WCCftech: Memory allocation problem with GTX 970 [UPDATE] PCPer: NVidia response

[Erenhardt]

from a neogaf thread
http://www.neogaf.com/forum/showthread.php?p=148763930#post148763930

Gotta love the amount of denial going from owners

 

[Spanners]

Thanks to the OP, there seems to be something to this.
ManuelG

 

[iiiankiii]

I hope this isn't hardware based. I have a GTX 970 in one of my rig. I would hate to get rid of it.

 

[amenx]

My GTX970 works fine!

So does mine, no matter what I've thrown at it. It seems 780ti also does poorly on its last 200mb of its 3gb vram in Nai's bandwidth benchmark.

 

[ShintaiDK]

Well it turns out GTX970 isnt 256bit its actually 208bit card
http://www.overclock.net/t/1535502/gtx-970s-can-only-use-3-5gb-of-4gb-vram-issue/170

One is tested with Aero, one is not. Stop comparing apples and oranges.

 

[Carfax83]

The last 300 or so MBs is Windows hardware acceleration. You need to run the test headless or with Windows running on the iGP for it to be accurate; Windows uses about 300MB of VRAM.

Which is why that 980 starts screwing up early.

Here's the problem I see with this. The primary contention the OP first came up with said that the GTX 970 couldn't utilize the full 4GB. In the second post of this thread, cmdrdredd easily disproved that by showing Crysis 3 running at 4K with 8x MSAA using 4GB of VRAM.

Then the inevitable goal post shift, where now the GTX 970's L2 cache and VRAM bandwidth both begin to drop off past a certain point of VRAM usage, which according to Nai's benchmark, seems to be happening at 3.2 GB.

However, plenty of people including myself play games where the GTX 970 is hitting well over 3.5 GB of VRAM, with no performance issues at all.

I just played AC Unity, and according to MSI Afterburner, the max VRAM usage was slightly over 3.7 GB, and this is at 60 FPS 1440p max settings. If Nai's benchmarks really did show a bandwidth decrease past a certain point in VRAM usage, then wouldn't my frame rate plummet as a result?

Here's a screenshot from NeoGaf where a user shows SoM using 3.6 GB of VRAM with very playable frame rates:

 

[ShintaiDK]

Gotta love the amount of denial going from owners

You can try run it with the same conditions with an AMD card. And you funna enough experience the same thing.

 

[Spanners]

One is tested with Aero, one is not. Stop comparing apples and oranges.

Nope. "Because this application runs in a window you have to either kill DWM or use your iGPU as primary GPU. Otherwise you possibly won't get the best result because desktop does use some VRAM."

This is from the guy who posted them so I'd assume he was running them correctly.

 

[HurleyBird]

While I'm initially sceptical of the results, I'm totally not surprised to see the usual Nvidia defense force members in full denial mode. The issue needs more investigation and less debating.

 

[ShintaiDK]

Nope. "Because this application runs in a window you have to either kill DWM or use your iGPU as primary GPU. Otherwise you possibly won't get the best result because desktop does use some VRAM."

Lets look on an example. I have run these within ~1 minute of one another on the same system.

 

[Gloomy]

You have a 980. What are you trying to prove?

 

[Carfax83]

While I'm initially sceptical of the results, I'm totally not surprised to see the usual Nvidia defense force members in full denial mode. The issue needs more investigation and less debating.

It's easy to deny, because the OP's premise has already been disproven. GTX 970 has no issues using 4GB of VRAM.

Then he changed the argument to "can't fully utilize the bandwidth," and now this is being shown to be faulty as well.

 

[ShintaiDK]

While I'm initially sceptical of the results, I'm totally not surprised to see the usual Nvidia defense force members in full denial mode. The issue needs more investigation and less debating.

Thanks for your "contribution".

 

[ShintaiDK]

You have a 980. What are you trying to prove?

That the entire thread is build upon BS with apples and oranges compares at best.

 

[Keysplayr]

While I'm initially sceptical of the results, I'm totally not surprised to see the usual Nvidia defense force members in full denial mode. The issue needs more investigation and less debating.

Why no comments about the attack force? I'm not saying I have enough data to make any judgement about this memory usage, but your comment was kind of one sided.

Seems to me, somebody started a rumor by showing specific findings in certain situations showing a memory deficit? Why? Who knows. Others ran with the ball. Now here we are.

I think we should conduct some tests of our own. What do you all think? And btw, the second the GTX970 uses more than 3.5GB in any situation shows that it can use more than 3.5GB and the claims that it cant, or the claims that it is less than a 256-bit interface gets flushed.

