• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

Nvidia responds to GTX 970 memory issue

cyberheater

PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 Xbone PS4 PS4
So you bought a card 4 months ago, you're happy with it, and it has been performing in line with your expectations, but suddenly now it isn't?

That's completely not the point. nVidia advertised the GTX 970 with a certain specifications. The reason why they do that is because consumers use those specs to make purchasing decisions. Today they have revised down those specifications and admitted it made a mistake.

Folks have the right to feel that they were mis-sold that card and seek redress.

this...

and this is over a .5 discrepancy? maybe?

GAF you lookin' real bitchy right now.

Some quality corporate ball licking you've got going there.
 

rav

Member
So you bought a card 4 months ago, you're happy with it, and it has been performing in line with your expectations, but suddenly now it isn't?

So as I've noticed in 4k, lower texture resolutions aren't that pretty. (CS:GO I'm looking at you.) Going forward developers expecting people to run at 4k are going to have even higher resolution textures. Guess what? Higher resolution textures take up more memory, and sooner than you realize you're constantly over that 3.5GB threshold because developers figure if you have a good card with 4GB of ram, they should use it!

Well, now in this one case, having a 4GB card is not great, because if it does use anything beyond 3.5GB as shown in benchmarks, you'll see a huge performance hit.

Edit: I tend to wait a few generations between buying graphics cards. (I upgraded from a 6870.) I now expect that this card in particular does not have the same lengthy cycle of my usual purchases.
 

Joejoe123

Neo Member
This could be a pretty serious issue if game developers over the next couple years target 4gb vram for their high/ultra spec.
 
I wonder what would have been the reaction of the same people if it had been AMD that had boondoggled their consumers. :p

Never had any good experience with AMD ever, so having it perform badly is expected. It is the price you pay for paying less, I guess.

That's why I stopped buying it, though.
 
I'd be interested in seeing an option to turn off the last 1/2 gig in the driver. But I guess it'd just swap to system ram at that point and be slower anyways.

I'm not so sure about that:

Anandtech said:
This in turn is why the 224GB/sec memory bandwidth number for the GTX 970 is technically correct and yet still not entirely useful as we move past the memory controllers, as it is not possible to actually get that much bandwidth at once on the read side. GTX 970 can read the 3.5GB segment at 196GB/sec (7GHz * 7 ports * 32-bits), or it can read the 512MB segment at 28GB/sec, but not both at once; it is a true XOR situation. Furthermore because the 512MB segment cannot be read at the same time as the 3.5GB segment, reading this segment blocks accessing the 3.5GB segment for that cycle, further reducing the effective memory bandwidth of the card. The larger the percentage of the time the crossbar is reading the 512MB segment, the lower the effective memory bandwidth from the 3.5GB segment.

If I understand it correctly: Any time the card is accessing the 0.5GB part it will completely halt access to the other 3.5GB, so it gimps the overall bandwidth significantly. I don't think it's clear cut that having the 0.5GB partition accessible is always better than having that data in system RAM instead. I'd guess it depends on the particular application/game.

I'd imagine having full speed access to 3.5GB + slow access from system RAM combined is faster than ONLY having access to the slow 0.5GB VRAM and nothing else in a given read cycle.

Can someone more tech-savvy on this comment on this?
 

Salaadin

Member
So you bought a card 4 months ago, you're happy with it, and it has been performing in line with your expectations, but suddenly now it isn't?
No.

I like my 970. It's treated me well over 4 months. But I don't like the idea of big corporations lying to consumers and reviewers and thinking they can get away with it.


this...

and this is over a .5 discrepancy? maybe?

GAF you lookin' real bitchy right now.
This isn't about .5 GBs. Why is this hard to understand?
 

Rafterman

Banned
I'm not so sure about that:



If I understand it correctly: Any time the card is accessing the 0.5GB part it will completely halt access to the other 3.5GB, so it gimps the overall bandwidth significantly. So I don't think it's clear cut that having the 0.5GB partition accessible is always better than reading it from system RAM instead. I'd guess it depends on the particular application/game.

You don't.

Even the .5gb, that is much slower than the 3.5gb, partition is significantly faster than system ram access and will always be so. If the entire card were equipped with only this slow vram it would still be significantly faster than accessing system ram.
 

