WCCftech: Memory allocation problem with GTX 970 [UPDATE] PCPer: NVidia response

[Final8ty]

Im more for making sure something like this does not happen again and that AMD dont start getting any bright ideas for doing the same.

If we all kept quiet you can be damn sure NV would do it again.

If the card had explicitly documented 3.5GB+512MB then i would have no issue as people knew what they were buying so could not complain.

 

[garagisti]

Who cares if they knew or not? That end result matters, that they sold a card with bogus specs. Either they were deceitful or oblivious, both aren't looking good for them. I really like Nvidia cards, but I'm not giving them a pass on this. I would have had a lot of respect for them if they came out on their own and fixed the issue before they were forced to. Now I don't know what to think. It makes me apprehensive of what I am buying in the future.

As a paying customer, you should properly care whether they knew of it, and well it's reasonable to think they did. You paid close to a thousand dollars to get their products, and surely will have purchased their products in past (as did i). You should be concerned how they are treating you as an individual, and collectively. It's a good start by Nvidia where their employee(s) is(are) helping people out on forums. Now may be it's time to take it to next level, by addressing other customers who are not on the various forums. That's what I'd like to see, if I'm to ever consider buying/ recommending their products. I know my recommendation is may be not many sales, but I resent being taken for a ride.

 

[garagisti]

To be false advertising. It would have to be on official boxes, ads and sites wouldnt it?

While nVidia may have given wrong information to review sites. Its actually the review sites that holds all responsibility for the information they publish. Unless some are secretly run by nVidia.

So are you suggesting that all the cards which were wrapped in nice shiny boxes suggesting 4gb has nothing to do with Nvidia? That it in no shape or form makes them liable? Is that what I'm to understand?

Well I'd like to see them try this in some European courts...

 

[garagisti]

Wouldn't they have to verify that the information they pass along is truthful and accurate? I imagine that many would copy&paste it in though.

When you will open a card, you will find the 4gb ram on the card, but pray tell how will you know that ROPs are cut and so on? Especially when speacialised tuning software etc also doesn't reveal anything? What do you propose to the review sites?

By the way, I'm completely appalled by "so what" attitude of some of the sites, but that is a separate issue.

 

[JoeRambo]

This is untrue.

As you can see only the last GB really has this problem. The memory controller can access the first 512 MB (which it does). However, whenever it accesses one 512 MB section, it cannot access the other. The other 3 64 bit controllers function normally.

I am not very familiar in how stuff works, but in my understanding you got 8 32bit controllers, 7 crossbars, and stripe accesses via those crossbars. The problem beeing that 8th chip is driven through its own crossbar, breaking the striping scheme. So this has following implications

1) 196GB/sec (7GHz * 7 crossbars * 32bit ) if those 3.5GB are read
XOR
2) 28GB/s (7Ghz * 1 crossbar* 32bit ) if that last 512MB are read, denying any striping, chewing it 32bits at a time.

Once "7"th crossbar is busy accessing 512megs, striping scheme is impossible, cause it can't generate access for 512 and stripe access at same time. Hence the "XOR" part.

 

[cmdrdredd]

When you will open a card, you will find the 4gb ram on the card, but pray tell how will you know that ROPs are cut and so on? Especially when speacialised tuning software etc also doesn't reveal anything? What do you propose to the review sites?

By the way, I'm completely appalled by "so what" attitude of some of the sites, but that is a separate issue.

Apparently people figured out something without opening the card and knowing the ROP count. None of the tech reviewers did.

 

[garagisti]

Apparently people figured out something without opening the card and knowing the ROP count. None of the tech reviewers did.

Ahem, because review sites don't specifically check CUDA performance. Well there may be a token test, but nothing too serious. These are gaming cards and due diligance has to go no further than,"so is this correct?" CUDA programmers were already not happy shortly after launch is what I read elsewhere, and were returning cards. It wasn't until Nai came along this became public knowledge.

 

[ShintaiDK]

So are you suggesting that all the cards which were wrapped in nice shiny boxes suggesting 4gb has nothing to do with Nvidia? That it in no shape or form makes them liable? Is that what I'm to understand?

Well I'd like to see them try this in some European courts...

The cards got 4GB, all accessible. So nothing wrong there. And it wouldnt even reach a european court if thats the issue you got. Because then there isnt any.

It was in relation to ROPs and L2.

