GeForce GTX 970 Design Flaw Caps Video Memory Usage to 3.3 GB: Report

[the54thvoid]

Having read the full thread at OCN there are lots of amateurs posting results from an improperly run benchmark. Quite funny.
Regardless, from all that debate I'm picking up conjecture, conspiracy and 'stabs in the dark'.
Issue needs addressed formerly, not by a few end users.
Nvidia will have a case to answer though. Regardless of the badly run benchmarks, people are reporting tanking frame rates when memory usage crosses a threshold that is reasonably below the cards hardware limit. Who knows, may be the texture compression algorithm uses Vram itself.

 

[Recus]

Yes, if it says 4GB in specs.

Just like 2 TB HDD is 1.81 TB or 4.7 GB DVD is 4.3 GB?

 

[qubit]

Hehe, I'm glad nvidia cards would never suffer such awkward problems and it's always AMD with their substandard products. Oh, wait...

 

[Pumper]

Just like 2 TB HDD is 1.81 TB or 4.7 GB DVD is 4.3 GB?

Nothing to do with missing memory, just different unit systems used. And as every single HDD, SSD, DVD, etc. used this, there is no issue. But when one GPU says 4GB and has only 3.3GB while every single other GPU says 4GB and has 4GB, we have a problem.

 

[CounterZeus]

Just like 2 TB HDD is 1.81 TB or 4.7 GB DVD is 4.3 GB?

No. That's actually fucking up the scientific names of the quantities.

1TB = 931GiB (Gibibyte, using 2^10=1024 instead of 10^3=1000)
If that was the case, 4GB of memory would be 3.725 GiB (4*1000*1000*1000/1024/1024/1024)

 

[64K]

Someone mentioned that Nvidia wouldn't do a recall on the other thread which is now closed so I'm moving my post over here

If Nvidia screwed the pooch on the GTX 970 then I don't see how they could avoid a recall if they sold a defective product. I suppose they could try but then all of those nasty legal fees will be added onto the cost of finally having to do a recall anyway (I'm looking at you Toyota) so I'm not worried. At most I will be inconvenienced.]

I'm using the word "If" for now and that's only if it couldn't be fixed through a patch or firmware upgrade.

 

[john_]

If this is a pcb design flaw a bios update most likely won't fix this, but i doubt that soo many custom PCB's will have the same "design flaw" and should be able to take it with a bios update, hopefully or i would return my card and buy a GTX 980 instead.

So you are going to pay Nvidia extra, for selling you a faulty card? Interesting.

 

[Recus]

So you are going to pay Nvidia extra, for selling you a faulty card? Interesting.

In Youtube you can't find gameplay videos where 290X not using all vram.

Also this http://www.overclock.net/t/1535502/gtx-970s-can-only-use-3-5gb-of-4gb-vram-issue/350#post_23451063

 

[john_]

In Youtube you can't find gameplay videos where 290X not using all vram.

Also this http://www.overclock.net/t/1535502/gtx-970s-can-only-use-3-5gb-of-4gb-vram-issue/350#post_23451063

Sorry. Didn't understood the comment about 290X.

As for the link, if he does replace his card that would mean that the problem was verified and it can't be fixed. Nvidia of course could replace all the cards, but I believe they will choose to give a coupon or something if it affects a big percentage of 970.

 

[Fluffmeister]

Certainly interesting, this dude with a 8GB 980M (GM204 with an additional SMM disabled compared to the 970) had no issues:

http://forums.guru3d.com/showpost.php?p=4998360&postcount=121

 

[The N]

they seemed was in hurry, to BLAST the whole market with 970s. though it will resolve through any latest update. as they 're working on it.
But how come they flawed there TOP product, while it was selling in market right now. the FLAW of memory is noitceable, 700mb isn't minor at all.

will this effect on 970 sales in market????

 

[qubit]

I'd be gobsmacked if this was a hardware fault somehow. It couldn't pass the various acceptance tests at the design stage if it was.

Most likely someone at NVIDIA forgot a comma or something silly like that when compiling the reference BIOS which an update will fix. It's still an embarrassing problem which shouldn't have happened, though.

 

[RejZoR]

So much for the best GPU ever and then it fails at something as basic as memory addressing... NVIDIA, I'm disappoint.

 

[AsRock]

If this is a pcb design flaw a bios update most likely won't fix this, but i doubt that soo many custom PCB's will have the same "design flaw" and should be able to take it with a bios update, hopefully or i would return my card and buy a GTX 980 instead.

