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

[wanderer27]

So I've read that the SM is really going to be the bottleneck here.

Also, even the 500MB section of Memory is 4x faster than System Bus speed.

I'm not sure I see a big problem here, it's not like the FPS etc. everyone's seen have all of a sudden regressed from what they were shown to be before this came to light.



If this 500MB section is really a problem, a test I'd like to see is one just testing over 3.5MB at stock Memory speed compared to (only) Memory speed overclocked.

Would probably need some kind of loop test demo so the results could be legitimately compared.




.

 

[Creig]

It's not a flaw. A flaw would signify something that is operating negatively outside design specifications. This card is operating as designed. The flaw was with the initial marketing.

"Initial" marketing would have been four months ago. The 970 was advertised as having the exact same memory subsystem as the 980 with just fewer SMMs and a lower clock speed. Now we find out that is also missing 8 ROPs and has less L2 cache. There's no way that every single person employed at Nvidia who read the reviews somehow 'missed' seeing that the specs being posted were incorrect.

Call it what you like, but I feel it is design flaw. The last 500MB of memory on the 970 cannot be read from all 8 memory controllers at once, unlike the 980. And this results in a measurable performance loss between 3.5GB and 4.0GB of memory usage.

There is still not a better card on the market for the price and you know that.

The 290/290X is a better card for the price. From the tests, it has no problem at all using its entire memory pool at full speed. And this 970 memory issue will only get worse as time goes on and more games start to push 3.5GB to 4.0GB.

I suspect most of the faux outrage is coming from people who want a freebie and fanboys. There probably is only a very tiny subset of players that are truly impacted by the cards design limitation. None of this changes my recommendation to friends and family to buy this card over competitor products.

Seems to me a lot of people don't like being lied to about purchases they made that cost several hundred dollars. Especially if it meant they could have purchased another product instead, had they known the truth.

 

[Sunaiac]

" Just hoping my AX850 can power the rig."

850W are largely enough for 2 290X, if rails are well organized.
The power needs of 290s are well overblown

(But of course, don't try to OC them to 1250 each with a CPU OC to 200W of power usage too)

 

[SteveGrabowski]

What's my suggested recourse then? Just eat close to $200 in paypal/shipping/fleabay/taxes to get rid of my 970 SLI setup? I've forwarded that conversation to the Nvidia rep on the geforce board to see if he's able to help me. If/when I'm able to get rid of these 970's, I'll go ahead with a 290x crossfire setup which is very well depreciated at this point (at least vs. the wildly overpriced 980) to get me through to the 390x launch. Just hoping my AX850 can power the rig.

In a perfect world EVGA would refund me I guess. Though I guess in a perfect world my card would have been 4GB.

 

[Cloudfire777]

Saying it is just a 5% performance loss is oversimplifying things. In the earlier links in the threads, minimum frame rates were plummeting in some games. Frame times are worse. But all of that only happens at some settings that use a lot of vram, and may not be a big deal to many gamers today. But it may be a problem down the road as games that are more and more vram heavy come to market. And it seems that drivers may have a big impact on how this card manages the vram on this card, and that makes people concerned as Kepler drivers seem to have taken a backseat now that Maxwell is on the market. What happens if Maxwell drivers take a backseat next year?

It isn't all doom and gloom as the card still has a number of things going for it and can still make sense for some people. And the problems may always be a setting or two away from being eliminated. That's why none of the performance problems are the worst part, in my opinion. What is to me is the fact that Nvidia knew how this card was being reviewed and marketed. I think it is fairly deceptive on their part not to be up front with this card's configuration. At least let people make an informed decision.

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.

 

[sontin]

The 290/290X is a better card for the price. From the tests, it has no problem at all using its entire memory pool at full speed. And this 970 memory issue will only get worse as time goes on and more games start to push 3.5GB to 4.0GB.

[/QUOTE]

Which Tests?
Here is one from PCGH.de: http://www.pcgameshardware.de/Geforce-GTX-970-Grafikkarte-259503/Specials/zu-wenig-VRAM-1149056/

The 290X is much more worse than the GTX970 in Watch Dogs with 1080p and 8xMSAA.

PCLabs.pl testet AC:Unity and Mordor: http://pclab.pl/art61614-2.html

The 290X is only as good as the GTX970.

Saying the 290/290X is "a better card for the price" is just wrong. They have enough problems which can result in the same behaviour.