 

[garagisti]

The cards got 4GB, all accessible. So nothing wrong there. And it wouldnt even reach a european court if thats the issue you got. Because then there isnt any.

It was in relation to ROPs and L2.

Well, there's also the matter of 256bit on the box, and that 4gb isn't described as two different pools. The card doesn't work as one would reasonably expect a card with 4gb of vram. Surely you could argue that it's not the norm, but the exception, and as such is misleading as it lacks any description what so ever on retail package.

 

[ShintaiDK]

Well, there's also the matter of 256bit on the box, and that 4gb isn't described as two different pools. The card doesn't work as one would reasonably expect a card with 4gb of vram. Surely you could argue that it's not the norm, but the exception, and as such is misleading as it lacks any description what so ever on retail package.

Its still 256bit and its still 4GB.

You wouldnt take AMD to court for selling the FX 8350 for example as an 8 core would you? Most likely not because it would just be plain silly and you would lose. Even tho we can debate the semantics.

 

[SlowSpyder]

The cards got 4GB, all accessible. So nothing wrong there. And it wouldnt even reach a european court if thats the issue you got. Because then there isnt any.

It was in relation to ROPs and L2.

Wouldn't the texture fillrate be different than what is on Nvidia's site due to the lower number of ROP's? I think even if so, that's really sort of splitting hairs. I'm not sure, but from what I've read I thought that number was tied to ROP's. I'm no expert, certainly could be mistaken.

But, I think seeing how Nvidia has responded, it's obvious that something was wrong in the way the GTX 970 was launched and advertised. They seem to be busy. I don't think a lawsuit will come of this, not sure it would fly. It isn't uncommon for a company to have incorrect specs on something, then immediately change it when that is found. I think for a lawsuit to go forward there would have to be evidence that Nvidia knowingly gave incorrect information and kept it there. I doubt it'd go that far or could be proven.

But, I wonder if Nvidia might be lawfully obligated to take returns (or their partners) because of the wrong specs, if someone wants a return..? Personally, I think that's the right thing to do. Of course I know about as much about business law as the most people, that is to say not much.

*edit - Just remembered this, it has some parallels. Cobra recall. Free upgrades to GTX 980's to get to the advertised spec's?

 

[garagisti]

Its still 256bit and its still 4GB.

You wouldnt take AMD to court for selling the FX 8350 for example as an 8 core would you? Most likely not because it would just be plain silly and you would lose. Even tho we can debate the semantics.

And how many months/ years did they kept their definition of core quiet? No, this is not semantics.

 

[ShintaiDK]

Wouldn't the texture fillrate be different than what is on Nvidia's site due to the lower number of ROP's? I think even if so, that's really sort of splitting hairs. I'm not sure, but from what I've read I thought that number was tied to ROP's. I'm no expert, certainly could be mistaken.

Isnt the texture fillrate the TMUs job? And its 109 vs 144 for GTX970 and GTX980.

 

[ShintaiDK]

And how many months/ years did they kept their definition of core quiet? No, this is not semantics.

I dont recall them talking about the scaling penalty in bulldozer. But it may just have been me.

Or the no compromise, 2 thread performance claim with SR.

 

[Grooveriding]

Wouldn't the texture fillrate be different than what is on Nvidia's site due to the lower number of ROP's? I think even if so, that's really sort of splitting hairs. I'm not sure, but from what I've read I thought that number was tied to ROP's. I'm no expert, certainly could be mistaken.

But, I think seeing how Nvidia has responded, it's obvious that something was wrong in the way the GTX 970 was launched and advertised. They seem to be busy. I don't think a lawsuit will come of this, not sure it would fly. It isn't uncommon for a company to have incorrect specs on something, then immediately change it when that is found. I think for a lawsuit to go forward there would have to be evidence that Nvidia knowingly gave incorrect information and kept it there. I doubt it'd go that far or could be proven.

But, I wonder if Nvidia might be lawfully obligated to take returns (or their partners) because of the wrong specs, if someone wants a return..? Personally, I think that's the right thing to do. Of course I know about as much about business law as the most people, that is to say not much.

*edit - Just remembered this, it has some parallels. Cobra recall. Free upgrades to GTX 980's to get to the advertised spec's?

Nvidia lists memory bandwidth as 224GB/s for the card. That is a false number when you are using enough memory with the hidden last .5GB memory issue. Not including the false specs they released to review sites.

