Jen-Hsun On GeForce GTX 970

H

HardOCP News

[H] News
Joined
Dec 31, 1969
Messages
0
Some of you are disappointed that we didn’t clearly describe the segmented memory of GeForce GTX 970 when we launched it. I can see why, so let me address it. We invented a new memory architecture in Maxwell. This new capability was created so that reduced-configurations of Maxwell can have a larger framebuffer – i.e., so that GTX 970 is not limited to 3GB, and can have an additional 1GB.

GTX 970 is a 4GB card. However, the upper 512MB of the additional 1GB is segmented and has reduced bandwidth. This is a good design because we were able to add an additional 1GB for GTX 970 and our software engineers can keep less frequently used data in the 512MB segment.

Unfortunately, we failed to communicate this internally to our marketing team, and externally to reviewers at launch. Since then, Jonah Alben, our senior vice president of hardware engineering, provided a technical description of the design, which was captured well by several editors. Here’s one example from The Tech Report.

Instead of being excited that we invented a way to increase memory of the GTX 970 from 3GB to 4GB, some were disappointed that we didn’t better describe the segmented nature of the architecture for that last 1GB of memory. This is understandable. But, let me be clear: Our only intention was to create the best GPU for you. We wanted GTX 970 to have 4GB of memory, as games are using more memory than ever.

The 4GB of memory on GTX 970 is used and useful to achieve the performance you are enjoying. And as ever, our engineers will continue to enhance game performance that you can regularly download using GeForce Experience. This new feature of Maxwell should have been clearly detailed from the beginning.

We won’t let this happen again. We’ll do a better job next time.

Jen-Hsun
 
MORE & SLOWER MEMORY ACCESS IS SUPPOSED TO BE A 970 "FEATURE" PEOPLE!

So there we have it.

Emm... Oye...

/facepalm
 
This is awesome. They are turning a marketing error into a technical engineering win. Is he considering a future in politics because he has amazing speech writing potential :D
 
4GB of RAM is 4GB, no matter how it's used. It depends on what the lawsuit claim is, because "Purposely misled for financial gain" and "Lied about product details" wouldn't pass muster in court.

I know it's fashionable to hate on NV for being an evil megacorp, but I don't think think fashion is a good replacement for fact.
 
Hmm. Maybe some of the user here need to understand that engineering always involves trade-offs. I'd rather have a 4GB card with a configurable caching algorithm for the last 512MB than a 3GB card (which almost guaranteed would have sold for the same price). The nice thing here is that the caching algorithm can be changed via drivers, to reduce the "noticeable" latency. I can fault Nvidia for lying in their specs, I cannot fault them for trying to engineer around limitations (probably imposed by the same upper execs).

/beginrant
What happened to when hardware/overclocking forums actually had knowledgeable users? Why are 90% of overclocking forum posts nowadays just bitching, ridiculous misinformation, and a general lack of knowledge? Is it because instead of spending days tweaking memory subtimings, perusing other users' findings, and then posting about them, people just use the newest omwtfbbqASUS awesome-button?
/endrant.
 
Oh, so now the CEO himself has something to say. He's telling us why this is a good design. Don't worry guys, we can take his word for it. And shame on us for not being more excited about it.
"Instead of being excited that we invented a way to increase memory of the GTX 970 from 3GB to 4GB, some were disappointed that we didn’t better describe the segmented nature of the architecture for that last 1GB of memory."

Wait, but hold on, it was for our benefit: "But, let me be clear: Our only intention was to create the best GPU for you."

And they promise to do a better job communicating with us in the future. They won't let this happen again.

Well I'll tell you wait. In the end, I paid significantly less for the card, so thank you... to Amazon.

Ha honestly though I understand they did it cuz of binning. I ain't got no problem with why they did it. Just gotta be up-front about it homes, before people actually spend money on the product.
 
schoolslave said:
Hmm. Maybe some of the user here need to understand that engineering always involves trade-offs. I'd rather have a 4GB card with a configurable caching algorithm for the last 512MB than a 3GB card (which almost guaranteed would have sold for the same price). The nice thing here is that the caching algorithm can be changed via drivers, to reduce the "noticeable" latency. I can fault Nvidia for lying in their specs, I cannot fault them for trying to engineer around limitations (probably imposed by the same upper execs).
Click to expand...

From my own experience, I did not notice any issue when my card used more than 3.5GB in Dying Light. Not sure if it's driver update or something, but it is certainly managing the last 512MB very well at least with Dying Light.
 
