Mikael_Wartooth

matthewdg1973

Mikael_Wartooth

Nereus

Mikael_Wartooth
If EVGA let me step up to a 980 I would send my 970 back in a heartbeat and I would also buy a second 980 this week when I purchase the last of my parts for my X99 build. If this issue was known before the deadline for my step up I would have sent it back already. 

I purchased my 970 back in September also. I noticed way back then the strange issue with the spike, and it was discussed on these forums. I returned the card for a refund to Newegg but more because of the noise. I since purchased 2 x 980 SC card and have had no issues. My point is, the subject about the memory only spiking to 4GHz and not remaining there was discussed back then too. Sorry you missed it, but I think you're stuck now.
 

Not stuck at all. EVGA is letting me step up and pay the difference for a 980 SC. Best Customer service in the biz. 

Are you sure its for the SC version and not the vanilla 980? Only asking because I was offered a step up but only to the vanilla 980.

Sorry, yes it's the vanilla version. Which I'm still ok with. 

Yeah I'm fine with it too. This card should hold me over for the time being.

LOL, I can't really get you guys.... You've just been f*cked up by Nvidia, and now you are buying a new, more expensive card from them again to fill their pockets even more. It's ridiculous...

Guys… of course it has 4GB that can be all accessed.
BUT only 3.5GB are connected in the regular and fast way. The other (last) 0.5GB are very slow. So any game using more than 3.5GB will have massive FPS drops!

Well… the conclusion in most test is that it is hard to test. Especially on such (already) low FPS.
Still, the facts are:
Nvidia needed to tackle that problem and did so by splitting the total 4GB memory into a primary (196 GB/sec) 3.5GB partition that makes use of the first seven memory controllers and associated DRAM, then there is a (28 GB/sec) 0.5GB tied to the last 8th memory controller.
 
Source: guru3d.com/news-story/middle-earth-shadow-of-mordor-geforce-gtx-970-vram-stress-test.html

Dschijn
Well… the conclusion in most test is that it is hard to test. Especially on such (already) low FPS.
Still, the facts are:
Nvidia needed to tackle that problem and did so by splitting the total 4GB memory into a primary (196 GB/sec) 3.5GB partition that makes use of the first seven memory controllers and associated DRAM, then there is a (28 GB/sec) 0.5GB tied to the last 8th memory controller.
 
Source: guru3d.com/news-story/middle-earth-shadow-of-mordor-geforce-gtx-970-vram-stress-test.html

The 3.5 GB portion is only up to 196 GB/sec when measured synchronously.  In other words, if you are transferring data equally among all the first 7 memory chips, you can achieve up to 196 GB/sec.  Each individual memory chip is only 28 GB/sec.  The 8th memory chip is also 28 GB/sec.  So, technically, they all have identical performance on their own.  The only difference is that the 8th memory chip can't do synchronous transfers with the other 7 chips.  If the 8th chip was synchronous with the rest, the total synchronous memory throughput for the card would have been up to 224 GB/sec.
 
So, you are correct but also sort of incorrect.  You are sort of incorrect because you are assuming that the first 7 chips will always be able to operate at 196 GB/sec which is not true at all.  How close the data throughput comes to the theoretical maximum 196 GB/sec is completely dependent on the game/application/drivers/operating system.  It is not possible to optimize all applications to use all resources in the best way at all times.  There are many mathematical and application situations where it is not possible to make everything perfectly synchronous and perfectly optimized. 
 
If you look at the problem "top down" -- the way I was arguing earlier -- the card has 4 GB of VRAM as advertised and all 8 chips operate at the exact same 28 GB/sec.  No chip is slower than any other chip; on their own.  Synchronous transfer ability is the only difference.  It is difficult to measure the performance "lost" because that is completely dependent on the game/application/drivers/operating system (as previously mentioned).  If the operating environment does not allow synchronous transfers to be utilized effectively, all chips will be running at 28 GB/sec and there will be absolutely no performance lost at all.  As some of the evidence shows in this thread, some people can definitively determine that in their case, there is zero noticeable performance lost due to the last 0.5 GB being "slow".
 