It's really indefensible. They lied outright or via omission, got caught and that is the end of story.

 

[KaRLiToS]

Just ignore that guy.
At the begginning of the thread he was accusing the OP of being wrong and now that Nvidia admitted their mistakes it is not their responsability. Pffff.

 

[SlowSpyder]

Isnt the texture fillrate the TMUs job? And its 109 vs 144 for GTX970 and GTX980.

Wikipedia article on fillrate.

The term fillrate usually refers to the number of pixels a video card can render and write to video memory in a second. In this case, fillrates are given in megapixels per second or in gigapixels per second (in the case of newer cards), and they are obtained by multiplying the number of raster operations (ROPs) by the clock frequency of the graphics processor unit (GPU) of a video card.

I'll let someone more knowledgeable step in if they want, I really don't know enough about the inner workings of GPU's to really know. But, it sounds like it, I thought it was. But again, hardly an expert.

 

[SlowSpyder]

Nvidia lists memory bandwidth as 224GB/s for the card. That is a false number when you are using enough memory with the hidden last .5GB memory issue. Not including the false specs they released to review sites.

It's really indefensible. They lied outright or via omission, got caught and that is the end of story.

I agree, they're 100% at fault here. And I don't believe that someone at Nvidia didn't know, they had to. Reviews and specs on sites like Newegg would have had to have been seen by someone familiar with the way the GTX 970 GPU worked. But, I just don't know how that'd stand up in a lawsuit. I think Nvidia should offer upgrades or refunds to people.

 

[ShintaiDK]

Wikipedia article on fillrate.

I'll let someone more knowledgeable step in if they want, I really don't know enough about the inner workings of GPU's to really know. But, it sounds like it, I thought it was. But again, hardly an expert.

http://en.wikipedia.org/wiki/Texture_mapping_unit#Texture_Fill_Rate

GTX980 128 TMU*1126Mhz=144G/s
GTX970 104 TMU *1050Mhz=109G/s

 

[garagisti]

I dont recall them talking about the scaling penalty in bulldozer. But it may just have been me.

Or the no compromise, 2 thread performance claim with SR.

Ahem, they made some disclosures on the architecture, benchmarks came, people decided what they will buy. Here, benchmarks came, people bought stuff, people made certain discoveries, then Nvidia made disclosures. The difference between two situations is plainly put, very obvious.

Just ignore that guy.
At the begginning of the thread he was accusing the OP of being wrong and now that Nvidia admitted their mistakes it is not their responsability. Pffff.

Good advice.

 

[Silverforce11]

Nvidia lists memory bandwidth as 224GB/s for the card. That is a false number when you are using enough memory with the hidden last .5GB memory issue. Not including the false specs they released to review sites.

It's really indefensible. They lied outright or via omission, got caught and that is the end of story.

According to AT and other sites, it's also just PLAIN FALSE. Because its only ~196GB/sec for the first 3.5gb segment and 28GB/sec for the 0.5gb segment.

It's not 224GB/s because if they are reading from the 0.5gb segment, they can't from the 3.5gb.

So that's just outright false advertisement.

TR, Hardware.fr noticed during the 970 launch that its bandwidth and fillrate is not as claimed, but they didn't know why, and just went with what NV told them (cos, why would a company lie about such specs??). Blame reviewers? Don't be daft please.

 

[B-Riz]

The full 4GB is not accessible across the advertised 256-bit bus at once, that is the issue at hand.

They are being shady and using software to make up for a known design limitation; only being able to access 3.5GB because of 56 ROPs.

 

[komatta]

Frametime analysis by PC Per is now up.

 

[Silverforce11]

Frametime analysis by PC Per is now up.

Its interesting they actually observe more frame time variance in BF4 which is on Frostbite, that is known to allocate extra vram for LOD purposes which isn't required*. In such a game, I would expect no effect.

The frame time spikes should only occur in games that load the vram with textures that it needs, ie Mordor, Skyrim + Mods, Arma 3. Or now Dying Light with texture pop ins.

* Because 3GB cards run it fine, and Johan demo their dynamic vram allocation awhile ago to show LOD changes, it will adapt to your vram capacity.

 

[cmdrdredd]

Frametime analysis by PC Per is now up.

Pretty much what I expected with a single card. Thanks for the link.