970's having performance issues using 4GB Vram - Nvidia investigating

[PayneUK]

I'll preface this by saying I own AMD GPU's and an Nvidia 980 GTX, so try not to pass this off as bias.

But there do seem to be a lot of people almost letting Nvidia off here, whereas if this was an AMD issue the outcry would be insane.

Possibly because more people are ‘pro’ NVidia by default?, or people who just want an answer are finding it hard to tell between the people who want to genuinely test this issue and the people who want or don’t want there to be a problem regardless of facts.

Hard to blame someone if you don’t know if there is anything to blame people for, personally from reading around the net I’ve come to the conclusion that there is an issue, but I’ve given up trying to determine how bad it is from the posts on any forum.

I’ve not been home for a while so I’ve not been able to try out any testing since this thread went up, all I know is that prior to coming away for work I swapped out my 970’s to box up and sell and put in a 780 I had. I got in a bit of time on Dragon Age before leaving and honestly while the FPS was a little lower on the same settings (I ran with SLI disabled due to it sucking in DA:I) the experience was MUCH smoother.

So I’m looking at a way to get my money back, I’m not sure what avenue I’m going to take yet but I’m certainly not going to keep the cards.

 

[LambChop]

Has NVidia said anything officially about this possible issue ?

 

[CAT-THE-FIFTH]

I don't even know what "Bumpgate" is

It was something that cost Nvidia(eventually) at least $200 million.

They basically chose a cheaper way to attach certain G80 series GPUs to the PCBs. In more thermally constrained situations with repeated heating and cooling cycles it would mean the GPU would come away from the PCB,causing it to stop functioning properly.

Nvidia try to deflect from solving the problem - I knew a lot of people with laptops affected by it.

The lack of coverage by the tech press was appalling - I knew loads of people who had the issue with their laptops. Luckily they knew me as I managed to help them get replacements and not get palmed off.

Companies tried to fix the problems by giving updates to laptops which ran the fans at higher speeds,and only after months of this happening Nvidia admitted there was a issue but not before class action lawsuits,some of which took years to end.

Charlie Demerjian who with The Inquirer at the time found out what the problem was:

http://www.theinquirer.net/inquirer/news/1004378/why-nvidia-chips-defective

Nvidia nevered confirmed it and I suspect the way Nvidia burned him over it,was why he ended up hating them ever since.

 

[old gamer]

if there is a real problem with the 970 and nvidia fob it off, then there is a danger of second hand prices taking a serious nose dive

 

[MOOGLEYS]

Has NVidia said anything officially about this possible 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.”

Tbh it sounds like an excuse to me.

 

[CAT-THE-FIFTH]

Its funny that if it were not for AMD's MOAR VRAM advertising Nvidia would have advertised the card with 3.5GB,just like the GTX570,GTX470 and GTX275 had less VRAM and memory bandwidth than the top end cards they were based on.

 

[LambChop]

Tbh it sounds like an excuse to me.

Thats not the best response from NVidia I have to admit, but it does explain it. Although I dont like that performance takes a drop when the second segment is accessed, why should it ? I don't fully understand what crossbar resources are either.

I'd need someone to explain exactly what NVidia means.

 

[Bonjour]

At least they admit the last 0.5GB operates differently to the rest.

 

[LambChop]

Its funny that if it were not for AMD's MOAR VRAM advertising Nvidia would have advertised the card with 3.5GB,just like the GTX570,GTX470 and GTX275 had less VRAM and memory bandwidth than the top end cards they were based on.

Stop guessing stuff.

35 pages of **** is hard enough to trawl through when looking for factual info.

 

[ubersonic]

Companies tried to fix the problems by giving updates to laptops which ran the fans at higher speeds,and only after months of this happening Nvidia admitted there was a issue but not before class action lawsuits,some of which took years to end.

I wish Dell had been one of those companies /sigh, I had an XPS M1730 and a XP£ 1330 back in the day and we were literally begging them to release a BIOS updates that either gave us control of fan speeds or just set a performance mode (they only ran the fans in silence mode because they figured thats all laptop users cared about).

The GPU's design was at fault OFC but with adequate cooling they would have lasted much longer, it put be off buying another Nvidia based Dell as it set the opinion in my mind that ATI could cope with Dell's **** cooling solutions better lol.

 

[CAT-THE-FIFTH]

Stop guessing stuff.