 

[SlowSpyder]

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.

http://www.guru3d.com/news_story/middle_earth_shadow_of_mordor_geforce_gtx_970_vram_stress_test.html

I can't explain why Hardware Canucks doesn't see what others are seeing. If this was a non-issue no one would have noticed anything going on, but they did and are in testing. Some people here are saying they can replicate problems, it seems especially with fast movement of the mouse. I guess it just depends on who's doing the testing. Had this issue not come to light now, it very well may have in the near future, seeing how hungry for vram many new games are.

 

[Spanners]

From the same paper:

Thanks, I did read it

I don't think that proves what you are claiming though. Again if Nvidia had achieved avg. 12.5%+ increases in VRAM usage efficiency I think we would have heard about it. I haven't seen any real-world game testing where the 980 acts like it has 4.5GB vs the R9 290X either.

Even if every bit of colour data stored in VRAM is compressed, which I don't think it is, how much is taken up with non-colour data(geometry data, driver data, etc)?

I've really only seen you making this claim and Google comes up dry also.

 

[skipsneeky2]

I wonder if from a pure performance and high fps point,is the 970 still a viable option?I often times drop settings if my minimums are hitting the 50's,i aim for 60s if i can.I don't do this 45fps min/65avg stuff these reviewers do.

I know future games could be impacted by 3.5gb but considering the insanely high resolutions and settings that is takes to hit that much memory right now,for 1080p could it offer enough still for a single gpu user?

 

[wand3r3r]

What's my suggested recourse then? Just eat close to $200 in paypal/shipping/fleabay/taxes to get rid of my 970 SLI setup? I've forwarded that conversation to the Nvidia rep on the geforce board to see if he's able to help me. If/when I'm able to get rid of these 970's, I'll go ahead with a 290x crossfire setup which is very well depreciated at this point (at least vs. the wildly overpriced 980) to get me through to the 390x launch. Just hoping my AX850 can power the rig.

If you are going to try get rid of them I'd go through NVidia so you don't lose money on the falsely advertised cards. I think they claimed that they will refund you, or help it happen.

I'm running 290x (reference) crossfire on the hx850 easily. It's around 700w while I've monitored it in BF4.

 

[Erenhardt]

Which Tests?
Here is one from PCGH.de: http://www.pcgameshardware.de/Geforce-GTX-970-Grafikkarte-259503/Specials/zu-wenig-VRAM-1149056/

The 290X is much more worse than the GTX970 in Watch Dogs with 1080p and 8xMSAA.

PCLabs.pl testet AC:Unity and Mordor: http://pclab.pl/art61614-2.html

The 290X is only as good as the GTX970.

Saying the 290/290X is "a better card for the price" is just wrong. They have enough problems which can result in the same behaviour.[/QUOTE]

If you pass a tip to PClab, to install amd drivers when they switch to testing amd gpus, I guess their result would change quite a bit. If you actually look anywhere else, you will see hawaii is faster.

Nice try with the smoke-screen thou...

The deal is, nvidia laid to their to their most loyal fans.
Now they want you to believe they can keep the performance (stuttering) in check via special (gtx970 exclusive) driver - that is after their kepler support have dropped from the roof (point in case: 660ti was a competitior to 7950. Look at graph above - miles behind 7870 <nosoap>)

 

[ShintaiDK]

Thanks, I did read it

I don't think that proves what you are claiming though. Again if Nvidia had achieved avg. 12.5%+ increases in VRAM usage efficiency I think we would have heard about it. I haven't seen any real-world game testing where the 980 acts like it has 4.5GB vs the R9 290X either.

Even if every bit of colour data stored in VRAM is compressed, which I don't think it is, how much is taken up with non-colour data(geometry data, driver data, etc)?

I've really only seen you making this claim and Google comes up dry also.

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?

 

[wand3r3r]

Which Tests?
Here is one from PCGH.de: http://www.pcgameshardware.de/Geforce-GTX-970-Grafikkarte-259503/Specials/zu-wenig-VRAM-1149056/

The 290X is much more worse than the GTX970 in Watch Dogs with 1080p and 8xMSAA.

PCLabs.pl testet AC:Unity and Mordor: http://pclab.pl/art61614-2.html

The 290X is only as good as the GTX970.

Saying the 290/290X is "a better card for the price" is just wrong. They have enough problems which can result in the same behaviour.

False, except maybe at 1080p and only when looking at avg fps without the frametimes to see spikes as the VRAM fills.