Sure because he's coming out now it sounds disingenuous, but back here in reality, if anything he says has a bit of truth to it, this WAS a good thing for consumers that turned bad because they definitely fucked up explaining it. Now i'm of the opinion that they purposefully fucked up explaining it, and just threw the card out there at a 4GB card and not a 3.5GB full speed + 512MB cache card.

If what he says is true though, this prevented it from being a 3GB 192-bit memory bandwidth card and allowed it to be a 3.5GB 224-bit bandwidth card. This DEFINITELY should have been communicated as an engineering win straight out the fucking gate though. I can't fathom why it was done any other way other than to mislead people.
 
schoolslave said:
/beginrant
What happened to when hardware/overclocking forums actually had knowledgeable users? Why are 90% of overclocking forum posts nowadays just bitching, ridiculous misinformation, and a general lack of knowledge? Is it because instead of spending days tweaking memory subtimings, perusing other users' findings, and then posting about them, people just use the newest omwtfbbqASUS awesome-button?
/endrant.
Click to expand...

I actually 100% agree. You forgot the other portion of user who claim to have uber hardware, so their e-peen size grows.
 
Hornet said:
From my own experience, I did not notice any issue when my card used more than 3.5GB in Dying Light. Not sure if it's driver update or something, but it is certainly managing the last 512MB very well at least with Dying Light.
Click to expand...

I to had great experience using the last 512MB in Dying Light. Evolve would get up to 3.7GB used and again, no problem. People like to hate and jump on the bandwagon. I am sure 50% of 970 haters aren't users and the other 40% are those that want something for nothing with this lawsuit. The other 10% are trying to get a midrange card to do something even the 980 struggles with. *Shakes head*

schoolslave said:
/beginrant
What happened to when hardware/overclocking forums actually had knowledgeable users? Why are 90% of overclocking forum posts nowadays just bitching, ridiculous misinformation, and a general lack of knowledge? Is it because instead of spending days tweaking memory subtimings, perusing other users' findings, and then posting about them, people just use the newest omwtfbbqASUS awesome-button?
/endrant.
Click to expand...

Right there with you brother, its sad...
 
Steve Jobs on Antenna Gate issue said:
We made a mistake in how our inidicate signal strength. It's a software bug
Click to expand...

Jen-Hsun on Memory Gate issue said:
We made a mistake on how we indicate memory performance.
Click to expand...

While few games will use that last 512 Megs of memory now, games are ever pushing the limit. And in 2,3 years, 3.5GB won't be enough. This make the obsolescence of the 970 that much quicker.

Lawsuit is justified IMHO.
 
I didn't know that marketing were the guys who would write down the specs on the video card BIOS, but now thanks to some posters here i know that!... i mean, it surely couldn't be that they tried to misrepresent the specs, it was just a marketing error.
:rolleyes:
 
Hornet said:
From my own experience, I did not notice any issue when my card used more than 3.5GB in Dying Light. Not sure if it's driver update or something, but it is certainly managing the last 512MB very well at least with Dying Light.
Click to expand...

The driver caps it to prevent it from using that last 512
 
Technically speaking the last 512MB is only a cache in the sense that it's a cache for system memory...and only if they program it correctly in the drivers in the first place to take advantage of it.
 
DigitalGriffin said:
The driver caps it to prevent it from using that last 512
Click to expand...

I dont think you read his post right, he said it USED more than 3.5GB of VRAM. And that's BS anyway as my card went over 3.5GB of VRAM in it too. So many forum members posting things they have never tried or don't understand.
 
DigitalGriffin said:
Technically speaking the last 512MB is only a cache in the sense that it's a cache for system memory...and only if they program it correctly in the drivers in the first place to take advantage of it.
Click to expand...

I don't want to pour salt on some wounds but regarding this, has anyone else seen the apparent performance drop off of Kepler cards in newer releases, meaning, releases since Maxwell came out?

Seeing that, does anyone really hold any hope of that cache being properly implemented and not ignored the second they come with something new?, will people be able to be pissed off at that point or would the concern still be invalid?
 
It's kind of like buying a car that advertises 400 hp, and then you later find out it really only has 350 hp. It still goes just as fast, but they still lied to you.
 
Stev3FrencH said:
I dont think you read his post right, he said it USED more than 3.5GB of VRAM. And that's BS anyway as my card went over 3.5GB of VRAM in it too. So many forum members posting things they have never tried or don't understand.
Click to expand...

