970 memory allocation issue revisited

yes but It is 4GB or 8GB version?

 

If you want to answer the question of whether or not an issue really does exist, we need to involve as many people as are willing to participate. We also need to be open to the opinions of others.

That tends to happen when you undertake such a large project on a forum.

Everyone should be using the latest release.

Unfortunately, no. There's no way for us to really be certain of anything at this point. We could simply be proving a flaw exists in CUDA itself and not the cards showing erratic behaviour. At this point, we really don't know.

Wrong. We need to see how the cards react with and without being the primary display adapter.

Wrong again. We need to know exactly how these cards react on different platforms.

Being able to code, doesn't really help if you're not familiar with the language/headers being used. Fox is just trying to learn as we go along.

I'm familiar with C++, Pascal, Javascript, VBScript, Basic and Visual Basic (pre-dot net). While I'm good with the code itself and understanding what the code appears to be trying to do, not knowing the header files being used causes me some issues. I haven't done any real coding since 2001.

Unfortunately, when coding using API's, you're at the mercy of the API and any flaws it may (or may not) have.

 

Very odd indeed.

High chance a VBIOS or driver update can fix the issue then.

Hmm, I didn't consider that possibility for that image.

 

there may be something in this, when I have ultra textures enabled on inquisition the game will sometimes hit 3.5gig usage and then I get stutters. After a while it clicked in my head it seemed to trigger when a certian vram usage got hit.

 

'tis the 4GB version in that picture (the image came from a review of my model of laptop by a member on another forum).

 

It looks like GTX780 suffers at least PART of the issue we're seeing with GTX970.

Now we're getting the kind of information we need. Now we need full GK104 and cutdown GK104.

 

@sykozis

He didn't actually give an opinion, and opinions don't solve anything that requires hard data, but let's move on from there.

First point, I see people running what appears to be different versions.

Second point, then we have no way of knowing if the results are even relevant.

Third point, is that while I agree with you, the fact that people CAN run it wither way without knowing whether they are supposed to or not and posting one or the other result without stating whether they ran it headless or not, or without confirmation that it was run headless or not, makes the results a mess and nigh on impossible to organize.

Fourth point, refer to my third. Standardization of data collecting must be used to acquire useful data. Not standard, not useful. This is basic troubleshooting/data comparison practice.

Lastly, I meant plainly that if I could write a standardized, verifiable, data collecting benchmark that did what this issue needs, I would.

But I can't.

 

I see some people posting results with the last few tests failing. It does that. It can also crash your video driver and cause flickering. Yay CUDA.

The old program did this as well, the difference with my version is that it actually notifies you if not all tests completed succesfully. Most of the times you can just run it again, don't touch anything until it's finished and it'll work. It's best to run this with SLI disabled, not having any other programs opened and using the Windows Classic Theme or when dwm.exe is not running.
The last part really influences the result. On the 970GTX it'll still look bad, while other cards may improve or don't show no slow down at all anymore.

 

Uhhh... well my 980M gets around the ~130GB/sec mark in that benchmark, despite its official bandwidth being advertised as 160GB/sec (and 'effective bandwidth' after Maxwell 2.0 3rd gen delta colour compression even higher still). So either all cut down NVIDIA GPUs are utterly borked or those metrics aren't really related to the issue at hand.

I too get micro-stutters in DAI too due to it being unable to run in fullscreen mode. I'm not sure if it is the best game to use to illustrate the issue as it still has numerous PC issues remaining after the 3rd patch (the lack of fullscreen being by far the biggest offender).

 

I doubt, it looks more like either win7 issue or driver allocation issue or this app benchmark issue..



I get drop @ 2.5gb too...

But if i run in headless mode, unplug main gpu cable and use igpu connection it doesn't have any drop.

DWM runs in background.