Maybe at lower resolutions but not at 4k or in crossfire where XDMA shines. http://www.techpowerup.com/reviews/Palit/GeForce_GTX_960_Super_JetStream/29.html

That 512MB of slow VRAM which cannot be accessed simultaneously with the rest of the 3.5GB has been demonstrated to have frametime spikes, and the fact that NVidia is offering a refund for their deceptive specs demonstrates the severity of the problem. We're past the stage of denial that the 970 has memory problems, it's just trying to determine the extent of the problem.

it can read the 512MB segment at 28GB/sec, but it cannot read from both at once; it is a true XOR situation

Basically the 970 is one of those GPUs which has 3.5 GB of memory with 512 MB of memory at nearly the slow speed of system RAM. Similar to a budget hybrid laptop (like TurboCache). I'd personally want the remaining 512 MB disabled so it doesn't hobble all of the VRAM as it requires the reduced speed across everything to use it.

If you happen to use less than the limit at low resolution with reduced settings then sure it's still a good GPU (still a little overpriced relative to the 290/x). The extent of 4k/SLI problems is still being investigated. As games require more VRAM it's basically a ticking time bomb until all new games would start to use it (most new ones already can even at 1080p).

Future proofing will not be good on the GTX 970.

 

[SlowSpyder]

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 don't understand where you're coming from here. There is a reason the GTX970 / 980 don't come with 2GB of vram. The GTX970 has less highspeed vram available to it than the GTX980. Since both use the same architecture and technologies, compression should be the same on both. But one has 3.5GB of highspeed vram available, one has the full 4GB available. So given the same compression, there is still less on the GTX970. How much difference that makes in real world gaming can be debated, but no matter how you look at it, there is less full speed vram on the GTX970.

 

[Magic Carpet]

I don't understand where you're coming from here. There is a reason the GTX970 / 980 don't come with 2GB of vram. The GTX970 has less highspeed vram available to it than the GTX980. Since both use the same architecture and technologies, compression should be the same on both. But one has 3.5GB of highspeed vram available, one has the full 4GB available. So given the same compression, there is still less on the GTX970. How much difference that makes in real world gaming can be debated, but no matter how you look at it, there is less full speed vram on the GTX970.

Actually, I think Shintai has a good point. Obviously, this needs to be further researched.

The perfect candidate, is however, AMD 285 2GB versus 270X 2GB. Both have the same amount of VRAM, but 285, if I am not mistaken, has even better compression than Maxwell. AT could do a nice article on this, imo.

 

[xthetenth]

But you know, too many people that feel they feel entitled to everything and milk this all they can to get an upgrade.

Yeah, I feel entitled to the card I bought or failing that the card I should have bought and quite possibly would have bought if I'd known.

I think that's pretty fair. Why do you think nvidia is entitled to the profits they made selling one card and shipping a worse one?

 

[caswow]

Which Tests?
Here is one from PCGH.de: http://www.pcgameshardware.de/Geforce-GTX-970-Grafikkarte-259503/Specials/zu-wenig-VRAM-1149056/

The 290X is much more worse than the GTX970 in Watch Dogs with 1080p and 8xMSAA.

PCLabs.pl testet AC:Unity and Mordor: http://pclab.pl/art61614-2.html

The 290X is only as good as the GTX970.

Saying the 290/290X is "a better card for the price" is just wrong. They have enough problems which can result in the same behaviour.

yea gamecraps titel in this test is the best choice D:

 

[tential]

I wonder if from a pure performance and high fps point,is the 970 still a viable option?I often times drop settings if my minimums are hitting the 50's,i aim for 60s if i can.I don't do this 45fps min/65avg stuff these reviewers do.

I know future games could be impacted by 3.5gb but considering the insanely high resolutions and settings that is takes to hit that much memory right now,for 1080p could it offer enough still for a single gpu user?

Then go for the GTX 970.

From what you've described, I really doubt this type of stuff is remotely relevant to you.

 

[amenx]

(as I mentioned elsewhere).. Ironic as it may sound, if the 970 from the beginning was presented as a 3.5gb card with 512mb L3 cache, and the tech logic behind it explained in reviews, it may have come off as more impressive sounding given the great benchmarks behind it. It may have been praised by the media as an ingenious feat.

 

[wand3r3r]

Then go for the GTX 970.

From what you've described, I really doubt this type of stuff is remotely relevant to you.