There are many types of calls to provision memory in D3D. Some are for user programs. Some are for shaders. Some are for texture memory. Some are for buffer memory. Some are for stencil buffer memory. Some are for depth memory. (related to buffer)....etc....

From the technical documentation it's looking like they are limiting the memory on some of these operations. Just because it's being used (>3.5), doesn't mean that it's being used by the D3D program, but by the driver as cache.

You can kind of think of it like virtual memory in windows. Sure you have access to 8GB of memory, but 4GB of it is virtual. And while the computer can hold 6GB of running programs, 2GB of that sits in a pseudo memory space which is a crap ton slower.

Understand the difference? Except with the 970 it sounds/looks as if that last 512 MB is hard locked as main system memory cache and not accessible to D3D buffers directly, which slows performance when D3D memory request exceeds 3.5GB.
 
Revdarian said:
I don't want to pour salt on some wounds but regarding this, has anyone else seen the apparent performance drop off of Kepler cards in newer releases, meaning, releases since Maxwell came out?

Seeing that, does anyone really hold any hope of that cache being properly implemented and not ignored the second they come with something new?, will people be able to be pissed off at that point or would the concern still be invalid?
Click to expand...

Technically speaking it shouldn't be that hard. It would be no more complex then any other caching algorithm. (As D3D drivers are responsible for memory allocation and there's a fixed set of memory allocation calls available)
 
The really funny--and strange--chapter in all of this is that the 980 is limited in exactly the same way as the 970; it, too, suffers mass performance drops in the last 500MBs of its on-board ram. It's slightly better because of the way the 970 configures its ram physically on the card, with performance drops of up to 50% for the 970 versus drops of 47% for the 980 in the same games. nVidia originally published the info, but for some reason few seem to understand that both the 980 & the 970's last 500MBs of on-board ram is hobbled & segmented in exactly the same way. Both cards guarantee 4GBs; both deliver 3.5GB that is accessible by the full width of each bus on each card.

Technically, nVidia cannot lose in court because they have delivered 4GBs of ram on each card--that is indisputable. All of it is available to the GPU; indisputable as well. All of it is not available to the 9x0's maximum bus width, however, and that's the problem. I have to say though that I think 99% of the people who own these cards will never see such huge slowdowns, because one really has to work to load the cards in such a way that the 3.5GB boundary is exceeded during a game. In normal gaming this will usually never happen.

nVidia would have had to increase the on-board bus width to access the entire 4GBs at the maximum possible throughput, which would have increased the cost and selling price of these cards somewhat--so nVidia opted to save a few bucks, and somebody noticed...;)
 
The Lamb said:
4GB of RAM is 4GB, no matter how it's used. It depends on what the lawsuit claim is, because "Purposely misled for financial gain" and "Lied about product details" wouldn't pass muster in court.

I know it's fashionable to hate on NV for being an evil megacorp, but I don't think think fashion is a good replacement for fact.
Click to expand...

According to that logic, one could tape a 1GB DIMM onto the heat sink to improve cooling
 
I really have to wonder if this type of memory segmenting has been done before on other NVidia cards (or AMD, for that matter).
 
my factory OC GTX970 rocks...

handles EVERYTHING I throw at it and barely makes any noise

I'm not seeing any issues
 
The Lamb said:
4GB of RAM is 4GB, no matter how it's used. It depends on what the lawsuit claim is, because "Purposely misled for financial gain" and "Lied about product details" wouldn't pass muster in court.
Click to expand...
What would you do if they said the GTX 970 had 4GB but turns out that it's all slow ass PC100 chips? What would you call that?
 
Stev3FrencH said:
I dont think you read his post right, he said it USED more than 3.5GB of VRAM. And that's BS anyway as my card went over 3.5GB of VRAM in it too. So many forum members posting things they have never tried or don't understand.
Click to expand...

I think what he means is that the driver locks out the VRAM by allocating it to nothing.
 
Stev3FrencH said:
People like to hate and jump on the bandwagon. I am sure 50% of 970 haters aren't users and the other 40% are those that want something for nothing with this lawsuit. The other 10% are trying to get a midrange card to do something even the 980 struggles with. *Shakes head*
Click to expand...

I agree 100% with you, and most of the complaints seem to come from people trying to use these cards with 4K or DSR or both in SLI configs. I have seen a lot of the same attitude about how the 960 is only 2GB/128-bit.

