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

[Mei]

lol ill believe u string!

 

[james.miller]

Just for you guys. watch in 1080p/60, obvs, and make your own minds up. Bear in mind shadowplay makes it a bit hitchier than it was in gameplay and...its AC:U anyway - hitch city.

 

[Kaapstad]

That's a hell of a drop going from 1080p to 4K. Ouch

That is actually very good.

A rough guide is to devide whatever fps you get @1080p by 4 to see what you are likely to get @4K. Just make sure your 1080p fps figure is not CPU bottlenecked to start with though.

 

[stooeh]

Posting a 4GB shot is easy - performance is terrible though but likely at this point it does need more than 4GB VRAM sadly

Spoiler

 

[Dirk Diggler]

lol ill believe u string!

 

[PayneUK]

Now I'm sorry if this has been answered, but after reading through the last dozen pages I'm about ready to make a Jackson Pollock on the wall using a shotgun and my brains.

Has there been any word from retailers, as while I’d be willing to give it some time to see if any issue can be resolved but NVidia could drag this out for an eternity.

Has anyone actually tried asking for a refund based on this yet? What am I saying, of course SOMEONE has, and what would people consider an appropriate time scale before it’s OK for people to take matters into their own hands?

 

[crinkleshoes]

Someone on Guru3D came up with the same hypothesis.

It seems quite clear on the speed of those memory benchmarks.

The memory bandwidth alone (and thus, transfer speed) should be the same on both 970 and 980 due to them both supposedly being 256bit @ 7GHz... the limitations in shaders should not make any effect to a benchmark relying solely on that speed.

So yeah... it is a separated bus & not complete as advertised.



I'm still happy with the current performance of these 970s, no doubt there...

In the game it just appears as a stutter. Powerplay recorded the glitch.

wasnt particularly relevant the first time that meme was posted.

Anyway we need more FC4 testing. i cannot replicate the hitching and nose dive in performance in other games. I dont have FC4 unfortunately.

... however this is exactly the type of thing I am concerned about and as games come to use this type of vram more and more, the more it will happen.

Had they sold the GTX970 as a 3.3GB 208bit card, it is likely that I would have still bought a pair of them for my latest build as they would still have been the best bang for buck.

I would actually have much preferred them to do this as it would avoid what is shown in that video. The colour-bar flickering seems extreme & I am surprised by that... to me that indicates a memory chip flaw rather than just a bus-width issue.

It's the stuttering that is the problem for me... that would/will drive me nuts... I can't stand gaming like that... especially when I've done all I can to eliminate that type of behaviour.

IF the cards were capped at 208bit/3.3GB... then that type of behaviour would not happen because the games/systems would not try to use above 3.3GB... it would see a hard/firmware limit and would keep within the buffer (generally, of course you can still try to use more if you wish). The software is actually better at coping with it in this way.

However what this mis-alignment of configuration is doing, is making the software see 4GB as the hard/firm limit and try to use it, with no way of compromising for the slow top-end.

This is why I'm frustrated by it & would like to see it fixed somehow.



For me, I would actually be happy enough (although still frustrated) with an updated bios that capped the card at the primarily usable 3.3GB. It would still be unacceptable business practice that should be compensated for, however. It's not ideal and I know others would not be happy with this option.

Of course most will be hoping for a bios update that unlocks the extra shaders and bus width, effectively makes the 970 into a 980. But then what would NV do for 980 owners other than them generally being able to clock higher with higher TDP?

It will be interesting to see what NV come back with.

 

[Dirk Diggler]

Latest Nvidia BIOS was released 2 days ago, and doesn't mention this fix

http://www.gigabyte.eu/products/product-page.aspx?pid=5209#bios

 

[Legend]

Of course most will be hoping for a bios update that unlocks the extra shaders and bus width, effectively makes the 970 into a 980. But then what would NV do for 980 owners other than them generally being able to clock higher with higher TDP?

But is that even possible? From what I'd read I didn't think it was, if hardware.

 

[crinkleshoes]

http://www.pcper.com/news/Graphics-Cards/NVIDIA-Responds-GTX-970-35GB-Memory-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.

There. What they haven't said, however, is what happens when you actually go over 3.5gb and try and use the last 500mb.

Wait, so the GTX970 is actually a 3.5GB card with 500MB of shared memory or something?

Erm, sort of. Basically the card sees 3.5gb as the main VRAM, then there's 500gb that runs like utter crap when you use it.

I just read that apparently Gigabyte have released a bios that fixes it, yet no real evidence as yet.

I'm not convinced that that is an official Nvidia statement as quite simply, they have the maths wrong.

It's possible there was just simple rounding to make the conversation simpler or it was communicated by someone who didn't fully understand.

For reference the split will be:
2048/1664 shaders
256/208 bit
4096/3328 MB
4GB/3.25 GB

So the segmented section is actually 0.75GB...

No no no dude, that's not true. The ram is all the same, be it Samsung, Hynix, whatever. It's the memory controller causing the issue. So for example, you use up to 3500mb ram and you get full 256 bit bandwidth. Then you over flow that and run into the last and you only get like 64 bit bandwidth..

Ugh, it's hard to explain but yeah, the ram is all the same. They don't fit the cards with 3.5gb of good ram then 500mb of crap ram.