In the real world, how much performance is lost due to the 8th asynchronous chip is directly proportional with how synchronous and optimized the application is.  And of course, in the real world, the application you are using must require more than 3.5 GB of VRAM in the first place; otherwise it is a moot point. The first 7 memory chips will operate somewhere between 28 GB/sec and 196 GB/sec.  Finding that unknown and unpredictable value of application synchronicity is the only way you can safely make any further claims regarding performance loss.
 
Since it is impossible to objectively determine the application's true synchronicity, the only claim you can make is that the 8th asynchronous chip will operate between 0% (no performance lost) and 700% slower than what the other 7 chips are capable of.  Any logical person will know that the theoretical maximum 196 GB/sec will probably never be achieved simply because no application is perfectly optimized and all data signals require some additional overhead for error correction and data corruption prevention.  According to the CUDA benchmark used as evidence in this subject, the 7 other chips might only operate at around 150 GB/sec in the real-world (at least in that application of course).  In that specific case, the 8th chip operates somewhere between 0% and 535% slower than what the other 7 chips may be capable of; depending on the application.  As most would agree, this CUDA application is very simple and designed specifically to show the worst case scenario.  Most would agree that any real-world application is highly unlikely to achieve anything close to that 150 GB/sec synchronous access number simply due to the complexity of the application and the inability to make all calculations in a synchronous manner.

ty_ger07

Rgallant

banzaigtv
3.5 GB of VRAM? I don't think so. It has 4 GB of VRAM. I nearly maxed it out and I can prove it. By running Dragon Age: Inquisition on maxed settings in 3840x2160 resolution.
 

just saying.
I see 3957mb ,  so not 4096 mb so not 4gb nor max vram

Are you serious?
 
3957 MB is more than 3584 MB.  So, it proves that the card does have more than 3.5 GB.  The original claim is that the cards only have 3.5 GB which we all know is false.  When he is using 3957 MB -- as shown in his screenshot -- he is definitely utilizing that 512 MB which supposedly doesn't exist.
 
And for those poo pooing his screenshot because the framerate is so low:
Look at the GPU utilization.  It is using 99% of the core's ability.  In other words, the low framerate has nothing to do with the supposed slowness of that last 512 MB of VRAM.  The core simply is not powerful enough in his demonstration.

No one ever said that 512MB was missing that I know of. And if you consider poo pooing a screenshot because someone asks if he plays at that rate then you need to take a look inside yourself and ask " why does this affect me so?".
 

matthewdg1973

ty_ger07

Rgallant

banzaigtv
3.5 GB of VRAM? I don't think so. It has 4 GB of VRAM. I nearly maxed it out and I can prove it. By running Dragon Age: Inquisition on maxed settings in 3840x2160 resolution.
 

just saying.
I see 3957mb ,  so not 4096 mb so not 4gb nor max vram

Are you serious?
 
3957 MB is more than 3584 MB.  So, it proves that the card does have more than 3.5 GB.  The original claim is that the cards only have 3.5 GB which we all know is false.  When he is using 3957 MB -- as shown in his screenshot -- he is definitely utilizing that 512 MB which supposedly doesn't exist.
 
And for those poo pooing his screenshot because the framerate is so low:
Look at the GPU utilization.  It is using 99% of the core's ability.  In other words, the low framerate has nothing to do with the supposed slowness of that last 512 MB of VRAM.  The core simply is not powerful enough in his demonstration.

No one ever said that 512MB was missing that I know of. And if you consider poo pooing a screenshot because someone asks if he plays at that rate then you need to take a look inside yourself and ask " why does this affect me so?".
 

Do you know what this thread is about?  Your comments don't seem to have anything to do with this topic.
 
The purpose of this topic is to:
1) Discuss whether or not the card only has 3.5 GB of usable VRAM.
2) Determine if the last 0.5 GB of VRAM is indeed slower that the first 3.5 GB.
 
You will see that my comments were in regards to both #1 and #2.  Your comments were completely out of left field and had nothing to do with either #1 or #2.

misiak
LOL, I can't really get you guys.... You've just been f*cked up by Nvidia, and now you are buying a new, more expensive card from them again to fill their pockets even more. It's ridiculous...