35 pages of **** is hard enough to trawl through when looking for factual info.

Increasingly look what has been posted here,Anandtech and on overclock.net and other forums it appears that the GTX980 lacks this problem,and its linked to the way the memory controller and SMXes are linked. Plus this is a discussion forum - imagine if we just kept to press releases. There would not be any threads.

Look at this post in the thread:

http://forums.overclockers.co.uk/showpost.php?p=27530983&postcount=897

Seems the best explanation so far TBH.

I wish Dell had been one of those companies /sigh, I had an XPS M1730 and a XP£ 1330 back in the day and we were literally begging them to release a BIOS updates that either gave us control of fan speeds or just set a performance mode (they only ran the fans in silence mode because they figured thats all laptop users cared about).

The GPU's design was at fault OFC but with adequate cooling they would have lasted much longer, it put be off buying another Nvidia based Dell as it set the opinion in my mind that ATI could cope with Dell's **** cooling solutions better lol.

The desktop cards tended to last longer(although my 8800GTS 512MB which was in a Shuttle die suddently,but I could have been unlucky as people have had 8800GT cards last for years).

I have an 8500GT lying around which probably works still.

 

[MOOGLEYS]

Thats not the best response from NVidia I have to admit, but it does explain it. Although I dont like that performance takes a drop when the second segment is accessed, why should it ? I don't fully understand what crossbar resources are either.

I'd need someone to explain exactly what NVidia means.

Without tools to read the 500MB of Ram Nvidia refer to it would be hard to prove.

I find it strange that this issue does not affect the 980M which shares most of its specs with the 970. I think there is more to it than just a separate 500MB portion tbh.

I'm sure someone will come up with a way to test Nvidia's claims at some point.

 

[james.miller]

Thats not the best response from NVidia I have to admit, but it does explain it. Although I dont like that performance takes a drop when the second segment is accessed, why should it ? I don't fully understand what crossbar resources are either.

I'd need someone to explain exactly what NVidia means.

Think of it like this: the 980 has 16 SMMs (Streaming Multiprocessor Maxwell). all 16 need to be in use at the same time to achieve the quoted bandwidth figures. Now, the 970 has 13 of them, the other 3 being disabled. Because of this, bandwidth and therefor memory performance is affected. It seems that nVidia have attempted to optimise bandwidth by segregating the memory in to two chunks, 3.5gb and 0.5gb, the 3.5gb chuck having a 'higher priority' according to nVidia. It seems like the right approach to me but it would appear to be causing some people more problems than others once you start using that smaller chunk.

This issue is muddied by the benchmark in the op, which apparently wasnt very well written and is possibly exaggerating the results, so further clarification needed there. In addition to this, there is a question over whether the various reporting programs (afterburner etc.) are reporting the right memory usage. It's possible that these programs are not reporting usage of the 500mb chunk, meaning people could be using 4gb of ram but afterburner etc reports <3.5gb. I'm not sure about this because we've had AB reporting 4gb and over so we need clarification there also; This would also be a good time for comments from the authors of these programs, such as Unwinder who wrote rivatuner (which afterburner is based on).

So, too many questions unanswered at the moment. But what does seem apparent is that these problems arent consistent from one 970 to the next. This could be down to which SMMs the 970 has disabled - most if not all 970s will be produced from rejected 980 cores, so disabling different parts of the GPU may have different affects on memory access.

 

[mattcrane]

34 pages......and, well......I am confused.

Do we actually have an issue or not? Seems that whatever is going on has indeed been over hyped.....but, I would like to know if there is a fundamental problem or not with the hardware itself.

Glad I am not the only one.....

I would like some advice though from people that do understand the issues if thats possible

I have a Ref 970 on Pre-order, now would I be better off cancelling that order and putting the money towards buying a Ref 980 or should I just stick with the 970.

Gaming is 24" @1080 with G-Sync Monitor, (being delivered today) if that helps any. But I want some future proofing as well.

Sorry if its off topic but I am that confused and really would like some peace of mind over it, as I don't have a lot of money to spare so I try to make the right purchase first time if I can.

 

[LambChop]

Think of it like this: the 980 has 16 SMMs (Streaming Multiprocessor Maxwell). all 16 need to be in use at the same time to achieve the quoted bandwidth figures. Now, the 970 has 13 of them, the other 3 being disabled. Because of this, bandwidth and therefor memory performance is affected. It seems that nVidia have attempted to optimise bandwidth by segregating the memory in to two chunks, 3.5gb and 0.5gb, the 3.5gb chuck having a 'higher priority' according to nVidia. It seems like the right approach to me but it would appear to be causing some people more problems than others once you start using that smaller chunk.

