Discussion in 'Hardware' started by lancer778544, 23 Jan 2015.
More here. Benchmark/VRam tester download here.
Yes it popped up in my thread. http://forums.bit-tech.net/showthread.php?t=281118 It is a bit worrying as I was close to flinging 300 quid at a 970. Now I am reluctant.
I've been hearing about this today but how much does it happen in the real world?
How often are people maxing it out to 4gb, I don't think I've seen anything happen yet with mine, not saying I'm happy about it but is this really going to affect me?
Ah crap, I should have known it would have been posted already. Oh well.
A lot of people are having issues though and if it turns out to be a hardware fault then it could prove very expensive for Nvidia. Lets hope they can get to the bottom of this sooner rather than later.
The cases where it is claimed to have an impact on games (generally at 4K+ and/or with DSR at 4K+) are where you're also hitting all sorts of other GPU bottlenecks. It's very difficult to tell if any real-world performance issues are down to having a few hundred mb vRAM running at a slower speed, or whether the bottleneck is elsewhere and having that extra section of vRAM access faster would make zero difference.
There are some rumours that the issue is down to uneven SMM binning for the 970's GM204 is starving the memory controller in one GPC, but the one benchmark that reliably shows the data rate drop issue is not particularly compute intensive (it's mainly shunting things in and out of vRAM as fast as possible).
If it turns out to be a hardware fault, surely they'll have to issue a recall/replacement?
I've had no problems with mine yet, but I've not really pushed it that hard.
Official word from Nvidia would be good. Same as above, I have not noticed any issue personally.
If true, I is a "not good enough" issue to be sure that warrants resolution. Games are certainly approaching that sort of vram usage. When I play Sharp of Mordor for instance, my vram usage is at 99%+ with presumed swapping from main ram.
As usual there is nothing yet though they must be aware of the issue by now?
Do you noticed any perceivable performance drops when vram usage gets that high? I don't much trust the data being shown in this benchmark as it's not representative of real world usage. I don't run a res high enough to push memory usage anywhere near that high, so probably wont run in to any problems.
Word is they 'are looking into it'.
This mirrors my thinking on the matter. As far as I can make out, it's only in one specific test that it does this. However it may be of concern for those who use CUDA in earnest, for example.
Can we stop this thread right here and look at if there is any truth at all to this outside of what 'some men on the internet' said.
What I'm after is confirmation from;
a) A well respected reviewer (Techpowerup, Bit Tech, Tech Report, Anandtech, Guru3d etc)
b) Nvidia themselves
c) An impartial third party such as a developer.
The dark arts of GPU memory allocation are far more complicated than the gun jumpers probably think. It might be as simple as Nvidias driver engineers reserving off 500mb for Shadowplay or something. All I'm saying is we all seem to have accepted this for fact, when that hasn't yet been sufficiently proven out.
on another forum , GTX 970 users are running kombuster right up the card headline ram limit with no slow downs , and the comments are along the lines of ` its a buggy benchmark`
If that were just the case why nothing from Nvidia?
I had to turn the textures down to the one below ultra as there were noticeable spikes in large fights. Enough to be irritating. Put it down and it seemed smooth 99% of the time.
That's still only with that weird benchmark tool. People have run it on all sorts of cards and it shows a slump for that last little chunk, including cards such as the 680/770 which have a full fat chip without any parts disabled.
In all the reviews I've read of the 970's, I've not seen the gap between 970 and 980 skew wildly when bumped up to 4k res, where vram would make the most difference.
Because exhaustive verifiable testing takes time, and it's the weekend.
Look at it this way: someone runs a compute benchmark and uses the results to claim that something is odd with memory bandwidth, people respond with (anecdotal, not once have I seen someone actually graph frame render times vs. vRAM allocation) reports that 'using lots of ram at extreme resolutions/settings makes the game stutter!'.
a) Jump to the conclusion that a serious hardware bug crept past all QC testing and authorise overtime for the engineering tem (pulling them off whatever they are working on) to find it
b) Actually do the work to verify the bug even exists
Personally, I'm going to wait until someone writes s a verifiable benchmark that is actually designed to test vRAM bandwidth (rather than using a benchmark designed to test something else entirely), and/or for someone with access to dedicated FCAT equipment to test frame render times in comparison to reported vRAM usage to see if there is any verifiable correlation.
Why don't we test it for yourselves? Anyone know a game that will easily use close to 4GB of VRAM? I've not experienced over 3GB yet, so I hadn't seen this problem during gaming.
Separate names with a comma.