XBP

Member
I understand people's frustration with this and can completely sympathize with it. Heck, I'm extremely disappointed as well that my card cannot access the full 4GB vram that the 980 can at all times but can we please stop saying that this gimps the 970 and makes it a underpowered card?

The 970 is still the same card that every single reviewer praised non stop about. Its still extremely wallet friendly and performs brilliantly even when compared to a 980. Despite this issue coming to light, I dont feel that the extra 200$ for the 980 is justified at all for the performance gain that it'll bring.

I also dont understand some of the posts saying that the 970 isn't that great for 4k anymore. From what I understood when I bought it, the card was never supposed to be great at that resolution (even the 980 in SLI wont give you 60+ FPS in every game @4k without sacrificing effects and thats a 1000$+ setup). At 1080p, a 970 will keep on performing great today and for the foreseeable future so can we please stop saying that this doesn't hold true anymore?

Oh and the people saying they wont buy the 970. What will you guys buy in this price range now that we have this information? The R9 290x? If yes, then that card's 8GB version doesn't perform any better than a GTX 970.

http://www.tomshardware.com/reviews/sapphire-vapor-x-r9-290x-8gb,3977-3.html

By the time that 8GB of VRAM is required for any game, a 970 and 290x will not be powerful enough to max out every game at 1080p anyway.
 
You don't.

Even the .5gb, that is much slower than the 3.5gb, partition is significantly faster than system ram access and will always be so. If the entire card were equipped with only this slow vram it would still be significantly faster than accessing system ram.

That I am aware of. Yes it is still 4x faster than system ram over PCI-Express.

But the article says the card can never access both the 0.5gb and the 3.5gb at the same time. So if in a given scenario the game engine needs to access assets that are located in both the 3.5gb part and the 0.5gb part, it will have to do one at a time.

My question is: Can the card access its 3.5gb part and system ram in the same read cycle? If yes, then can't there be a scenarios where it would be advantageous to have continuous access to the full bandwidth of the 3.5gb while occasionally dipping into system ram for lower priority stuff, instead of completely stopping any transfer from the 3.5gb every time it needs to make a call to the 0.5gb part?

I'm not a hardware expert and I haven't crunched the numbers to see how the math turns out in both cases. I'd like to see it though.

Of course it all falls apart if the answer to my bolded question is No.
 

XBP

Member
NVIDIA CAO David M. Shannon dumps nearly $0.5 million in stock 3 days ago.

Coincidence?

http://www.wkrb13.com/markets/468447/nvidia-cao-david-m-shannon-sells-21400-shares-nvda/

I'm pretty sure it is. This is not going to cause Nvidia any major problems at all. The most they will do is offer people free games (which will satisfy the majority of the people) or some sort of a step up program to get a GTX 980. If the RROD fiasco didn't cause microsoft to go under I dont see how a problem with a single GPU can result in Nvidia's CAO fearing for the company's financial state.
 

Scarlego

Neo Member
I was inspired today.
waUNUhR.png
 

LilJoka

Member
You don't.

Even the .5gb, that is much slower than the 3.5gb, partition is significantly faster than system ram access and will always be so. If the entire card were equipped with only this slow vram it would still be significantly faster than accessing system ram.

So why don't nvidia use 20GB/s VRAM on GPUs? Why is 3.5Gb of the VRAM specced to 150GB/s? Sounds like nvidia are wasting money? /sarcasm

Even if it's faster than system ram it's irrelevant since it's GDDR5, we expect the specified 224GB/s bandwidth.

And since when do we ever want to use system ram as VRAM? The point is it's only slightly quicker than system ram, that's the whole point of having GDDR5, for much much faster speeds...
 

Bastables

Member
I understand people's frustration with this and can completely sympathize with it. Heck, I'm extremely disappointed as well that my card cannot access the full 4GB vram that the 980 can at all times but can we please stop saying that this gimps the 970 and makes it a underpowered card?

The 970 is still the same card that every single reviewer praised non stop about. Its still extremely wallet friendly and performs brilliantly even when compared to a 980. Despite this issue coming to light, I dont feel that the extra 200$ for the 980 is justified at all for the performance gain that it'll bring.

