GTX 970: 3.3GB?! Whaa?!

[Elapo]

I read someone's post in which they mention the Nvidia drivers making the 970 NOT use anything over 3.5 gigs now in games to alleviate the problem, is that true or no? Don't shoot the messenger.

 

yes. Anyone who claims the .5GB is making their game lag is full of shit. My 970 goes up to 3.5GB, never higher, but the drivers do load stuff into the remaining .5GB. 

 

It only impacts tests that were made to fill up VRAM, it does NOT impact real world applications (such as games.)

[techtoday]

Sorry for the late reply. http://www.extremetech.com/extreme/198214-198214is our source here. Is there really a memory bug?! Judging by what you guys are saying there does not seem to be.

[Zulkkis]

yes. Anyone who claims the .5GB is making their game lag is full of shit. My 970 goes up to 3.5GB, never higher, but the drivers do load stuff into the remaining .5GB. 

 

It only impacts tests that were made to fill up VRAM, it does NOT impact real world applications (such as games.)

 

The cases are narrow but it can pose a bottleneck. The bigger question was nVidia having months and months of time selling the product with false specs and only coming out to correct it when the users found out. The whole debacle is more about ethics than it is about the real-world performance implications.

 

It mostly just stops caching anything above the limit. The drivers won't do much magic, and game devs are not going to accommodate for a single card on the market. nVidia can do a thing or two there but I doubt they'll go much beyond optimizing a few major titles. The performance just tanks above ~3.5GB, people have tried rendering with the thing (it is a market they are selling to, otherwise it wouldn't be bundled into the GTX cards) and the impact is noticeable.

 

Thinking it as a 3.5GB card is the simplest way to think about the implications of the whole design. There aren't many cases where you would go above 3.5GB but still stay under 4GB. It doesn't really make the card much less unappealing unless you think about the ethics, so there is no reason to exaggerate it, but people who still go on about telling that the card is "4GB" or that it doesn't have any impact at all are also being dishonest or just not too well informed.

[Hieb]

The performance issues some people have from GTX 970s comes from the fact that the last two 512MB modules share a single 'pipeline' (so to speak), whereas every other 512MB pool of VRAM gets its own memory controller, L2 cache, etc. So basically the last gigabyte of VRAM on the GTX 970 shares a pipeline, so when you go over 3.5GB the last gigabyte loses some performance as it can't get the full bandwidth out of both pools of VRAM simultaneously.

 

Now I'm not 100% sure on this, but I believe driver updates have adjusted video memory allocation for the GTX 970 to reduce the issue.

[Elapo]

The cases are narrow but it can pose a bottleneck. The bigger question was nVidia having months and months of time selling the product with false specs and only coming out to correct it when the users found out. The whole debacle is more about ethics than it is about the real-world performance implications.

 

It mostly just stops caching anything above the limit. The drivers won't do much magic, and game devs are not going to accommodate for a single card on the market. nVidia can do a thing or two there but I doubt they'll go much beyond optimizing a few major titles. The performance just tanks above ~3.5GB, people have tried rendering with the thing (it is a market they are selling to, otherwise it wouldn't be bundled into the GTX cards) and the impact is noticeable.

 

Thinking it as a 3.5GB card is the simplest way to think about the implications of the whole design. There aren't many cases where you would go above 3.5GB but still stay under 4GB. It doesn't really make the card much less unappealing unless you think about the ethics, so there is no reason to exaggerate it, but people who still go on about telling that the card is "4GB" or that it doesn't have any impact at all are also being dishonest or just not too well informed.

 

It has no impact on performance. Ever. There's not a single game that ever uses over 3.5GB vram, believe me, I've tried (in multiple games, using DSR and the likes). Usage won't pass that 3.5 GB barrier, only specific VRAM benchmarks are able to access that remaining .5GB of memory, which is how people found out in the first place. Even rendering using iRay works just fine for me.

 

I know what all the fuzz was about, people being surprised a company lied to them. Like that never happened before.

[R_Golden]

No need for that. Especially when Nvidia acknowledged the issue. You're just trying to sway the discussion.

No need for that? What? Until you've tested the card in real world use you have no idea how this "issue" actually affects the card.

[BirdyTheMighty]

No, it's very usable. I would know.