So all of you running in normal mode or both gpus active & connected to display - it will do that, GL running in circles :infinity:

 

780 TI test

Nai's Benchmark, edited by VultureX
Device: GeForce GTX 780 Ti (3.00 GB)
Memory Bus Width (bits): 384
Peak Theoretical DRAM Bandwidth (GB/s): 336.000000

Allocating Memory . . .
Chunk Size: 128 MiByte
Allocated 22 Chunks
Allocated 2816 MiByte
Benchmarking DRAM
DRAM-Bandwidth of Chunk no. 0 (0 MiByte to 128 MiByte):281.65 GByte/s
DRAM-Bandwidth of Chunk no. 1 (128 MiByte to 256 MiByte):287.00 GByte/s
DRAM-Bandwidth of Chunk no. 2 (256 MiByte to 384 MiByte):288.21 GByte/s
DRAM-Bandwidth of Chunk no. 3 (384 MiByte to 512 MiByte):287.00 GByte/s
DRAM-Bandwidth of Chunk no. 4 (512 MiByte to 640 MiByte):286.09 GByte/s
DRAM-Bandwidth of Chunk no. 5 (640 MiByte to 768 MiByte):286.05 GByte/s
DRAM-Bandwidth of Chunk no. 6 (768 MiByte to 896 MiByte):286.60 GByte/s
DRAM-Bandwidth of Chunk no. 7 (896 MiByte to 1024 MiByte):287.06 GByte/s
DRAM-Bandwidth of Chunk no. 8 (1024 MiByte to 1152 MiByte):287.64 GByte/s
DRAM-Bandwidth of Chunk no. 9 (1152 MiByte to 1280 MiByte):287.40 GByte/s
DRAM-Bandwidth of Chunk no. 10 (1280 MiByte to 1408 MiByte):292.11 GByte/s
DRAM-Bandwidth of Chunk no. 11 (1408 MiByte to 1536 MiByte):293.44 GByte/s
DRAM-Bandwidth of Chunk no. 12 (1536 MiByte to 1664 MiByte):294.25 GByte/s
DRAM-Bandwidth of Chunk no. 13 (1664 MiByte to 1792 MiByte):294.36 GByte/s
DRAM-Bandwidth of Chunk no. 14 (1792 MiByte to 1920 MiByte):294.38 GByte/s
DRAM-Bandwidth of Chunk no. 15 (1920 MiByte to 2048 MiByte):294.22 GByte/s
DRAM-Bandwidth of Chunk no. 16 (2048 MiByte to 2176 MiByte):294.15 GByte/s
DRAM-Bandwidth of Chunk no. 17 (2176 MiByte to 2304 MiByte):293.62 GByte/s
DRAM-Bandwidth of Chunk no. 18 (2304 MiByte to 2432 MiByte):295.11 GByte/s
DRAM-Bandwidth of Chunk no. 19 (2432 MiByte to 2560 MiByte):295.07 GByte/s
DRAM-Bandwidth of Chunk no. 20 (2560 MiByte to 2688 MiByte):294.48 GByte/s
DRAM-Bandwidth of Chunk no. 21 (2688 MiByte to 2816 MiByte):16.67 GByte/s
Benchmarking L2-Cache
L2-Cache-Bandwidth of Chunk no. 0 (0 MiByte to 128 MiByte):575.76 GByte/s
L2-Cache-Bandwidth of Chunk no. 1 (128 MiByte to 256 MiByte):575.75 GByte/s
L2-Cache-Bandwidth of Chunk no. 2 (256 MiByte to 384 MiByte):576.31 GByte/s
L2-Cache-Bandwidth of Chunk no. 3 (384 MiByte to 512 MiByte):576.13 GByte/s
L2-Cache-Bandwidth of Chunk no. 4 (512 MiByte to 640 MiByte):575.76 GByte/s
L2-Cache-Bandwidth of Chunk no. 5 (640 MiByte to 768 MiByte):576.72 GByte/s
L2-Cache-Bandwidth of Chunk no. 6 (768 MiByte to 896 MiByte):575.84 GByte/s
L2-Cache-Bandwidth of Chunk no. 7 (896 MiByte to 1024 MiByte):576.70 GByte/s
L2-Cache-Bandwidth of Chunk no. 8 (1024 MiByte to 1152 MiByte):575.89 GByte/s
L2-Cache-Bandwidth of Chunk no. 9 (1152 MiByte to 1280 MiByte):575.88 GByte/s
L2-Cache-Bandwidth of Chunk no. 10 (1280 MiByte to 1408 MiByte):576.60 GByte/s
L2-Cache-Bandwidth of Chunk no. 11 (1408 MiByte to 1536 MiByte):575.69 GByte/s
L2-Cache-Bandwidth of Chunk no. 12 (1536 MiByte to 1664 MiByte):575.84 GByte/s
L2-Cache-Bandwidth of Chunk no. 13 (1664 MiByte to 1792 MiByte):577.01 GByte/s
L2-Cache-Bandwidth of Chunk no. 14 (1792 MiByte to 1920 MiByte):575.75 GByte/s
L2-Cache-Bandwidth of Chunk no. 15 (1920 MiByte to 2048 MiByte):575.78 GByte/s
L2-Cache-Bandwidth of Chunk no. 16 (2048 MiByte to 2176 MiByte):575.72 GByte/s
L2-Cache-Bandwidth of Chunk no. 17 (2176 MiByte to 2304 MiByte):575.87 GByte/s
L2-Cache-Bandwidth of Chunk no. 18 (2304 MiByte to 2432 MiByte):576.48 GByte/s
L2-Cache-Bandwidth of Chunk no. 19 (2432 MiByte to 2560 MiByte):576.31 GByte/s
L2-Cache-Bandwidth of Chunk no. 20 (2560 MiByte to 2688 MiByte):576.19 GByte/s
L2-Cache-Bandwidth of Chunk no. 21 (2688 MiByte to 2816 MiByte):16.99 GByte/s
Press any key to continue . . .

