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

I'm going to reserve judgement until more information is available, specifically what Nvidia says on the matter once they have done some testing, but it does seem something odd is happening.

A blog post from Scott Wasson from the Tech Report shows there is a difference between the 970 and 980, but like the benchmark program people are using to test their cards it seem to only exhibit it's self in certain scenarios.

 

I thought sleeping dogs with high res texture pack would do it but I only hit 1.7gb, cod advanced warfare maxed out at 2560x1440 hit 3.5gb but I did not notice any slow down, not sure what I have that would push the ram up to 4gb

 

Question: Has anyone run this Nai's memory benchmark program for themselves ?

I only ask because i just ran it on a 670 and got the same results, although I'm guessing like many people i didn't run it on my GPU in headless mode.

 

Given the 970 uses the same (albeit slightly gimped) GPU as the 980, and they have the same memory bus, why isn't this evident with 980s?

And this on some unproven benchmark?

 

NVIDIA Responds to GTX 970 3.5GB Memory Issue. pcper

 

In laymans terms does that mean it is OK tp buy a 970 then?

 

I've ran the rec.exe file same result as OP's article screenshot. But it's clearly a CUDA benchmark tool, not real world usage.

Interesting you are getting the same result, also on a SM cut-down GPU.

I suspect the reason is that the benchmark tool allocates one SM cluster to one part of memory, so looking at the screenshot, it means half way into 26th cluster things go pear shape for 970. Considering 970 is cut down from 980's 32 to 26 SM clusters, it supports my theory. The PCPer article also supports this theory to some degree: the first 26/32 memory are directly allocated to video memory whereas the last few hundred MB are distributed between SM's. Assumption is that the memory controller is very simple, a set amount (128MB) is dynamically allocated to each SM, or only in this benchmark.

Basically I'm saying the benchmark is trying to use disabled SM clusters to test performance. Unless we can test using an actual game or nVidia/benchmark show us their design/code, it's impossible to know what's happening with 1 benchmark.

 

Yes. It turns out that, unless Nvidia's benchmarking is somehow wildly incorrect, the limited bandwidth of the reserved portion has about a 1% effect on framerate. Games that don;t use more than 3.5GB will not be affected in the slightest.

If you;re overly cautious, wait for one or more third parties to run additional tests. But so far nobody has shown any case where this has a noticeable impact other than one specific CUDA benchmark, so it's unlikely that after several months of general release nobody would realise until now the there was a severe real-world performance impact.

 

Thanks. As I am still in the process of getting the cash together it is likely to be another two weeks or so before I am ready to purchase. Hopefully by then we will have a definitive answer.

 

So all the fuss was just internet BS with a bit of nerd rage thrown in then. Thought as much, and also thought for NVidia to release something that would have been so fundamentally flawed was not very likely... I'm sure they learned that lesson already..

 

I'm talking about playing the game. Actually playing the game, during combat, with a lot of enemies around brought a bit of stutter here and there. It all cleared up when I started playing with the texture setting on the next peg down. Nothing else was reduced.

Now, maybe it is an AMD issue? I can't be sure as I don't have an equivalent Nvidia card, but it certainly pretty much fixed it for me. I've not spent a lot of time evaluating whether other people had the same issue. I did briefly read something to that effect whilst searching for a fix to a bug or two that would occur - that weird mouse one where you can't seem to do anything - and had seem comments about frame rate peculiarities regarding swapping from main memory to gpu.

Other games have been generally very smooth with applicable settings. Other games though are not maxing vram usage currently that I've noticed. I don't think we can expect high end gaming not to use all that vram going forward given everything from larger resolutions to the current generation consoles pushing higher res textures themselves.

Yes. Shadow of Mordor will easily use more than 3GB of memory. On my card I see reported usage on afterburner over 4000MB. I suspect though that frame time variance is what we want to measure though, and I don't know how to do that... something about fcat.

I am aware though that memory usage and afterburner has been seen to be not the best indicator of actual memory usage.

 

so if NVidia says its ok then it must be?

 

doesn't matter - that's false advertising - you pay for 4GB and expect it to work properly

 

It;s got 4GB vRAM, you can access all 4 GB.

 

No.. I mean if they are to be believed (easily provable) then in reality, when gaming at high resolutions, it seems to not really cause an issue at all.

 

I was alarmed initially as I have two of these cards running @ 4k, however it sounds like a non-issue aside from a poorly optimized benchmark.

 

Im the same sli msi versions with 4k gaming I see no problems. Im keeping mine

 

https://forums.geforce.com/default/topic/803518/geforce-900-series/gtx-970-3-5gb-vram-issue/

verifiable reports of this happening in a number of games , includinf farcry 4 and unity to name 2. that's from an NVidia rep.

 

I still have another couple of weeks or so before I hhave the necessary funds, shoukd I go ahead and fling 300 quid at a 970?

 

Just wait until you have the money, mate. See what the situation is then.