting they made which is unique for this GPU architecture. And yes it has 4GB addressable memory. This is something we can see in later architectures and cards from Nvidia.

You may want to put this on the main post.

[Lewellyn]

If I'd really want to convince such an irrespective person who straight up resorts to ad hominems I would actually bother to show you some benches on Survarium which pretty much always maxes out my VRAM without dropping more than an average of 2-3 frames. Have fun with yourself from this point forward.

 

Oh, you're one of those people. Sorry to burst your wikipedia-logic crusade but calling someone an asshole is not an ad hominem. You just keep moving farther and farther away from my original point, probably because you don't have a leg to stand on. Data testing with only three points is not useful. End of.

[BiG StroOnZ]

Your attempts to insult my intelligence over the internet are small and only deserve a response merely because of the order I chose to respond is in relation to your statements.

 

And I've hear that answer before, hmm where was that? Oh yeah, everywhere else. Very original and I like how you added nothing more to the previous speculations.

 

That's better, but only just. 8/Eight.

 

And you question my intelligence... or did you state it as fact? :unsure:

 

The definition and use are already predetermined: http://en.wikipedia.org/wiki/Horsepower https://www.wordnik.com/words/horsepower The way you used them is in the informal sense and is therefore lacking in specific quantity. In short, it leaves my questioning the validity of your statements, preferring something a more accurate to the current seriousness of the topic at hand.

 

Anything else?

 

You don't understand basic answers to questions, then you get offended easily when people respond to you in a direct manner. Somehow considering it being disrespectful. 

 

Then when someone goes into detail and explains the answer to you. You tell me it's unoriginal, is merely a hypothesis, and you already knew it. Firstly, it's not a theory or speculation. Anyone with basic GPU knowledge knows this. It happened before the 970 and it will happen after the 970. TurboCache has been out for many, many years, they even titled it "TurboCache"

 

I question your intelligence when you don't understand a reply to you, when every single other person in the thread understood it crystal clear. What does that say? Clearly not something that could be objectified at that point.  Also, I like how you focus merely on that aspect of the conversation instead of the actual discussion that took place. Another Straw man argument entirely. 

 

Why are you linking the actual definition to Horsepower? It's figurative speech? What don't you understand here.  Do you really not know what figurative speech is?

 

Like I said before either you are a troll or you are just not very smart or a combination of both. I'm saying both. 

[Victorious Secret]

I get what your saying, and I still agree, but only to a point.  If people bought sli 970s for star citizen,  then on what marketing did they make their choice?

Do we even know what the minimum requirement for half decent graphics on SC is yet?

 

 

No, we don't. We won't know requirements for SC till much later. Right now theres only pie in the sky "tiers" of hardware for optimal experience, but no actual requirements. 

Also, LOL at people who are investing thousands into computer hardware for a game that won't enter its BETA till this summer/fall and won't have full on release till next year. I'll buy my hardware then, when its cheaper and even more powerful. 

[CreativeTHC]

they should just have said is a 3.5gb same as 480 is a 1.5... but "green team talks trash" such a shame.....

[PopReference]

You don't understand basic answers to questions, then you get offended easily when people respond to you in a direct manner. Somehow considering it being disrespectful. 

 

Then when someone goes into detail and explains the answer to you. You tell me it's unoriginal, is merely a hypothesis, and you already knew it. Firstly, it's not a theory or speculation. Anyone with basic GPU knowledge knows this. It happened before the 970 and it will happen after the 970. 

 

I question your intelligence when you don't understand a reply to you, when every single other person in the thread understood it crystal clear. What does that say? Clearly not something that could be objectified at that point. 

 

Why are you linking the actual definition to Horsepower? It's figurative speech? What don't  you understand here.  Do you really not know what figurative speech is?

 

Like I said before either you are a troll or you are just not very smart or a combination of both. I'm saying both. 

Okay, believe what you will. But I won't deal with a fool like you anymore.

[BiG StroOnZ]

Okay, believe what you will. But I won't deal with a fool like you anymore.

 

Why even respond then with a comment of that magnitude? That just opens up dialog further, clearly.

[TheMidnightNarwhal]

Meh, card is still great so I don't care.

[Lewellyn]

Meh, card is still great so I don't care.

 

What?! you mean you don't want a full refund + a free 980 and a Ferrari? How could you possibly not want to sue Nvidia for everything they own, that's the American way!... oh I see you're just a damn fanboy!

 

/s

[Danielm888]

This seems like a dick move to straight up lie about the specifications of your device and then say there was an "error" really? 

 

False Technical Specifications Memory Size: 4GB VRAM at 224GB/S

 

What Nvidia should have told us, Technical Specifications Memory Size: 3.5GB VRAM Primary at 224GB/s and Secondary reserve 0.5GB VRAM at 6-25GB/S as pointed out in the WAN show.

 

I don't own a 970 but If I was one of those people to drop a bunch of cash on 2 970's in SLI for high res gaming just to have FPS fall off the cliff and games stuttering past 3.5GB VRAM I'd be pretty disappointed to.

[mr moose]

This seems like a dick move to straight up lie about the specifications of your device and then say there was an "error" really? 

 

False Technical Specifications Memory Size: 4GB VRAM at 224GB/S

 

What Nvidia should have told us, Technical Specifications Memory Size: 3.5GB VRAM Primary at 224GB/s and Secondary reserve 0.5GB VRAM at 6-25GB/S as pointed out in the WAN show.

 

I don't own a 970 but If I was one of those people to drop a bunch of cash on 2 970's in SLI for high res gaming just to have FPS fall off the cliff and games stuttering past 3.5GB VRAM I'd be pretty disappointed to.

 

What they did wasn't ethical, absolutely, however anyone who spent $$ one two of these for better than 2560x1440 at highest quality settings didn't do their research.  Ram cliff or not, all the reviews said they couldn't do it so there is nothing really to get disappointed about.

[beebskadoo]

What they did wasn't ethical, absolutely, however anyone who spent $$ one two of these for better than 2560x1440 at highest quality settings didn't do their research.  Ram cliff or not, all the reviews said they couldn't do it so there is nothing really to get disappointed about.

hey moose <3

[mr moose]

hey moose <3

Hey o/

 

Come to join the discussion?

[Ryan Leech]

I'm thankful to have my single GB of vram, meanwhile people are floundering over an insignificant 512MB out of 4GB. This certainly isn't the first time a GPU hasn't been able to use all of its vram..and it certainly won't be the last.

[osbios]

Meh, card is still great so I don't care.

 

NVidia didn't think so.

[bogus]

What they did wasn't ethical, absolutely, however anyone who spent $$ one two of these for better than 2560x1440 at highest quality settings didn't do their research.  Ram cliff or not, all the reviews said they couldn't do it so there is nothing really to get disappointed about.

It's not just unethical - it's illegal.

In more competitive markets, NVIDIA would have had alot of lawsuits for their practices and missinformation. This isn't the first time they do it, neither will be the last - but this one is completely off the charts, this is the kind of stuff that get products banned from markets.