 

Have you used one before?

in skyrim yes bought the 970 cause of the "4GBVRAM" boy wrong choice 

[brokencross]

No need for that? What? Until you've tested the card in real world use you have no idea how this "issue" actually affects the card.

The card has already been tested in real world situations and it performed awfully when the last .5GB cluster was used.

Pull your heads out of your asses and quit being smug every time someone states a well proven fact.

I've never said the 970 was a bad card, it's actually a great card as long the last cluster isn't used, which pretty much excludes any type of 4K gaming and even some 1440p @ ultra. Also some ultra hd texture mods.

[Majestic]

That's because you aren't running the test in headless mode. Windows snoops memory aswell.

 

/thread.

[zappian]

It has 4 GB of VRAM but 500 mb are on a small bit bus ( something like 24 bits) .
So the actual bandwidth on that section is significantly lower .

Doesn't impact you unless you want to play games at 4k with something like gtx 970 SLI .

Thats a really bad idea.

For other uses the gtx 970 is still an amazing card , don't let that shake your opinion on the card.

Its the best card at the price point .

[typographie]

The card has already been tested in real world situations and it performed awfully when the last .5GB cluster was used.

 

If you can link some reputable benchmarks actually showing something like that, I don't know why you're holding back. Actual real-world gaming benchmarks showing that are both the Holy Grail and Loch Ness Monster of this topic.

 

For you, OP, or anyone else who actually wants to read some real information on this that didn't come from a fellow forum poster:

http://www.anandtech.com/show/8935/geforce-gtx-970-correcting-the-specs-exploring-memory-allocation

[Majestic]

 

Or anyone else who actually wants to read some real information on this that didn't come from a fellow forum poster:

http://www.anandtech.com/show/8935/geforce-gtx-970-correcting-the-specs-exploring-memory-allocation

 

I think everyone, including you, misunderstood the topic. He noticed an incline at 3.3GB instead of the proclaimed 3.5GB. My guess is he ran the Nai benchmark in desktopmode and saw the performance hit before 3.5GB and failed to understand that 200-300MB was already taken up by windows.

 

Hence I said;

That's because you aren't running the test in headless mode. Windows snoops memory aswell.

 

/thread.

 

This entire topic has been a giant red herring.

[brokencross]

If you can link some reputable benchmarks actually showing something like that, I don't know why you're holding back. Actual real-world gaming benchmarks showing that are both the Holy Grail and Loch Ness Monster of this topic.

 

For you, OP, or anyone else who actually wants to read some real information on this that didn't come from a fellow forum poster:

http://www.anandtech.com/show/8935/geforce-gtx-970-correcting-the-specs-exploring-memory-allocation

Not going to google that for you. There are lots of benchmarks from sites like anandtech and such.

[typographie]

Not going to google that for you. There are lots of benchmarks from sites like anandtech and such.

 

http://www.anandtech.com/SearchResults?q=GTX+970

Well, here's the search results for 'GTX 970' on Anandtech. In order of appearance, I see:

1.) the article I linked in my last post, cautiously concluding that the memory partitioning was unlikely to be a material problem;

2.) their original news article reporting the existence of the memory discrepancy; and

3.) their original launch review of the GTX 970.

 

If you're going to make a claim like that, be prepared to back it up. Don't expect others to find your source for you.

[brokencross]

http://www.anandtech.com/SearchResults?q=GTX+970

Well, here's the search results for 'GTX 970' on Anandtech. In order of appearance, I see:

1.) the article I linked in my last post, cautiously concluding that the memory partitioning was unlikely to be a material problem;

2.) their original news article reporting the existence of the memory discrepancy; and

3.) their original launch review of the GTX 970.

 

If you're going to make a claim like that, be prepared to back it up. Don't expect others to find your source for you.

Material problem? What do you mean? The last .5GB cluster is indeed phisically capped. Those are Nvidia's words, not mine.

[typographie]

Material problem? What do you mean? The last .5GB cluster is indeed phisically capped. Those are Nvidia's words, not mine.