I also dont understand some of the posts saying that the 970 isn't that great for 4k anymore. From what I understood when I bought it, the card was never supposed to be great at that resolution (even the 980 in SLI wont give you 60+ FPS in every game @4k without sacrificing effects and thats a 1000$+ setup). At 1080p, a 970 will keep on performing great today and for the foreseeable future so can we please stop saying that this doesn't hold true anymore?

Oh and the people saying they wont buy the 970. What will you guys buy in this price range now that we have this information? The R9 290x? If yes, then that card's 8GB version doesn't perform any better than a GTX 970.

http://www.tomshardware.com/reviews/sapphire-vapor-x-r9-290x-8gb,3977-3.html

By the time that 8GB of VRAM is required for any game, a 970 and 290x will not be powerful enough to max out every game at 1080p anyway.

The 290x Seems to trade pretty well at 1440p + resolutions vs the 970, is cheaper but is louder and greater power draw, but then it also has all of it's 4 gigs, which might explain why it does better at the higher resolutions http://www.anandtech.com/bench/product/1059?vs=1355

Im now looking at a 290x or 290.
 
I understand people's frustration with this and can completely sympathize with it. Heck, I'm extremely disappointed as well that my card cannot access the full 4GB vram that the 980 can at all times but can we please stop saying that this gimps the 970 and makes it a underpowered card?

The 970 is still the same card that every single reviewer praised non stop about. Its still extremely wallet friendly and performs brilliantly even when compared to a 980. Despite this issue coming to light, I dont feel that the extra 200$ for the 980 is justified at all for the performance gain that it'll bring.

I also dont understand some of the posts saying that the 970 isn't that great for 4k anymore. From what I understood when I bought it, the card was never supposed to be great at that resolution (even the 980 in SLI wont give you 60+ FPS in every game @4k without sacrificing effects and thats a 1000$+ setup). At 1080p, a 970 will keep on performing great today and for the foreseeable future so can we please stop saying that this doesn't hold true anymore?

Oh and the people saying they wont buy the 970. What will you guys buy in this price range now that we have this information? The R9 290x? If yes, then that card's 8GB version doesn't perform any better than a GTX 970.

http://www.tomshardware.com/reviews/sapphire-vapor-x-r9-290x-8gb,3977-3.html

By the time that 8GB of VRAM is required for any game, a 970 and 290x will not be powerful enough to max out every game at 1080p anyway.

I didn't buy a 970 just for today's games, I also bought it for TOMORROW'S. This isn't hard.
 

cyberheater

PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 Xbone PS4 PS4
I'm pretty sure it is. This is not going to cause Nvidia any major problems at all. The most they will do is offer people free games (which will satisfy the majority of the people) or some sort of a step up program to get a GTX 980. If the RROD fiasco didn't cause microsoft to go under I dont see how a problem with a single GPU can result in Nvidia's CAO fearing for the company's financial state.

It really depends how bad it gets, how nVidia continues to respond and the markets assessment and reaction to it.

Remember. IBM once had a relatively healthy hard drive division and one product with a high failure rate essentially put it under and they sold up to Hitachi. Just saying...
 

Mikeside

Member
It doesn't really matter why anyone bought the card.

It was advertised as something that it isn't & that's not OK. There's no legitimate way to turn the blame onto the consumer here.


NVIDIA misled about exactly what this card is & they should definitely have to compensate people who bought them on the promise that wasn't delivered.


I've got 2x970s in SLI and I really hope I can upgrade them to 980s for a reduced price, get a partial refund or similar.

I fully expect we'll just get a gratis game or something, but it's not REALLY good enough.
 

XBP

Member
The 290x Seems to trade pretty well at 1440p + resolutions vs the 970, is cheaper but is louder and greater power draw, but then it also has all of it's 4 gigs, which might explain why it does better at the higher resolutions http://www.anandtech.com/bench/product/1059?vs=1355

Im now looking at a 290x or 290.

At 1440p

Metro LL = Same
COH2 = 290x
Bioshock Infinite = 970
Battlefield 4 = Same
Crysis 3 = Same
Warhead = 290x
Total War Rome = 970 (but almost same)
Thief = Same
Grid 2 = 970 (but almost same)

290x = 2 clear win
970 = 1 clear and 2 barely wins

Everything else is same between the cards. If you take into account the latest MFAA tech that Nvidia added to almost every DX10 and 11 game in the last driver update, you get a further 10-15% FPS boost in all games with the same IQ.

