Support NeoGAF

[ashecitism]

update 5:

so, computerbase did some testing.

They showed that games will stutter heavily or freeze if you force them to use the slow memory.
1080p gamers are most likely not affected (yet). SLI and/or multi-monitor users will suffer from problems much most often.

Case in point: http://www.computerbase.de/videos/2015-01/gtx-970-vs-gtx-980-2-monitore-far-cry-4/
FarCry 4 before and after (51s) connecting a second monitor. Second monitor is showing aero desktop.
https://www.youtube.com/watch?v=MTYd9_fe4iI
Talos principle long duration freezes

update 4:

PSA: AMD drops price of 290X to $299 ($279 with rebate from Newegg) and offers discounts to 970 owners who want to return their card.

update 3:

Retailers and AIC partners are taking all the heat right now. NV really needs to come up with a plan.

Perfectly Functional GTX 970 Cards Being Returned Over Memory Controversy

update 2:

http://www.pcworld.com/article/2876...r-update-for-memory-performance-concerns.html

Important update 1/29/15: The Nvidia employee who said the company was looking into a GTX 970 driver that would "tune what's allocated where to further improve performance" has updated his post to remove the claim after it was covered by several publications, including PCWorld, PC Gamer, and PC Perspective. As that changes the entire thrust of this article, its headline has been updated to reflect that. We still stand by our recommendation of the GTX 970 and you can read a summary of the memory allocation firestorm here.

update:

PCPer has the first update with info from nvidia: http://www.pcper.com/reviews/Graphics-Cards/NVIDIA-Discloses-Full-Memory-Structure-and-Limitations-GTX-970

Apparently Nvidia "accidentally" mislabeled the 970 as having the same number of ROPs and L2 cache as the 980 in all the review material they sent out to tech sites.

They confirm the last 500MB is 1/7th the speed of the first 3.5GB also. Expect detailed benchmarks this week sometime.

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

--

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

The GeForce GTX 970 is equipped with 4GB of dedicated graphics memory. However the 970 has a different configuration of SMs than the 980, and fewer crossbar resources to the memory system. To optimally manage memory traffic in this configuration, we segment graphics memory into a 3.5GB section and a 0.5GB section. The GPU has higher priority access to the 3.5GB section. When a game needs less than 3.5GB of video memory per draw command then it will only access the first partition, and 3rd party applications that measure memory usage will report 3.5GB of memory in use on GTX 970, but may report more for GTX 980 if there is more memory used by other commands. When a game requires more than 3.5GB of memory then we use both segments.

We understand there have been some questions about how the GTX 970 will perform when it accesses the 0.5GB memory segment. The best way to test that is to look at game performance. Compare a GTX 980 to a 970 on a game that uses less than 3.5GB. Then turn up the settings so the game needs more than 3.5GB and compare 980 and 970 performance again.

Here’s an example of some performance data:

On GTX 980, Shadows of Mordor drops about 24% on GTX 980 and 25% on GTX 970, a 1% difference. On Battlefield 4, the drop is 47% on GTX 980 and 50% on GTX 970, a 3% difference. On CoD: AW, the drop is 41% on GTX 980 and 44% on GTX 970, a 3% difference. As you can see, there is very little change in the performance of the GTX 970 relative to GTX 980 on these games when it is using the 0.5GB segment.

edit:

Maybe you want to add techreport article to the op since there is a interesting part at the end about rop usage: http://techreport.com/news/27721/nvidia-admits-explains-geforce-gtx-970-memory-allocation-issue

I guess it's not as big of a problem as it looked like but they shouldn't have kept customer ignorant of it.

 

[The Llama]

So it IS a hardware issue. Oh boy. Not a good look.

 

[Dictator93]

Cross posting my comment:
Their testing methodology is not necessarily good at showing what occurs. Better would be
1. to put the game with low textures at something like 4K. Thereby not filling the VRAM to even 3.5GB, then measure performance.

2. Then putting the textures to ULTRA, thereby filling to beyond 3.5GB at 4K, and measuring the performance. This would show the actual cost of filling that last 512 MB.