Because AMD is garbage? If there was a better alternative I would go the route, but there isn't so... 

Mikael_Wartooth

misiak
LOL, I can't really get you guys.... You've just been f*cked up by Nvidia, and now you are buying a new, more expensive card from them again to fill their pockets even more. It's ridiculous...

Because AMD is garbage? If there was a better alternative I would go the route, but there isn't so... 

I don't see where you have specified whether or not this potential performance issue is actually affecting you.  If this issue is not affecting you, you should continue as you originally planned.  There is no reason to stress out over this potential issue or upgrade to a higher priced card if unnecessary.
 
As has been discussed, this issue does not appear to really affect the majority of people.  As has been discussed, in situations where this issue could potentially manifest itself, performance is already so low (low framerate due to overloaded core) that the majority of people don't even enter into that territory.
 
The one interesting caveat is that you are considering upgrading to SLI.  In that configuration, it will take a lot longer to reach the overloaded core scenario and therefore maybe that does increase the chances of reaching a situation where this issue does manifest itself.  I am not sure that there is enough data out there to really make an educated decision at this time in regards to an SLI setup.
 
As has been discussed, this issue only manifests itself when every chip is attempted to be utilized synchronously ("at the same time").  Whether or not that will occur is very much dependent on the application which is utilizing the graphics card.  Additionally, NVIDIA says that they plan to better distribute (via driver updates) synchronous versus asynchronous data calls to ensure that it is less likely that the 8th chip is subjected to a situation where synchronous data calls are being made to that chip.  This further highlights why making an educated decision at this time is difficult.
 
Lastly, if your usage (determined by your observations in the games/benchmarks/applications you use) does not require greater than 3.5 GB of VRAM, you will never have to worry about the potential performance implications of the 8th asynchronous chip in the first place.
 
 
I hope that you read what I wrote here:
http://forums.evga.com/FindPost/2288720
I also hope that what I wrote there is understandable.  And of course, I hope that what I wrote is completely accurate.

jasonwatkins

Dschijn
Guys… of course it has 4GB that can be all accessed.
BUT only 3.5GB are connected in the regular and fast way. The other (last) 0.5GB are very slow. So any game using more than 3.5GB will have massive FPS drops!

On Battlefield 4, at 120% resolution scaling w/ max settings @ 4k, both the GTX 980 and GTX 970 use around 3.4 GB of VRAM.  GTX 980 FPS: 22.8 FPS, GTX 970: 19.2 FPS.  970 is 18 percent slower, which is about what you'd expect since the 980 is a faster card.
Now if you bump up to 150% resolution scaling to increase the memory usage, both cards use around 4GB of VRAM.  GTX 980 FPS: 15.0 FPS, GTX 970: 12.8 FPS.  970 is 17% slower, which is roughly the same speed difference versus 3.4 GB of usage.
(source for the above values)
 
So I ask, where is this MASSIVE FPS drop you're referring to?!?!?
 
 
 
 

The problem is not about massive drops but about stuttering when app utilize more than 3.5GB with extensive data because of slow bandwidth of the last 0.5GB block of memory. This can be easily observer by analysis of frames time. You would have very high peaks during that time...

misiak

-ZS-Carpenter

Scarlet-Tech

misiak
LOL, I can't really get you guys.... You've just been  DELETED FOR FOUL LANGUAGE up by Nvidia, and now you are buying a new, more expensive card from them again to fill their pockets even more. It's ridiculous...

You keep adding a * into the vulgarities you want to post. Stop. You are going to get banned from the forum for being foolish enough to constantly break the rules about language.

STOP.

you got it backwards, he needs to add more profanity to his posts so the rest of us need not run upon the constant nonsense he has been posting

@Scarlet, sorry but a better word did not come into my mind, English is not my primary language and I didn't know there are such rules here. I've used asterisk if you have noticed :)
 
@Carpenter, could you please be more specific which posts exactly were nonsenses because from you backlog seems you have answer for everything...

Per the forum rules, and I do understand that English is not your first language, vulgar language, even if abbreviated or symbols replace such language, will not be tolerated and you may receive an infraction.