I didn't buy a 970 just for today's games, I also bought it for TOMORROW'S. This isn't hard.

Yes, I didn't either. This however does not significantly change how tomorrow's games will perform on this card. There is a change yes, maybe 5-10%, but that isn't something which Nvidia cant work around with better drivers and performance.

It really depends how bad it gets, how nVidia continues to respond and the markets assessment and reaction to it.

Remember. IBM once had a relatively healthy hard drive division and one product with a high failure rate essentially put it under and they sold up to Hitachi. Just saying...

Apples and oranges, not comparable at all.


It doesn't really matter why anyone bought the card.

It was advertised as something that it isn't & that's not OK. There's no legitimate way to turn the blame onto the consumer here.


NVIDIA misled about exactly what this card is & they should definitely have to compensate people who bought them on the promise that wasn't delivered.


I've got 2x970s in SLI and I really hope I can upgrade them to 980s for a reduced price, get a partial refund or similar.

I fully expect we'll just get a gratis game or something, but it's not REALLY good enough.

I agree with this. Its pathetic of Nvidia to lie to its customers when they could have been upfront about this whole situation. Expecting compensation is completely fine and there is nothing wrong with that.
 

rav

Member
I'm not so sure about that:

Originally Posted by Anandtech

This in turn is why the 224GB/sec memory bandwidth number for the GTX 970 is technically correct and yet still not entirely useful as we move past the memory controllers, as it is not possible to actually get that much bandwidth at once on the read side. GTX 970 can read the 3.5GB segment at 196GB/sec (7GHz * 7 ports * 32-bits), or it can read the 512MB segment at 28GB/sec, but not both at once; it is a true XOR situation. Furthermore because the 512MB segment cannot be read at the same time as the 3.5GB segment, reading this segment blocks accessing the 3.5GB segment for that cycle, further reducing the effective memory bandwidth of the card. The larger the percentage of the time the crossbar is reading the 512MB segment, the lower the effective memory bandwidth from the 3.5GB segment.

If I understand it correctly: Any time the card is accessing the 0.5GB part it will completely halt access to the other 3.5GB, so it gimps the overall bandwidth significantly. I don't think it's clear cut that having the 0.5GB partition accessible is always better than having that data in system RAM instead. I'd guess it depends on the particular application/game.

I'd imagine having full speed access to 3.5GB + slow access from system RAM combined is faster than ONLY having access to the slow 0.5GB VRAM and nothing else in a given read cycle.

Can someone more tech-savvy on this comment on this?

So what it appears to me is they're saying the memory speed calculation is bogus. At least for read speed. It might be correct for writing speed.
You can't get 196GB/s and add the other 28GB/s and get 224GB/s. (The picture they show, has an 8th memory controller for that last 0.5GB. And you're right about having exclusive read access.) However, AnandTech goes on to say NVidia prioritizes what things they will need most often in the 3.5GB partition, and use some secret sauce in their driver to figure out what is most used and least used.

All in all it's giant mess.

I'd still prefer a driver option to report that there is only 3.5GB available that way when polled about how much a developer can use, they'll only see the fast 3.5GB.

(Sidenote: That's what Microsoft tried to claim with the Xbox One esram and the ddr3 ram.)
 

Joejoe123

Neo Member
It really depends how bad it gets, how nVidia continues to respond and the markets assessment and reaction to it.

Remember. IBM once had a relatively healthy hard drive division and one product with a high failure rate essentially put it under and they sold up to Hitachi. Just saying...

The good ole "Death Star" drives. To be fair to nvidia, I doubt this memory issue will rise to the level of wide spread drive failures.
 

TSM

Member
So what it appears to me is they're saying the memory speed calculation is bogus. At least for read speed. It might be correct for writing speed.
You can't get 196GB/s and add the other 28GB/s and get 224GB/s. (The picture they show, has an 8th memory controller for that last 0.5GB. And you're right about having exclusive read access.) However, AnandTech goes on to say NVidia prioritizes what things they will need most often in the 3.5GB partition, and use some secret sauce in their driver to figure out what is most used and least used.