 

I don't trust anything right now. The random game tests popping up are forgetting how VRAM is dynamically allocated these days. I remember my 1GB 5850s handling games at 1250MB usage and not exhibiting performance drops. So either there is nothing wrong, or many just don't know what to look for.. or obviously every scenario isn't going to show the symptoms of bandwidth drops.

 

without driver instaled show 3gb memory size its normal? and with driver show 4gb. 1gb is fake?
this screen without driver ir real memory size gtx970?

 

GTX 670 2GB version.
Nal's and VultureX test:

Spoiler

 

And that may (or may not) be an issue in itself. It would be best if everyone was running the same version so we can compare the outputs from only 1 version.

Now you see the problem we have here and why we need results from previous generations. We need to valid the results as best we can, which means using older generations and seeing exactly how they react. That's the best we can do at this point.

We really need people to run it in multiple configurations and post the results, while letting us know what configuration they used. Win7 users need to run with DWM disabled. All users need to run the test both with the display adapter active and in headless mode. We need to know which result is which and what platform they're on, whether it's Win7, Win8.x or Win10.

If you want valid data, you have to consider all operating conditions and test those conditions. Doing otherwise, invalidates your results.

If there was a way to do that, it would be great. Unfortunately, using software, it's impossible to write such a benchmark and not leave some degree of doubt. Software code will never be flawless as humans are not without flaws. Since humans create software, software will have errors.

 

That looks about right with Aero left enabled.

 

Yeah after looking around some more, this whole accusation seems really sketchy. People are consistently pointing to Shadow of Mordor with Ultra textures as the smoking gun, but Dragon Age and Watchdogs run a 4K with heavy AA apparently don't have issues...

 

You'll see that with Nai's benchmark the last two lines have really weird performance numbers 1.J#? This is actually, because the kernel launch failed as my version indicates.
Read this please to get accurate results:

 

Plus there is one user (GTX 980M 8GB) here who posted his result and it doesn't exhibit the issue.

EDIT: Anyone with GTX970M 6GB willing to run the bench?

 

You do realize, that most importantly, the 970 and 980 were sold as having the same memory systems, the same bandwidth and the same capacity.

The most important test is a direct comparison between the two by the same person(s) on an otherwise same rig.

Finding out if previous generations were gimped in a similar fashion is interesting, and can show a pattern, but for the sake of the current issue at hand, the only comparison that matters is 970 to 980, as again, they were sold as having identical RAM systems.