(Now "people" think I have results all ready to go showing exactly that and are working out excuses ahead of time for why my results are inconclusive.)

So I won't. I'll leave it up to our neutral forum members. I have a bias toward Nvidia products. I think they're better all around. I think you all know this. So I'll ask anybody here that isn't a usual suspect in the AMD vs. Nvidia war continuum to take their GTX 970's and 980s and their 290 and 290Xs and run the same tests and lets see where we stand.

This should be good.

 

[Spanners]

Lets look on an example. I have run these within ~1 minute of one another on the same system.

I'm not sure what your example proves? I wasn't saying "nope" Aero makes no difference, I was quoting the person who posted the benches to show they knew the correct way to run them. Edited my post to be more clear.

that GTX970 is not 256bit card.
Thats i think is why GTX970 have some problems utilizing full 4096 Vram in games.

It's easy to deny, because the OP's premise has already been disproven. GTX 970 has no issues using 4GB of VRAM.

Then he changed the argument to "can't fully utilize the bandwidth," and now this is being shown to be faulty as well.

Those seem to be his main points which haven't been edited for the last two weeks. If the cards bandwidth performance is due to not having full access to the entire bus width then his bolded statement may be true.

If these synthetic tests turn out to be representative then I think "have some problems utilizing full 4096 Vram in games." might be an accurate statement as well.

I don't see any goalpost shifting.

 

[Keysplayr]

I'm not sure what your example proves? I wasn't saying "nope" Aero makes no difference, I was quoting the person who posted the benches to show they knew the correct way to run them. Edited my post to be more clear.





Those seem to be his main points which haven't been edited for the last two weeks. If the cards bandwidth performance is due to not having full access to the entire bus width then his bolded statement may be true.

If these synthetic tests turn out to be representative then I think "have some problems utilizing full 4096 Vram in games." might be an accurate statement as well.

I don't see any goalpost shifting.

To be fair, it "looked" like you said "nope" as if to say Aero makes no difference. But thanks for clearing that up.

Next, I think it has been shown that 970 can and does use more than 3.5GB of VRAM which would indicate it indeed does have a 256-bit memory bus.

And how is this.....

"It's easy to deny, because the OP's premise has already been disproven. GTX 970 has no issues using 4GB of VRAM.

Then he changed the argument to "can't fully utilize the bandwidth," and now this is being shown to be faulty as well."

...not shifting goalposts?

 

[Gloomy]

I'd be happy to run tests on my R9 290, but this little app is written in CUDA.

AMD's driver has no issues allocating 4GB though. In fact this kind of setup that Nvidia is using wouldn't even work, since Mantle games manage their own VRAM directly (sometimes to their own detriment, lol).

 

[iiiankiii]

Here's my result. Drop off at 3200mb. Ran with igpu as primary with no display attached to the 970. Afterburner stated the gtx 970 was using 200mb before the test was ran. But GPUZ stated the 970 was using 0mb dynamic 10mb dedicated. I'm not sure how to get rid of the overhead.

 

[ShintaiDK]

Here's my result. Drop off at 3200mb. Ran with igpu as primary. Afterburner stated the gtx 970 was using 200mb before the test was ran. I'm not sure how to get rid of the overhead.

Close everything else. Sidebar, aero, apps etc.

 

[Carfax83]

Those seem to be his main points which haven't been edited for the last two weeks. If the cards bandwidth performance is due to not having full access to the entire bus width then his bolded statement may be true.

If these synthetic tests turn out to be representative then I think "have some problems utilizing full 4096 Vram in games." might be an accurate statement as well.

I don't see any goalpost shifting.

But it isn't true. We have solid proof of the GTX 970 using 4GB of VRAM, and we also have solid proof of gaming whilst using high amounts of VRAM (over 3.2 GB) without any of the associated performance drops you would expect from a bandwidth reduction if it were occurring..

Then to top it off, you have the GTX 980 showing the same "problem." That benchmark is bogus, and this whole issue is bogus if you ask me.

It just doesn't add up. If these problems were real, then it would have been discovered during the reviews where the cards were subjected to bandwidth tests and 4K gaming, which is VRAM intensive.

 

[Spanners]

Nvm.

 

[iiiankiii]

Close everything else. Sidebar, aero, apps etc.

system is running windows 8.1. Nothing else opened. No sidebars or apps.

 

[Carfax83]

system is running windows 8.1. Nothing else opened. No sidebars or apps.

He may be saying you need to disable the visual effects as well..

 

[sontin]

Hm, has somebody a GTX980m?!
This card has the full 256bit interface but only 12SMM. If there is a hardware problem the GTX980 would only use 75% of the memory with full speed - so around 3GB/6GB.