All in all it's giant mess.

I'd still prefer a driver option to report that there is only 3.5GB available that way when polled about how much a developer can use, they'll only see the fast 3.5GB.

(Sidenote: That's what Microsoft tried to claim with the Xbox One edram and the ddr3 ram.)

The way I read what they said is that they use the .5GB as a third pool of memory that is faster then the main system memory, but much slower then the main 3.5GB pool of vram.
 

Bastables

Member
At 1440p

Metro LL = Same
COH2 = 290x
Bioshock Infinite = 970
Battlefield 4 = Same
Crysis 3 = Same
Warhead = 290x
Total War Rome = 970 (but almost same)
Thief = Same
Grid 2 = 970 (but almost same)

290x = 1 clear win
970 = 1 clear and 2 barely wins

Everything else is same between the cards. If you take into account the latest MFAA tech that Nvidia added to almost every DX10 and 11 game in the last driver update, you get a further 10-15% FPS boost in all games with the same IQ.



Yes, I didn't either. This however does not significantly change how tomorrow's games will perform on this card. There is a change yes, maybe 5-10%, but that isn't something which Nvidia cant work around with better drivers and performance.



Apples and oranges, not comparable at all.




I agree with this. Its pathetic of Nvidia to lie to its customers when they could have been upfront about this whole situation. Expecting compensation is completely fine and there is nothing wrong with that.

I like how one 290x win disappears in your final tabulation, you also miss the two bear wins at 1440p are reversed at 4k, again indicating that the 290x actual 4gigs does start making a difference. Price is still cheaper amd have not lied to me or reviewers so i'm going red this time round.
 

Rafterman

Banned
That I am aware of. Yes it is still 4x faster than system ram over PCI-Express.

But the article says the card can never access both the 0.5gb and the 3.5gb at the same time. So if in a given scenario the game engine needs to access assets that are located in both the 3.5gb part and the 0.5gb part, it will have to do one at a time.

My question is: Can the card access its 3.5gb part and system ram in the same read cycle? If yes, then can't there be a scenarios where it would be advantageous to have continuous access to the full bandwidth of the 3.5gb while occasionally dipping into system ram for lower priority stuff, instead of completely stopping any transfer from the 3.5gb every time it needs to make a call to the 0.5gb part?

I'm not a hardware expert and I haven't crunched the numbers to see how the math turns out in both cases. I'd like to see it though.

Of course it all falls apart if the answer to my bolded question is No.

The card would literally have to be using only the .5gb partition for seconds at a time before system ram would be more beneficial, and since ram cycles are done in nanoseconds there would never be a case where system memory was the better solution.

So why don't nvidia use 20GB/s VRAM on GPUs? Why is 3.5Gb of the VRAM specced to 150GB/s? Sounds like nvidia are wasting money? /sarcasm

Even if it's faster than system ram it's irrelevant since it's GDDR5, we expect the specified 224GB/s bandwidth.

And since when do we ever want to use system ram as VRAM? The point is it's only slightly quicker than system ram, that's the whole point of having GDDR5, for much much faster speeds...

Reading is fundamental.

The post i replied to said:

If I understand it correctly: Any time the card is accessing the 0.5GB part it will completely halt access to the other 3.5GB, so it gimps the overall bandwidth significantly. So I don't think it's clear cut that having the 0.5GB partition accessible is always better than reading it from system RAM instead. I'd guess it depends on the particular application/game.

And I replied that that it is clear cut that the .5gb partition is always better than system ram...that's it. No one said anything about wanting to use system ram, or that video card manufactures didn't need ram as fast as they use, or whatever else you pulled out of your ass to argue against.
 

LilJoka

Member
The card would literally have to be using only the .5gb partition for seconds at a time before system ram would be more beneficial, and since ram cycles are done in nanoseconds there would never be a case where system memory was the better solution.



Reading is fundamental.

The post i replied to said:



And I replied that that it is clear cut that the .5gb partition is always better than system ram...that's it. No one said anything about wanting to use system ram, or that video card manufactures didn't need ram as fast as they use, or whatever else you pulled out of your ass to argue against.