 

Yes, it is a slower memory partition, that's already known. By "a material problem," I mean a realistic gaming scenario where the user is going to run into a serious performance issue that can be isolated to the use of that partition. At this point I've given up on the possibility you're just going to read the linked article, but I'll pull a quote out from it to underscore the point:

 

"If NVIDIA’s heuristics and driver team do their job correctly, then the performance impact versus a theoretical single-segment 4GB card should only be a few percent. Even in cases where the entire 4GB space is filled with in-use resources, picking resources that don’t need to be accessed frequently can sufficiently hide the lack of bandwidth from the 512MB segment. This is after all just a permutation on basic caching principles."

 

It's not as simple as "these figures are bad on paper so it's going to perform badly in X, Y, and Z predictable situations." There's a lot more going on than that.

[brokencross]

Yes, it is a slower memory partition, that's already known. By "a material problem," I mean a realistic gaming scenario where the user is going to run into a serious performance issue that can be isolated to the use of that partition. At this point I've given up on the possibility you're just going to read the linked article, but I'll pull a quote out from it to underscore the point:

 

"If NVIDIA’s heuristics and driver team do their job correctly, then the performance impact versus a theoretical single-segment 4GB card should only be a few percent. Even in cases where the entire 4GB space is filled with in-use resources, picking resources that don’t need to be accessed frequently can sufficiently hide the lack of bandwidth from the 512MB segment. This is after all just a permutation on basic caching principles."

 

It's not as simple as "these figures are bad on paper so it's going to perform badly in X, Y, and Z predictable situations." There's a lot more going on than that.

But I've never said it would perform badly in realistic scenarios, such as 1080p gaming. But it does perform badly in some 1440p games and some heavily modded 1080p ones.

[Majestic]

Why are you guys perpetuating this red herring? The problem is headless mode, not the 970 vramgate.

[R_Golden]

But I've never said it would perform badly in realistic scenarios

Yes you did.

[IceCold007]

ive got enough money in the bank at the moment to get two of the fastest 970 cards available but this 3.5 vram issue is just putting me off spending the money. I want to go back to playing skyrim modded with shit loads of mods on it and I just don't want to make the same mistake when I brought my 2 670 and ran out of vram.

 

Holding off and see "IF" the 980ti comes out with the increase vram.

[WhatComesAround]

Its ironic how I am playing gta v at 4k on low settings and enjoying the full capabilities of my 4 gb of vram (gtx 980).

[typographie]

But I've never said it would perform badly in realistic scenarios, such as 1080p gaming. But it does perform badly in some 1440p games and some heavily modded 1080p ones.

 

The quote I gave was talking about cases where all 4 GB of VRAM was in use, it intentionally did not specify resolution. Again, if you have benchmarks that can show those situations as a VRAM-specific performance problem, post them. I'm as eager as anyone to see those cases play out.

Its ironic how I am playing gta v at 4k on low settings and enjoying the full capabilities of my 4 gb of vram (gtx 980).

 

That's exactly what the Anandtech article predicts to be the case, again, provided Nvidia's driver optimizations work the way they claim.

[brokencross]

Yes you did.

Don't misquote me.

 

But I've never said it would perform badly in realistic scenarios, such as 1080p gaming.

 

And,

 

I've never said the 970 was a bad card, it's actually a great card as long the last cluster isn't used, which pretty much excludes any type of 4K gaming and even some 1440p @ ultra. Also some ultra hd texture mods.

 

 

 The quote I gave was talking about cases where all 4 GB of VRAM was in use, it intentionally did not specify resolution. Again, if you have benchmarks that can show those situations as a VRAM-specific performance problem, post them. I'm as eager as anyone to see those cases play out.

You just basically answered your question.

Games at 4K do generally use all the vram available.

[R_Golden]

Don't misquote me.

I never quoted you. Go back a page and re-read what you wrote.

[brokencross]

I never quoted you. Go back a page and re-read what you wrote.

 

You deliberately misquoted me, but whatever. Don't really give a fuck. 

I know exactly what I said. 

 

 

There's a L2/ROP cluster disabled, making the last .5GB virtually unusable. So you don't really need to sugar coat it.

 

Sure it is. 1/7 speed of the remaining clusters is indeed very usable. Whatever, really. 

It's an issue more than proved, even acknowledge by Nvidia themselves and yet there are some of you who just deny it.

I didn't say it was a bad card, not even by a long shot. It just happens it doesn't perform like Nvidia announced it would.

Where did I say the 970 was a bad card? 

Just quit watching chinese cartoons and learn to read instead.