This issue is muddied by the benchmark in the op, which apparently wasnt very well written and is possibly exaggerating the results, so further clarification needed there. In addition to this, there is a question over whether the various reporting programs (afterburner etc.) are reporting the right memory usage. It's possible that these programs are not reporting usage of the 500mb chunk, meaning people could be using 4gb of ram but afterburner etc reports <3.5gb. I'm not sure about this because we've had AB reporting 4gb and over so we need clarification there also; This would also be a good time for comments from the authors of these programs, such as Unwinder who wrote rivatuner (which afterburner is based on).

So, too many questions unanswered at the moment. But what does seem apparent is that these problems arent consistent from one 970 to the next. This could be down to which SMMs the 970 has disabled - most if not all 970s will be produced from rejected 980 cores, so disabling different parts of the GPU may have different affects on memory access.

Good post, thanks james. I agree with a lot of that.

 

[Bonjour]

Stop guessing stuff.

35 pages of **** is hard enough to trawl through when looking for factual info.

I wouldn't bother. It's mostly rubbish conjecture. Hopefully we'll see some solid info this week.

 

[CAT-THE-FIFTH]

I wouldn't bother. It's mostly rubbish conjecture. Hopefully we'll see some solid info this week.

You mean like almost every thread on here which is about rumours or any degree of discussion. I will remember that in the future if you ever go off topic!

Good post, thanks james. I agree with a lot of that.

And next time instead of whining at me,bother to read the thread.

This was on page 30:

http://forums.overclockers.co.uk/showthread.php?p=27530983#post27530983

To the contrary, that s quite possible when looking at the GPU diagram, the memory controllers that are connected to the clusters have 16 bits of addresses, on a total of 64, that are indeed not used since they would feed disabled units, and the simplified MCs cant swap the data from the 0.5GB partitions RAM to feed a nearby unit, actually it s quite possible that those 0.5GB are not functional at all even if they can be addressed, that is you can fill it with datas but you have no mean to send those datas to functional units because the design lacks the relevant crossbars.

Yes you can load 4GB but only 3.5 can be sent to the SMMs, the data in the remaining 0.5GB cant be sent for execution to the SMMs used for executing the datas that are within the 3.5GB, it s only a theory but the GPU design and the presence of a separated partition point to this implementation, what transpire is that each SMM has a fixed adress space in the RAM and that data meant to be executed by a given SMM must be retired in the relevant adress space, if this given SMM is disabled there s no mean to send the data to another SMM, so as said the whole 4GB is adressable but only 3.5GB can be executed by the GPU computing units...

Someone I know on another forum suggested the problem could be dependent on what part of the GM204 die is defective? After all the GTX970 does use GM204 dies which cannot be used for GTX980 cards.

That was all in the last five pages of the thread.

 

[RanxZy]

I have a Ref 970 on Pre-order, now would I be better off cancelling that order and putting the money towards buying a Ref 980 or should I just stick with the 970.

If I was you I would go with a 980. Or if you have no issues with AMD a 290X is a great card for the price.

 

[StrongSad]

This all reads like doom n gloom.... still a great card for me though...

However my question

Have Nvidia changed the game optimal settings in Nvidia experience to take account of this?

It would be interesting to find out what the allocated memory was for AC U etc on the old software vrs new and if it was over 3.5 or it is below.

 

[james.miller]

sorry cat but there is an awful lot of crap in this thread, comment after comment trashing nvidia and/or people panicking. I don't blame people for not want to read though all this rubbish.

This all reads like doom n gloom.... still a great card for me though...

However my question

Have Nvidia changed the game optimal settings in Nvidia experience to take account of this?

It would be interesting to find out what the allocated memory was for AC U etc on the old software vrs new and if it was over 3.5 or it is below.

Well, i'm on the latest drivers and you can see in the AC:U video i posted that AB reads over 4gb usage at points. This means either i was using 4gb or, if these programs really dont read the last 500mb, i was using 4.5gb and paging to the system ram. you can see from the video that the hitching was no worse than is usual for ACU so i'm reluctant to believe i was exceeding 4gb of usage.