Graphics Nvidia’s GTX970 has a rather serious memory allocation bug

Do you have a 970? Or are you running AMD by chance? You are taking such a keen interest.

I unfortunately do not have a copy of COD:aw to test on, but as the numbers and percentages are incomprehensible and the (one would assume) standardized test setup unverifiable, I will continue to believe there is no problem. If someone discovers a test or game which we all can run that shows the 970 to be a dud, by all means share it. That youtube clip could be direct from AMD fanboy headquarters for all the Russian we can read.

 

neither - I have a GTX 670 (and actually interested in a titan)

with the Russian website showing an issue , are they not a verified source? or do you only read western websites ?

 

NVidia have already said the card has 3.5GB of full speed ram , with the extra 500mb addressed separately - pretty much it was `tacked on` to make it a 4GB card.

they should have just called it a 3.5GB card

 

Dunno, mate, perhaps you could give us all a verbatim translation from the Russian so that we can all see what you're seeing here?

 

Are they showing an issue? You read Russian? The link on the video leads to this website, and I can see as of today they have posted Nvidia's official response. Aside from that I need google translate to tell me what they are saying, which leaves a lot to be desired.

I have over £1,000 invested on two GTX 970s and a 4k monitor. So this rumour is particularly significant to me. You however are posting plenty of hearsay and, now, a dubious Russian video, in the mean time you own a 670 and are considering a Titan?

 

You may actually want to read PCPER's article:

 

I've re-watched the video now you have provided some context to what it was for and honestly I'm not sure i see any micro stuttering or not, i think maybe i do but TBH i probably wouldn't even notice something like that if it was staring me in the face.

Didn't some web site run frame by frame analysis to identify when it was suspect AMD cards suffered from micro stuttering ?

This has also nothing to do with being biased, it's to do with people using a program that's not fit for purpose, and is next to useless.

I'm not sure how anyone can "totally disregarding PCPER - to the point of calling them out right liars", all they have done is post an article that pretty much says the same thing as Gareth's news article this morning, that being some ill-informed people have used a program in a way that it wasn't intended to be used and are now claiming there is some kind of problem.

 

yes it was the same PCPER website using FCAT..... the same one which is in contact with NV and printed NV`s corporate PR reply and the green team fanboys love it.

and yes I do have an NV card (and prefer it to AMD) . im just rather stunned by the hypocrisy = the same website which forced AMD to sort crossfire drivers are being ignored and berated when they find an issue (which potentially could be aweful)


https://forums.geforce.com/default/...tx-970-3-5gb-vram-issue/post/4432672/#4432672

^^ official NV response

higher priority = lower latencey

 

New article on pcper

NVIDIA Discloses Full Memory Structure and Limitations of GTX 970

 

anandtech article on it is interesting read.

Seemed odd regarding the 13/16 sm split relating to a 3.5/4 GB split. But turns out nvidia disabled 1/8 of the rope and l2 cache. Moreover, it can use 3.5gb or 512 mb segment, but not at same time. So would have impact on performance.

Will be interesting to see what FCAT results produce, but still nothing on pcper website

 

worst case - if you have textures needed by the game stored in the 500mb part of the ram , and they get accessed , the entire card would have to wait for it to be accessed @ 1/7th the performance - hence potential stuttering oO

 

The anandtech article looks detailed

Probably wont get any proper tests from the big sites until later this week

 

forum tests on similar games on R9`s coming out showing uniform performance drop as the detail/ resolution scales.

 

In all honesty, I see a great deal of bias too, but some of it is from yourself. Why are you so active in the thread? Do you own a 970?

In the interest of being non-bias however, the video you posted does raise a question: If that's the same rig, then why is the 980 accessing all of it's RAM and the 970 is not?

 

from anandtech:

IF they are correct and the 512MB is alarmingly slower - that would explain the increased CPU useage as its trying to do a lot more work...


would make for interesting FCAT.

 

So please provide us with the FCAT testing PCPER have done on the 670, AFAIK so far all we have to go on is the widely discredited Nai's benchmark and some dodgy YouTube videos.

 

widely discredited? go read anandtech and PCPER - they are all saying its THE reason this was discovered and not some game bugs or dodgy drivers!

 

but his benchmark is correct, when reading the first 3.5GB its showing 150GB/s and then for the final 0.5GBits down to 22.35GB/s ~ 150 / 7 = 21.42

therefore its showing the speeds when the card stops reading from the other chips and linked memory controller and goes for the single read of single chip though 1 memory controller.

 

From what i have read since i posted that, they say that with competent OS heuristics the performance difference should be negligible in real-world gaming scenarios.

No it's not.
It using CUDA, it is not using the DirextX or OpenGL memory management.
If you get into a situation like the following...

You have more to worry about than a 0.5GB portion if you have a game that doesn't have 0.5GB of data that is used less frequently or in a less latency-dependent.

 

and didn't AMD say something similar regarding the Bulldozer and core scheduling