At least that would be consistent instead of also measuring the performance drop associated with increasing resolution.

 

[DreadPirateRoberts]

Cross posting my comment:
Their testing methodology is not necessarily good at showing what occurs. Better would be
1. to put the game with low textures at something like 4K. Thereby not filling the VRAM to even 3.5GB, then measure performance.

2. Then putting the textures to ULTRA, thereby filling to beyond 3.5GB at 4K, and measuring the performance. This would show the actual cost of filling that last 512 MB.

At least that would be consistent instead of also measuring the performance drop associated with increasing resolution.

They're pretty clearly just trying to downplay the issue since the only fix would be a recall.

 

[JimmyRustler]

So wait, the 980 has a similar problem as well or what?

 

[gregor7777]

Yikes

 

[The Llama]

So wait, the 980 has a similar problem as well or what?

No, it doesn't. Basically the 970 uses a cut-down version of the same chip as the 980. Part of what they cut out has to do with memory management, and that's what causes the poor bandwidth.

 

[diablos991]

So wait, the 980 has a similar problem as well or what?

The 989 doesn't have the problem of being a 3.5gb card sold as a 4gb card.

 

[Stevey]

3% isn't a lot.

 

[CassidyIzABeast]

AMD 300 series cards will feature "FULL 4GB HBM" on the box

 

[BennyBlanco]

God damn. I'm so glad I didn't buy one.

 

[IAmRandom31]

So wait, the 980 has a similar problem as well or what?

Doesnt sound like it, since the 970 has a different setup then the 980. Sounds like the 980 is altogether while the 970 is split into 1 3.5GB section with its own memory bandwidth pipeline and then another .5GB section with its own pipe. When a game doesn't need 4GB of RAM the whole thing runs off the 3.5GB pipe. If it needs 4GB it will use both, and I gues s that is the cause for some issues in some games.

 

[Krakn3Dfx]

So what I'm reading is that Nvidia basically kneecapped memory usage above 3.5GB on a card with 4GB of VRAM.

Nvidia gonna Nvidia.

Still happy with my GTX970, but as someone who normally buys a new card Day One, this is going to make me more cautious going forward for sure.

 

[Terbinator]

So it IS a hardware issue. Oh boy. Not a good look.

Define issue, I guess. The chip/RAM sub-system is behaving just how Nvidia planned/expected it to, given it's cut-down nature relative to a 980.

 

[Serandur]

We need frametime graphs, not average FPS and not from Nvidia themselves. That the 970s try to avoid utilizing the last 0.5GBs can alone causes issues in some situations.

 

[JimmyRustler]

No, it doesn't. Basically the 970 uses a cut-down version of the same chip as the 980. Part of what they cut out has to do with memory management, and that's what causes the poor bandwidth.

Well, the graph in the OP shows similar FPS drops with both 970 and the 980. so whats the biggy? I don't get it.

 

[Seanspeed]

Cross posting my comment:
Their testing methodology is not necessarily good at showing what occurs. Better would be
1. to put the game with low textures at something like 4K. Thereby not filling the VRAM to even 3.5GB, then measure performance.

2. Then putting the textures to ULTRA, thereby filling to beyond 3.5GB at 4K, and measuring the performance. This would show the actual cost of filling that last 512 MB.

At least that would be consistent instead of also measuring the performance drop associated with increasing resolution.

Well I think the point of what they're showing is that the difference in performance between this and a 980 when it goes over 3.5GB is pretty similar. So unless this is an issue that both the 980 and 970 have, then its performing as expected.

Obviously that is only one test, and it would be interesting to see if 980's are actually affected by some issue as well, but what they're saying makes sense.

 

[The Llama]

Define issue, I guess. The chip/RAM sub-system is behaving just how Nvidia planned/expected it to, given it's cut-down nature relative to a 980.

How is it not an issue that 12.5% of the card's VRAM performs so poorly??