48bit... the maths are simples

But in essence yes... it the a limitation in the bandwidth that the GPU is able to access that memory.

But even then, the benchmarks of that top-end memory do not add up... even at 48bit, the speed of the memory should be faster than the bench is indicating. There must be more in the way they have bodged the internal cross-wiring to get the other SMMs to access the extra memory that slows it down even more.

So the ram itself is the same... only the 13/16ths of the GPU that are enabled in a 970 only have very slow badly wired access to the remaining 3/16ths of the memory.

 

[Dervious]

My Nvidia experience with a GTX460 was far from good (a nightmare).

This issue, combined with numerous coil whine reports, suggests that I should stick with AMD.

I'd switch sides tomorrow if Nvidia provided evidence to convince me that it would be a good move.

 

[Bonjour]

*snip*

So the ram itself is the same... only the 13/16ths of the GPU that are enabled in a 970 only have very slow badly wired access to the remaining 3/16ths of the memory.

I doff my cap to your numerical prowess, good sir.

 

[crinkleshoes]

But is that even possible? From what I'd read I didn't think it was, if hardware.

Don't know... just saying I know that that is what people will be hoping for, I imagine

Can't say I would complain.

I don't know how they have disabled that "part of" 3/16 of the core... (note, not whole 3/16 is disabled as they have cross-wired some).

If it was simply a firmware update, I think we would have already seen unlocked bioses by now... like there was with ATI X800->X800XT mods (I think, or was it another similarly aged series?)

I know the TDP limit is hardware based... put in place by a resistor and I was curious how to unlock that to play with overclocks... you basically bypass the resistor for a TDP... but of course that's a completely different part of the card.

 

[Dash RipRock]

Seen this posted on the geforce forums by someone,sounds quite interesting and a possible reason maybe ?




"OK, this makes more sense:

There seems to be a lot of confusion about this, so I'll just post how I interpretted, the statement:
What they're saying is that there are two memory modules, 3.5GB + 0.5GB, they both perform the same, but the 3.5GB is the primary memory pool, so it will always access that first, and then it will access the 0.5GB if it needs to.

The problem is that memory testing programs, and monitoring programs, cannot access that 0.5GB module, so when a monitoring program like MSI Afterburner is monitoring memory use, it only sees the 3.5GB memory module, if the card is using >3.5GB, monitoring programs will only show 3.5GB, as they cannot monitor the 0.5GB.

Similarly, the memory tester that was hot on this issue, cannot access the 0.5GB, but knows that is a 4GB card, so when it tries to get the final 0.5GB, it gets "redirected" to system RAM, which has much less bandwidth than video memory.

When people see >3.5GB reported by monitoring programs, I'm going to guess what's happening there is they are beginning to use >4GB, causing system memory to be allocated, which is then added to the monitoring program's measurement, making it look like >3.5GB is causing massive slowdowns, when in reality it's >4GB that's causing massive slowdowns.

Anyway, that's my interpretation, I could be completely wrong, so... Grain of salt and all that.

Any comments on this? "


Does sound like a possible reason to me but who can say at the moment.One things for certain ,Nvidia have to make some proper response and fast if they dont want to hurt sales more than they possibly have been already.

 

[james.miller]

Anyway, that's my interpretation, I could be completely wrong, so... Grain of salt and all that.

Any comments on this? "

check the usage in my ac:u video, if it was swapping ram i dont think i'd manage 10fps, let alone 30 lol

 

[Dash RipRock]

the " " was the quote from a poster on the geforce forums,not mine.Your footage does show that afterburner is recording memory usage above 3.5 GB and performance doesn't appear consistent with paging to system memory.Its a puzzler alright .

 

[Bonjour]

Is Unity an NVidia optimised game?

 

[r7slayer]

Is Unity an NVidia optimised game?

is a ubisoft game and they tend to be nVidia games. And yea sorta optimised. As optimised as you can get with a broken buggy game.

 

[moogle]

Didn't the benchmark run fine with aero disabled? Surely that would mean the ram is running normally if that was the case.

 

[crinkleshoes]

I doff my cap to your numerical prowess, good sir.

LOL... it was just nagging at my (relatively minor) mathematical OCD.

check the usage in my ac:u video, if it was swapping ram i dont think i'd manage 10fps, let alone 30 lol

I thought you had posted that video because you had noticed the stuttering.

I don't think you understand the issue & I actually see it, albeit minor, in your video... does that microstutter every couple of seconds not bother you? THAT is the issue... it is more visible on the previously posted FC4 video (ignore the colour splashes).

Using the extra memory / paging 10% of the textures will not cause the FPS to suddenly drop by 70%.

It may cause the FPS to drop by 5-20% but it's the stutters every few seconds that are occurring... that is the swapping and/or use of extra slow memory bank.

It's not going to load all the textures from swap over and over again and 85%+ of the textures are still being read/written at full speed.

The stutter is clear in the video you posted of AC:U.

the " " was the quote from a poster on the geforce forums,not mine.Your footage does show that afterburner is recording memory usage above 3.5 GB and performance doesn't appear consistent with paging to system memory.Its a puzzler alright .

I don't think he's accurate as the relationship won't be as he is referring to it.