Even 1080p with high settings is hitting that area. DSR would for sure. Future games will easily use 4 GB unless you drop settings. It's about the least future proof card of the top contenders. I'd personally take a 3GB 780 ti over a 970 which slows down the whole VRAM to access that last 512 MB, but if I was buying today it'd be a 290/x without question.

 

[wand3r3r]

(as I mentioned elsewhere).. Ironic as it may sound, if the 970 from the beginning was presented as a 3.5gb card with 512mb L3 cache, and the tech logic behind it explained in reviews, it may have come off as more impressive sounding given the great benchmarks behind it. It may have been praised by the media as an ingenious feat.

Maybe, but unfortunately the current implementation is terrible!

It slows down ALL the VRAM when you are utilizing that slow section.

If it weren't for that quite questionable implementation then I agree it would only be a bonus as long as there wouldn't be adverse side affects on the frametimes like there are now.

The only solution (strictly imo) is to disable it entirely since it has adverse affects when used.

 

[Pneumothorax]

After going through 3 newegg reps, I got one who battled with her supervisor and got me store credit on my 970's. 290x CF here I come!

 

[n0x1ous]

After going through 3 newegg reps, I got one who battled with her supervisor and got me store credit on my 970's. 290x CF here I come!

Nice! Apparently Nvidia's Social Media guy has joined the forum "eatgamer" and has offered to help anyone with refund issues.

http://forums.anandtech.com/showpost.php?p=37120512&postcount=280

 

[ocre]

http://www.guru3d.com/news_story/middle_earth_shadow_of_mordor_geforce_gtx_970_vram_stress_test.html

I can't explain why Hardware Canucks doesn't see what others are seeing. If this was a non-issue no one would have noticed anything going on, but they did and are in testing. Some people here are saying they can replicate problems, it seems especially with fast movement of the mouse. I guess it just depends on who's doing the testing. Had this issue not come to light now, it very well may have in the near future, seeing how hungry for vram many new games are.

The difference is these random people are pushing their 970 up and passed unplayable settings.

The difference is HW Canucks is comparing their gtx970 vs a gtx980, this is the only way to truly measure the effect of nvidia's disabling..
You have to have a point of reference or it is useless. Seeing as the 980 is an uncut, full flat gm204........it's the one true way to see the penalty.

Every other example of proof is actually flaw.. The only way to measure the effect is to compare the penalty against a gtx980 with before and after 3.5gb settings.

I still am not happy about the misinformation and am in talks with nvidia.

But anyone who wants to see the real effect of their decision to cut the cache and segment the ram, they must look at it relative to the gtx980.

On a side note, on the nvidia forum mod Pete says nvidia is working on changing the driver and working on utilizing the ram in the best possible ways. He said that there is a lot of things that load into VRAM other than game textures. The drivers, OS, and apps you have running take ram and they are working on using the ram to allow non critical stuff differently, I guess they are gonna stuff more stuff in the extra 500mb and let games have access to the 3.5gb. Actually, I have no idea why they were not already doing that. Windows and apps are always using up VRAM on any card. 300-500mb is not all that common, your OS can eat up that much VRAM no problem. Apparently these things were occupying the faster ram. If nvidia driver team can stuff this stuff in the slow ram, the gaming performance for the 970 would be much closer to the 980 even in high VRAM scenarios. See, the 980 also has to deal with windows and driver stuff occupying its 4gb as well.

There seems to be some improvements that can be made. We will see if it helps but it really appears like it should.

I still am waiting for nvidia to respond to me though, so its not like I fine with the situation. Just trying to keep it real.

As it should be

 

[tential]

Even 1080p with high settings is hitting that area. DSR would for sure. Future games will easily use 4 GB unless you drop settings. It's about the least future proof card of the top contenders. I'd personally take a 3GB 780 ti over a 970 which slows down the whole VRAM to access that last 512 MB, but if I was buying today it'd be a 290/x without question.

Ok but we're talking about people who are completely brand agnostic who are gaming at 1080p, with 60 FPS minimums.

Nvidia shines at 1080p and this user will turn down settings to get to a 60 FPS minimum.

If he had expressed interest in DSR/Ultra Textures then I'd say go with the R9 290x.

Also, lets be real here, a lot of big announcements recently are gameworks titles. They perform horribly on AMD GPUs at launch. Yes, I get it already, the R9 290x is a great card as a whole. But it doesn't fit every gamers needs and certainly if you're not going above 1080p and you want to push as many frames as possible, you get a GTX 970.