 

[Seanspeed]

How is it not an issue that 12.5% of the card's VRAM performs so poorly??

Did you not read the OP? :/

 

[Terbinator]

How is it not an issue that 12.5% of the card's VRAM performs so poorly??

A design decision/trade-off causing only a 3% drop in performance compared to a card that costs £100 more?

You tell me where the issue is.

 

[Elios83]

So it IS a hardware issue. Oh boy. Not a good look.

I wouldn't call it a issue.
It's a cheaper card with more limited resources, there's a compromise to use more than 3.5GB of memory and it performs worse. No shock here, I understand that experienced users would like to be informed even of low level technical details though and nVidia has not been transparent on the matter.

 

[The Llama]

Did you not read the OP? :/

Yes, I did read about how nVidia split off the problematic 0.5GB from the rest of the VRAM because they know it's slower.

A design decision/trade-off causing only a 3% drop in performance compared to a card that costs £100 more?

You tell me where the issue is.

Frankly, I don't trust nVidia's benchmarks. Of course they're going to try to make the problem look as small as possible! Plus, it's frametimes (stuttering) that's the problem, not the overall framerate.

 

[JoeyJungle]

So the bandwidth of the last .5GB is 3% slower? Not really an issue if I'm understanding this.

 

[Serandur]

A design decision/trade-off causing only a 3% drop in performance compared to a card that costs £100 more?

You tell me where the issue is.

An average FPS number tells you nothing about the low points, stuttering, and frametime inconsistencies. It's just an average, not representative of the actual performance and issues.

 

[Naked Snake]

#ramgate

 

[Kayant]

I will wait for independent benchmarks thank you but I guess that confirms that theory from GURU3D iirc.

 

[orava]

So it was a non-issue after all. What a surprise.

 

[Thaedolus]

Seems like a negligible "issue." Had I a 970, I wouldn't really care.

Spoiler

*Hugs 980*

 

[Objectively Bad Opinion]

Cross-posting:

Why even bother testing? We know what the performance for that last .5GB is, and it's super shit, that has been tested extensively already. Whatever they do in their drivers to juggle VRAM around to mitigate that is beside the point. They're just skirting the issue with no guarantee that they can avoid it for all cases.

I just don't understand why they had to lie when this card would have been perfectly fine sold as a 3.5GB GPU.

I'd also like to add that I'm wondering how NVidia will report this card to developers who will have to manage the memory themselves using DX12. I wonder if any of them will give a damn about this hamstrung SKU.

 

[R_Deckard]

Dat damage control,

Spoiler

not really a massive issue in reality

 

[Qassim]

So.. a paired down GPU performs slightly worse than a higher spec GPU? Would it be fair to say that people are just worried and concerned because a limitation is brought to their attention rather than an actual legitimate issue with the GPU?

Or am I being unfair and misrepresenting the issue? I haven't really kept up with this.

 

[Seanspeed]

Yes, I did read about how nVidia split off the problematic 0.5GB from the rest of the VRAM because they know it's slower.

So you didn't read the OP, got it.

 

[dalin80]

So the bandwidth of the last .5GB is 3% slower? Not really an issue if I'm understanding this.

Nope, the bandwidth on the last 0.5 is about 90% slower but as they have used an average test it has been nicely disguised as a minor issue. Nvidia trying to pull a con with percentages.

 

[The Llama]

So you didn't read the OP, got it.

Ah yes, thank you for telling me what I have and have not done, I truly appreciate it.

 

[Zafir]

That seems to agree with my experience with Shadow of Mordor on my 970. It mostly kept to 3.5GB, but sometimes it went up to 4GB during certain situations. One being when it was rainy/stormy. I only really had stuttering on loading in to an area, but I kind of expect that, because it has to load in the assets.

 

[Alasfree]

Maybe you want to add techreport article to the op since there is a interesting part at the end about rop usage: http://techreport.com/news/27721/nvidia-admits-explains-geforce-gtx-970-memory-allocation-issue

I guess it's not as big of a problem as it looked like but they shouldn't have kept customer ignorant of it.