Regardless of his choice of language he does bring up the point that keeps getting brushed aside. It's not a fps issue, it's a stuttering issue from what I've learned is frametime, not framerate.
 
That's what I encounter in my 970 SLI setup@4k. FPS will be steadily above 40 but there is a persistent yet random stutter in some games once the 3.5gb threshold is crossed(honestly it's right before 3.5 when it starts). I've said it before it's not all games. BF4 is great even above 3.5. Unity, FarCry4 and DA:I are terrible for me.

much ado about a minor thing as i run at 1920x1080 it is a zero "problem" for my ssc gtx970.

misiak
 
 
@Carpenter, from what you said I expect you own a GTX970 card, do you ? If not, then we don't need to talk anymore. Obviously you are an nvidia fanboy and trust everything they will say. Look, I'm nvidia customer for maybe 10 year and only once I had ATI card. I was always satisfied with Nvidia. But do you really think they are so stupid to admit they fault ? Do you really think than engineers did not communicate with marketing the specs they gave to reviewers. And that no-one from engineers noticed this discrepancy in specs for more than 4 months ?? If you do, then fine. The last few days I spent lot of time investigate the issue and I had simillar findings than lot of users who tested it as well and encountered an issue. COD:AW was great example. May I ask if you test any game with 970 ?
 

 
I don't own a 970 I own 3 of them and when i bought them over 2 months ago, long before this memory issue surfaced, I knew that in a couple games at very high res I would have to sacrifice a couple settings to keep things in check. With this ever so slight issue I may need to maybe turn 1 more setting down 1 notch to keep playable frame rates. The sky is not falling, your house isn't burning down, you just have to settle for 2x AA over 8x AA....cry me a river
 
And you can stuff your fanboy comment up your COD:AW. You have done nothing but hunt out the most absurdly bad console trash ports to show any kind of real world performance hit. Everyone(except you) already knows that poor console ports run like crap. They suck up resources like a vacuum, it's nothing new and your proving nothing by using them as your "source" for condemning the 970 and nvidia. Yea mistakes were made but in reality it is not as big of a deal as you are trying to make it out to be.
 
   

ty_ger07

Mikael_Wartooth

misiak
LOL, I can't really get you guys.... You've just been f*cked up by Nvidia, and now you are buying a new, more expensive card from them again to fill their pockets even more. It's ridiculous...

Because AMD is garbage? If there was a better alternative I would go the route, but there isn't so... 

I don't see where you have specified whether or not this potential performance issue is actually affecting you.  If this issue is not affecting you, you should continue as you originally planned.  There is no reason to stress out over this potential issue or upgrade to a higher priced card if unnecessary.
 
As has been discussed, this issue does not appear to really affect the majority of people.  As has been discussed, in situations where this issue could potentially manifest itself, performance is already so low (low framerate due to overloaded core) that the majority of people don't even enter into that territory.
 
The one interesting caveat is that you are considering upgrading to SLI.  In that configuration, it will take a lot longer to reach the overloaded core scenario and therefore maybe that does increase the chances of reaching a situation where this issue does manifest itself.  I am not sure that there is enough data out there to really make an educated decision at this time in regards to an SLI setup.
 
As has been discussed, this issue only manifests itself when every chip is attempted to be utilized synchronously ("at the same time").  Whether or not that will occur is very much dependent on the application which is utilizing the graphics card.  Additionally, NVIDIA says that they plan to better distribute (via driver updates) synchronous versus asynchronous data calls to ensure that it is less likely that the 8th chip is subjected to a situation where synchronous data calls are being made to that chip.  This further highlights why making an educated decision at this time is difficult.
 
Lastly, if your usage (determined by your observations in the games/benchmarks/applications you use) does not require greater than 3.5 GB of VRAM, you will never have to worry about the potential performance implications of the 8th asynchronous chip in the first place.
 
 
I hope that you read what I wrote here:
http://forums.evga.com/FindPost/2288720
I also hope that what I wrote there is understandable.  And of course, I hope that what I wrote is completely accurate.

Right now the 970 is fine for what I am doing, but I am in the middle of a X99 build and will be going 1440 or 4k so I am going to SLI 980's instead of 970's. Not a big deal.