Sorry I missed that post!
 

LiquidMetal14

hide your water-based mammals
Called Gigabyte support and they were very helpful. They basically can't do anything yet as they have to wait from word from Nvidia but the gentlemen that I spoke to both were helpful and one of them even took my contact information as if there is a major update on this I will probably be receiving a call.

I kind of just want to get ahead of the game because I want the performance that was advertised and I know it might not seem like a lot but when you talk about somebody who plays at high resolutions and high refresh rates and then you have other stuff along the way like virtual reality headsets coming, then you would understand why I still want to continue to support these companies but what's right is right. I don't even mind sending the card back so long as I can get a better card because to me this is not really something I would want to live with. I'm trying not to sound hyperbolic but it is what it is and as a consumer and enthusiast builder I cannot be any nicer about this issue. I want all the performance that I paid for is all.
 
...in the meantime in Europe EVGA 970s have received a nice +40 euros to the price. Just today.

Meaning that a 970 now costs around 50 euros more than a 290x. I'm glad I'm waiting to buy a new card, because right now it's a really bad time to spend money. European prices are skyrocketing up after being much, much higher than US prices in the first place.

Close to 400 euros for a 970 is not anymore a good price. It's the equivalent of $432.
 

XBP

Member
I like how one 290x win disappears in your final tabulation, you also miss the two bear wins at 1440p are reversed at 4k, again indicating that the 290x actual 4gigs does start making a difference. Price is still cheaper amd have not lied to me or reviewers so i'm going red this time round.

Really? I missed that by mistake. Not like I didn't mention the win a couple of lines above. I wasn't comparing the performance at 4k. These cards are not meant to be used at 4k resolution. Unless someone has a 4K gsync monitor, those frame rates will result in terrible gameplay.
 
Called Gigabyte support and they were very helpful. They basically can't do anything yet as they have to wait from word from Nvidia but the gentlemen that I spoke to both were helpful and one of them even took my contact information as if there is a major update on this I will probably be receiving a call.

I kind of just want to get ahead of the game because I want the performance that was advertised and I know it might not seem like a lot but when you talk about somebody who plays at high resolutions and high refresh rates and then you have other stuff along the way like virtual reality headsets coming, then you would understand why I still want to continue to support these companies but what's right is right. I don't even mind sending the card back so long as I can get a better card because to me this is not really something I would want to live with. I'm trying not to sound hyperbolic but it is what it is and as a consumer and enthusiast builder I cannot be any nicer about this issue. I want all the performance that I paid for is all.

Did they know about it?
 

Bastables

Member
Really? I missed that by mistake. Not like I didn't mention the win a couple of lines above. I wasn't comparing the performance at 4k. These cards are not meant to be used at 4k resolution. Unless someone has a 4K gsync monitor, those frame rates will result in terrible gameplay.

So 30fps + is now unplayable and at 4k you need a nvidia compatible monitor . . .

Sure ok, but I'm still not getting a 3.5 gig card initially sold as a 4 gig card that early trades with a cheaper card with actually 4gigs of video memory, not just 4gigs pr.
 

Vamphuntr

Member
I'll contact ASUS support tomorrow. Just double checked online and the card is registered properly with its serial to my account. Hopefully they will be helpful.

I feel disappointed a lot. You can argue all you want the card is still fantabulous but it still not what they advertised. If it's a 4 GB card it should be able to use all 4 GB at full speed and not only 3.5 GB. It doesn't help that GPUZ also displays the wrong info. Feels really shady and more than an honest mistake.
 

rav

Member
The way I read what they said is that they use the .5GB as a third pool of memory that is faster then the main system memory, but much slower then the main 3.5GB pool of vram.

Yes, but if accessing that extra 'faster than main system memory' pool slows down the other 3.5GB, no thanks.
 

ruiner0

Neo Member
So at this point, The take away I am getting is the following. Nvidia lied about the cards being 4GB they are actually 3.5GB with a shitty over flow of 500MB. When it hits the 500MB of overflow, The game running becomes a slide show for a few seconds? And if you want a Nvidia card with true 4GB buy the GTX 980? If so these has made my purchase in the next 2 weeks a simple one.
 
Top Bottom