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

Page 33 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Status
Not open for further replies.

Spanners

Senior member
Mar 16, 2014
325
1
0
Give me a break. You'll tell me nobody from Nvidia noticed that in 5 months? 5 freakin months man.

There are 8,800 employees at Nvidia.

Honest mistake? Yeah right. Even if I was the greatest Nvidia fanboy I would just turn my back on them. I can tell you that my next GPUs purchase, unlike my previous, I will not consider Nvidia.

Hanlon's razor man, although it is hard to not be skeptical.
 

rgallant

Golden Member
Apr 14, 2007
1,361
11
81
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.
it's only accessible if the drivers and windows all work like they should on paper , that seems a bit like a dice roll for a issue [3.5 +.5] that was created to bin more defective chip and nothing more.

some peeps seem to have problems with the last 512mb
could be funny if win 10 does not play nice.
 
Last edited:

KaRLiToS

Golden Member
Jul 30, 2010
1,918
11
81
I'm sorry if I was trolling. PCper is a great website. And I truly believe they do awsome work.

I removed my negative comment about them.
 
Last edited:

Haserath

Senior member
Sep 12, 2010
793
1
81
it's only accessible if the drivers and windows all work like they should on paper , that seems a bit like a dice roll for a issue that has created to bin more defective chip and nothing more.

some peeps seem to have problems with the last 512mb
could be funny if win 10 does not play nice.

Seems like it'd be far better to not share between memory controllers and cut out the L2/Rops/Mem all together and save on the extra Ram.
 

Enigmoid

Platinum Member
Sep 27, 2012
2,907
31
91

Well, I deleted my post because I thought I was mistaken but it turns out I am not.

AdvWarfare_2560x1440_PLOT_0.png


AdvWarfare_2560x1440_STUT_0.png


Pcper is not normalizing their data. Relative Variance is far greater at high than ultra on the 980. However, this is not seen on the percentile frametine chart because there is no normalization to performance.

Essentially, for a constant distribution variance with scale with magnitude. Double the value of each point in the data set and you will double the absolute variance, however they are the same data set just one is scaled. Therefore the 980 with a 20-25% advantage over the 970 will enjoy 20-25% lower frame time variances.

Think of it this way. 980 vs. 970 on ultra, 980 is 25% ahead with 25 vs 20 fps. On high the 980 is still 25% higher, 50 vs. 40 fps. The gap between the two cards has widened in terms of absolute fps yet has remained constant in terms of relative performance.

The variance graph pcper gives us is only the absolute framerate variance, not the relative framerate variance.

(The higher variance on high vs. ultra settings is likely an effect of driver granularity and game engine execution (ie player triggers some animation in the background by crossing a line).

Likewise the bf4 tests are pretty useless.

980-BF4_3840x2160_PLOT.png


BF4_3840x2160_PLOT.png


-970 better than 980 at 150% scaling.
-1 spike at 970 @ 130% is meaningless, especially when the 140% spike is smaller

You can look at the rest of the charts but the few couple spikes throw the charts right off. Which is bad because they appear to be completely random.

The variance charts also do not normalize to performance.
 
Last edited:

TheRyuu

Diamond Member
Dec 3, 2005
5,479
14
81
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. o_O

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.

I believe the problem is that the driver is oblivious to the specifics of an engine's memory allocations. It has to do with how the frostbite engine allocates memory (en masse) so it can store things that aren't in use. The drivers are unaware of certain specifics that may help it in that case so the heuristics may not be able to as easily create a more optimal layout.
 

garagisti

Senior member
Aug 7, 2007
592
7
81
From AMD for those who may want "a full 4gb card"
https://forums.geforce.com/default/...tx-970-3-5gb-vram-issue/post/4440406/#4440406

Hey guys,

Just wanting to let you know that AMD is offering a deal to those who return their GTX 970 to get a "REAL 4GB Graphics card."


https://twitter.com/amd_roy/status/560462075193880576


"Anyone returning their GTX 970 and wanting a great deal on a Radeon with a full 4GB please let us know."

Return your GTX 970 and get a AMD R9 290/X 4GB GPU with the deal.

E-mail: roy.taylor@amd.com

Looking forward to it. Spending $330+ on a GPU that has issues with current games that UNACCEPTABLE! Not to mention nvidia totally lied until they got caught and couldn't hide anymore. Imagine what's going to happen couple of years down the road when more and more games are going to take advantage of memory exceeding 3.5GB. That's totally un-acceptable for a high end GPU.
 