Mountain <----------------- molehill
 
Yakk said:
MORE & SLOWER MEMORY ACCESS IS SUPPOSED TO BE A 970 "FEATURE" PEOPLE!

So there we have it.

Emm... Oye...

/facepalm
Click to expand...

Cerulean said:
What would you do if they said the GTX 970 had 4GB but turns out that it's all slow ass PC100 chips? What would you call that?
Click to expand...

The problem with both these statements is that it is not slower memory. It has 3.5GB of higher bandwidth memory than the 3GB it would have had otherwise. Then the last 512 is cached. So it does not have 4GB of the full bandwidth, but its still more than it would have had and cards of that nature had in the past. So all in all, it is an improvement.

What I find funny is how many people are rage ranting about a marketing statement. It isn't even that bad compared to all the other marketing crap that goes on out there all the time. It just shows how ridiculous our society has come to that they are suing based on a claim that is 100% true, but not in the way they expect it... I mean come on. And as others have said 90% or more of those complaining won't even notice the difference. It is a sad sad day in our society that this is what we have come to...
 
To paraphrase, "Let me be clear! That 400 horsepower car you bought? When it hits 350 horsepower, the car will slow down as it gets up to 400 horsepower utilization."

Frankly, they are just sounding more stupid the more they open their mouths.
 
This shit is ridiculous.

If you ever sit around and wonder why developers abandon the PC gaming market or think of it as an after thought.... look no further than this forum/thread.

It is shit like this, making a mountain out of a molehill. So now there will be a lawsuit and a law firm will get a gigantic check. Meanwhile here is you free game.
 
Cerulean said:
What would you do if they said the GTX 970 had 4GB but turns out that it's all slow ass PC100 chips? What would you call that?
Click to expand...

If you build something like that, it would perform with single digit fps in review and nobody will buy the card. That's really the whole reason why people purchase hardware, we don't purchase a hardware because it has a fancy looking hardware architecture, we purchase hardware based on it's performance.

Knowing how a GPU works doesn't change the performance of the GPU, all the figures you saw in [H]'s review remains valid today as they do when the card was released, and we make our purchased based on those performance figures, not how many TMU was present in the GPU.
 
Assuming that this was in fact initially an error in communication, my problem with NVIDIA, and I'm guessing what will be argued in court is the following, after the review dropped on launch day, it is extremely unlikely that someone this error was not observed by an employee and then communicated to a higher up. Then they just remained silent, or were glad to make it through Q4 without this coming to light and affecting sales.

I bought on launch day, and a big part of upgrading from a gtx 780 was the extra bit of ram, since performance was a wash. I would not have upgraded had I known that the card only had 512mb of ram more.
 
Mr. Bluntman said:
I agree 100% with you, and most of the complaints seem to come from people trying to use these cards with 4K or DSR or both in SLI configs. I have seen a lot of the same attitude about how the 960 is only 2GB/128-bit.

Mountain <----------------- molehill
Click to expand...

However, the 960 was never advertised as having 4GB of VRAM with a 256-bit bus that turns out to really be segmented into a 3.5GB block of VRAM with a 224-bit bus and 512MB of VRAM just piggy-backing for the ride. I realize that specs do not necessarily equate to performance, but thanks to the way that these 970s were marketed, we have a situation where a lot of people had expectations of longevity or better higher-resolution performance (better SLI performance even) for a lower price based upon the specifications that were being advertised.

As a 970 owner myself, I am disappointed that the cards were marketed incorrectly, but I am not disappointed with the performance of my card for my uses. I am not actively engaging in the class-action lawsuit nor am I expecting something for nothing. I am less inclined to be an early-adopter going forward and that is about it.
 
My BS meter just pegged into the red zone. Maybe the solution is for Nvidia to install some oversight into its marketing department.
 
amddragonpc said:
To paraphrase, "Let me be clear! That 400 horsepower car you bought? When it hits 350 horsepower, the car will slow down as it gets up to 400 horsepower utilization."

Frankly, they are just sounding more stupid the more they open their mouths.
Click to expand...

FFS .... My 5.7L Hemi Ram is advertised as having 400HP. You know what I get at the wheels? 315HP.

*breaks out pitchforks and anal lube* I WANT MY 85 HORSEPOWER!!

Back in reality the advertised horsepower of an engine in a vehicle is NEVER what you are putting to the pavement. Using a car engine analogy does not work here.
 
Sounds like a spin. He would've been better off not adding fuel to the flame.
 
You must log in or register to reply here.
Back
Top