 

[eot]

Well, this is a bit shitty. I'm still tempted to buy one though.

 

[Zane]

Nope, the bandwidth on the last 0.5 is about 90% slower but as they have used an average test it has been nicely disguised as a minor issue. Nvidia trying to pull a con with percentages.

It doesn't matter if the bandwidth on that last bit of RAM is whatever percentage slower if it only results in an overall performance downgrade of a max of 7%. Video cards don't live or die by their memory bandwidth on the last 512MB of RAM.

 

[ashecitism]

Maybe you want to add techreport article to the op since there is a interesting part at the end about rop usage: http://techreport.com/news/27721/nvidia-admits-explains-geforce-gtx-970-memory-allocation-issue

I guess it's not as big of a problem as it looked like but they shouldn't have kept customer ignorant of it.

put it in the OP

 

[Dictator93]

In those Nvidia presented tests, is the GTX 970 reading more than 3.5GB of RAM in the ">3.5GB setting = 3456x1944" test?

Because if not, that means it is NOT accessing that extra amount even though the 980 may. Once again, I think there needs to be more transparent testing regarding reported VRAM amounts and frametimes. Very specically in a game that does not just fill VRAM given to it, but rather, one with scalar VRAM costs based upon resolution or texture settings.

 

[blastprocessor]

I'm not seeing an issue here given the stats.

 

[teiresias]

It doesn't matter if the bandwidth on that last bit of RAM is whatever percentage slower if it only results in an overall performance downgrade of a max of 7%

An AVERAGE performance delta of single digit percentage is misleading because the instantaneous performance degradation that leads to that lower average actually results in sporadic frametimes and hence annoying tearing and jutter - as people with 970s are reporting to be the case.

Nvidia is reporting the average delta so it sounds better in a PR statement.

 

[samn]

They did something like this with my GTX 660 2GB. 0.5GB of it is shitty slow DDR3 RAM or something.

 

[Dictator93]

I'm not seeing an issue here given the stats.

An AVERAGE performance delta of single digit percentage is misleading because the instantaneous performance degradation that leads to that lower average actually results in sporadic frametimes and hence annoying tearing and jutter - as people with 970s are reporting to be the case.

Nvidia is reporting the average delta so it sounds better in a PR statement.

Yes, an average is not good at showing what could be happening. FCAT or ms readings would be better.

 

[dalin80]

It doesn't matter if the bandwidth on that last bit of RAM is whatever percentage slower if it only results in an overall performance downgrade of a max of 7%. Video cards don't live or die by their memory bandwidth on the last 512MB of RAM.

Because an average doesn't take into account stuttering.

60 60 60 60 0 60 60 0 60 60 60 60 60 60 60 0 60 60 60 60 60

Total average is still high but will be awful to play. On average during the course of you life you will not be on fire in fact probably closer to over 99% of your life, so why bother about being on fire for less then 1%! Fun with averages.

 

[maniac-kun]

So they addressed the RAM in 32 bit. What kind of noobs work there?

 

[Reallink]

So you didn't read the OP, got it.

I'd caution against taking the claims of a corporation at face value who has an interest in maintaining sales and avoiding potential threats of lawsuits/recalls. I don't know what the truth of the situation will turn out to be, but any company will first try to talk their way around an issue or put out tilted explanations and evidence to the contrary. They may very well be truthful and on the up and up, but you should really wait for third party verification and explanations/tests from third party experts before you come out with the "Booya, told ya! Read the OP MORAN!" responses.

 

[Henrar]

It's not the first time they pull something like this. I think 460s used similar memory configuration in 768MB variants.

Also, ITT: lot's of 970 owners defending their purchase and believing corporation that the hardware issue in not an issue, but a feauture.

 

[espher]

Good way to massage numbers to downplay an issue. Sheesh. Hot garbage.

 

[Dictator93]

On average during the course of you life you will not be on fire in fact probably closer to over 99% of your life, so why bother about being on fire for less then 1%! Fun with averages.

LöL