Final8ty

Golden Member
Jun 13, 2007
1,172
13
81
Did some more testing using Shadow of Mordor, SLI enabled and disabled, sub-3500mb and over 3500mb VRAM consumption. Frametimes stay within normal variation/acceptable consistence when in single-card mode regardless of VRAM, but going to over 3500mb in SLI causes wild and rampant stutters/hitches with vastly fluctuating frametimes to match.

mhOevfQ.png
Long story short, I have to agree for single-card that while it is a big false advertisement and spec change it may not have a giant practical impact (at least from what I can see so far... unless you may want to go dual-card later for example), but in SLI it is a very real and major issue.
http://hardforum.com/showthread.php?t=1849838&page=21
 

3DVagabond

Lifer
Aug 10, 2009
11,951
204
106
HardwareCanucks used FCAT and measured frame times and included that in the result (Average FPS).

They tested 4 games and found very little difference. How do you explain that?

I agree that Nvidia should have been honest from the start and given the reviewers and listed that on the product page. I guess they thought nobody would notice. And they wouldnt either, until the german dude made the benchmark that measured bandwidth.

What I find revealing from HWC's testing is how the 290X just holds up better when the memory is stressed than all 3 of the nVidia cards period. Consider that HWC is far from AMD biased too. Maybe it's not just the 970 that's having issues accessing memory?
GTX-970-MEMORY-1.jpg

The 290X goes from being 2% faster to 17% faster than the 970

GTX-970-MEMORY-2.jpg

Goes from 3% to 10% faster

GTX-970-MEMORY-3.jpg

Goes from 11% to 30% faster

GTX-970-MEMORY-4.jpg

Goes from 10% to 24% faster

I every instance the performance percentage loss is less on the 290X than the three nVidia cards. And as far as the 5% not mattering, what's the big difference that makes the nVidia cards worth more money? Especially when you look at what has happened to GK110 as it's aged vs. Hawaii.
 

SimianR

Senior member
Mar 10, 2011
609
16
81
pcper review is very good, the only thing i dont like is that they go from 4k to 4k+ res
i would like it more if they do the same testing as the reddit guy at 1080p and then going up

https://www.reddit.com/r/buildapc/comments/2tu86z/discussion_i_benchmarked_gtx_970s_in_sli_at_1440p/

and as long as NVIDIA keep updating their driver it wont be a problem

You can't magically fix everything with a driver though. At some point there will be a situation where nearly all of that 4GB of VRAM will be necessary, and having 500mb of it that is basically inaccessible without a performance hit will have an impact.
 

3DVagabond

Lifer
Aug 10, 2009
11,951
204
106
... only 970 owners should really be concerned about this "issue".

Not true. This type of marketing is everyone's concern if we don't want to see it happen again. Maybe if there is a next time it will effect those who didn't buy 970's.
 

kagui

Member
Jun 1, 2013
78
0
0
You can't magically fix everything with a driver though. At some point there will be a situation where nearly all of that 4GB of VRAM will be necessary, and having 500mb of it that is basically inaccessible without a performance hit will have an impact.

Thats true
 

Abwx

Lifer
Apr 2, 2011
11,892
4,876
136
What I find revealing from HWC's testing is how the 290X just holds up better when the memory is stressed than all 3 of the nVidia cards period. Consider that HWC is far from AMD biased too. Maybe it's not just the 970 that's having issues accessing memory?

I every instance the performance percentage loss is less on the 290X than the three nVidia cards. And as far as the 5% not mattering, what's the big difference that makes the nVidia cards worth more money? Especially when you look at what has happened to GK110 as it's aged vs. Hawaii.

Too bad a 290 is missing on thoses charts, it s still the card with the best pef/price ratio 15 months after it was released, on thoses graphs it would had fared better than the 970 when memory stressed.
 

3DVagabond

Lifer
Aug 10, 2009
11,951
204
106
If nVidia was honest upfront about the memory config, relased it at 3.5GB and why the card had 3.5GB, that having and using 4GB was not ideal, I think people would be fine with that.

It still offers great performance, only with an asterisk now.

I'm not sure if I'm the only one, I hope not, but I was very confused at the 980 / 970 launch as to why the 970 mirrored the 980 ROP and memory specs.

From past launches, the x70 was always cut down.

Well, now we know the 970 was cut down, but that information was not able to leave nVidia until these past few days...