Well done, so nVidia screw up you reward them, well done you.

I believe you need to step away from PC''s for a moment.

they seemed was in hurry, to BLAST the whole market with 970s. though it will resolve through any latest update. as they 're working on it.
But how come they flawed there TOP product, while it was selling in market right now. the FLAW of memory is noitceable, 700mb isn't minor at all.

will this effect on 970 sales in market????

Not really, due to people like Puma99dk|.

 

[GorbazTheDragon]

Sorry, don't have time to read through the whole thread, but my guess is that either the missing portion is already being used by the OS, I heard several people suggesting that the benchmark was only supposed to be run with the display on the iGPU. Otherwise I am guessing this is a problem with the benchmark, or NVidia's (and possibly AMDs) memory allocation, as several people were getting similar results on different cards including 670s, 560Tis, etc. as the last percentage appears to be behaving in a similar way that a HDD slows down once it gets to the last ~1/8th of the capacity.

 

[BiggieShady]

It's not about addressing the whole 4 GB. You can use the whole 4 GB, but when moving data in the last 512 MB od VRAM memory bandwidth plummets from 150 GBps to 20 GBps

 

[sergionography]

Now I'm no expert on this front but if I remember correctly nvidia in the past used to cut down on the memory width with cut down chips, for example a gtx580 had 384bit memory width, while the gtx570 had 320bit width along with its one disabled sm and by default having less overall vram chips on the board. So was this just to seperate it in performance from the higher end part or was it an actual engineering restriction in the Fermi design meaning that one sm being disabled results in losing 32bit of that memory interface width(meaning parts of the memory interface is addressable only by a specific sm). Because by this theory if we look at gtx980 it has 256bit width with 4096mb vram and 2048 cores spread between 16sm of 128 cores each, so if we do the math each sm gets Fed by 16bit of memory width and 256mb vram. So in comparison if we look at the gtx 970 u have a total of 13sm enabled on the chip for a total of 1664 cores, now if we do the math that each sm is addressed by 16bit memory width and 256mb vram then you end up with 13sm x 256mb = 3328mb which seems to be in line with the reports. Now I am almost certain Kepler for example had cut down sm units disabled without effecting memory so Maxwell shouldn't be different but since there is a problem as mentioned it looks like there is 1 more things users might wanna test to verify this theory and that is by checking if the bandwidth is also taking a hit, meaning along with getting only about 3.3gb vram is the memory width intact or is it cutting 48bit(16bit per disabled am) of bandwidth ?


Edit: corrected sone of the math(16bit per sm)

 

[GorbazTheDragon]

It's not about addressing the whole 4 GB. You can use the whole 4 GB, but when moving data in the last 512 MB od VRAM memory bandwidth plummets from 150 GBps to 20 GBps

The low bandwidth would be a symptom of the GPU swapping out the overflow onto the system RAM...

 

[xfia]

Ahhh...so that's why I can play Battlefield 4 @4K resolution on a 970 SLI setup, using 3.6GB of VRAM at a smooth 50-60fps for hours with no problems is it?

..oh wait a minute.

someone with 295 quadfire said bf4 pushes them over 3.8gb at ultra 4k

 

[BiggieShady]

The low bandwidth would be a symptom of the GPU swapping out the overflow onto the system RAM...

The way this benchmark is coded, there is no overflow ... it allocates only the available vram. It all happens in the memory controller in the gpu, the benchmark is almost cuda only.

 

[TheoneandonlyMrK]

I think in situations like this it is possible to theorize a vast array of possible causes ,conspiracies and what not but at the end of the day it is what it is and will come to be what it was.

yet.


wow some argue crazey stand points that benefit no one yet delude a few, if NV fecked up fine ,they all do sometimes ,,

,its what they do next that really matters imho ie to appease or fix the issue.

 

[Fluffmeister]

Yeah I've been more than happy with my 970 and this apparent issue hasn't affected my enjoyment of the card at all since I've owned it.

With that said, I'm very interested to see if and or how they intend to address it.

 

[ZoneDymo]

damn Nvidia, y u no quality check?

 

[D007]

Nice QC.. Ugh..

 

[BiggieShady]

Yeah I've been more than happy with my 970 and this apparent issue hasn't affected my enjoyment of the card at all since I've owned it.

With that said, I'm very interested to see if and or how they intend to address it.

What's also interesting is that no-one wonder's if this is CUDA issue. Someone should build the same memory bandwidth benchmark using DX11 and some huge textures.