The Memory spec is not the only falsehood though. The ROP's and L2 cache are pure lies. No way to try and spin that as simply accessed differently.
 

3DVagabond

Lifer
Aug 10, 2009
11,951
204
106
I honestly don't consider an Nvidia Rep encouraging people to return the cards to the seller as rectifying the problem. Unless they're offering a $40 or $50 credit per card returned to offset the money the seller is losing having to sell it open box now, all they have really done is punish sellers for their own error. And I don't understand how anyone in his right mind can think this was an accident. I'm really supposed to believe that every engineer who worked on these cards didn't go read a single online review? That the people who wrote the drivers who knew this was a segmented memory architecture never saw a review? That's nonsense and I can't believe anyone would be so gullible.

This solution is also pretty sleazy. Rather than come out and offer refunds to customers they say, "If you can't get one let us know and we'll try and do something on the side." Blaming it on the marketing department is pretty sad too. Like Marketing made up the specs or something. They would have obviously used the specs they were given.
 

Carfax83

Diamond Member
Nov 1, 2010
6,841
1,536
136
What I find revealing from HWC's testing is how the 290X just holds up better when the memory is stressed than all 3 of the nVidia cards period. Consider that HWC is far from AMD biased too. Maybe it's not just the 970 that's having issues accessing memory?

It's not surprising to me, given that the 290x has a 512 bit bus and a lot more bandwidth. There's only so far that a 256 bit bus can get you, even with a massive L2 cache and third generation color compression.

I every instance the performance percentage loss is less on the 290X than the three nVidia cards. And as far as the 5% not mattering, what's the big difference that makes the nVidia cards worth more money? Especially when you look at what has happened to GK110 as it's aged vs. Hawaii.

The only reasons why AMD isn't an option for me right now, are the much higher power consumption and the high CPU overhead for DX11.

If it wasn't for these two issues, I'd sell my GTX 970s (or try to get a refund) and buy a single R9 290x to tide me over until GM200 and R9 300 parts start shipping..

I'm done with multi GPU. I want to move over to a single card solution desperately, but currently nothing is fast enough for 1440p max settings at the moment.

I hope AMD's 390x is really good, because that's the only way NVidia is going to be checked on their pricing.
 

Spanners

Senior member
Mar 16, 2014
325
1
0
Ever wondered why a 2GB card can run something like Dying Light for example with the same settings as a 4GB card that utilize over 3GB and no stuttering from both?

I hadn't wondered that and I don't even agree that it's the case. Again though it doesn't address anything I asked in either of my posts, I'll drop it now.

http--www.gamegpu.ru-images-stories-Test_GPU-Action-Dying_Light-test-dl__vram.jpg


Seems the 980 and 290x are trading blows when it comes to VRAM usage depending on the resolution.
 

Carfax83

Diamond Member
Nov 1, 2010
6,841
1,536
136
Ever wondered why a 2GB card can run something like Dying Light for example with the same settings as a 4GB card that utilize over 3GB and no stuttering from both?

That's really dependent on the engine, and how efficient it is with resources. On PCs, all of the graphics data (including textures) is stored in both VRAM and system memory.

It's the shuffling and swapping of data between these two memory pools that's responsible for a lot of the stuttering. But some engines are better at mitigating this effect than others.

Examples of engines with exceptional resource handling are CryEngine and Frostbite 3. I first played Crysis 3 on my overclocked GTX 580s with 1.5 GB of VRAM each, and my average frame rate was about 35 at max settings with 1440p. Frame rate was laggy at times, but there was no stuttering.

On the flipside, trying to play AC Unity with less than 3GB with ultra textures will lead to tons of stuttering, because the engine just has horrible resource management.
 

Haserath

Senior member
Sep 12, 2010
793
1
81
The Memory spec is not the only falsehood though. The ROP's and L2 cache are pure lies. No way to try and spin that as simply accessed differently.

Not sure why people are making a fuss about the rops/L2. The only thing that will cause an issue is the memory segmentation.
 

Pantalaimon

Senior member
Feb 6, 2006
341
40
91
Not true. This type of marketing is everyone's concern if we don't want to see it happen again. Maybe if there is a next time it will effect those who didn't buy 970's.

Yup, it's not only current GTX970 owners, but also potential customers who are considering buying a GTX970 need to know this.

While I agree that the performance of the card is still there, it's longevity outlook has dimmed, and those who bought the card might have decided otherwise, because some customers don't buy only for current performance but also for the longevity potential.
 
Status
Not open for further replies.