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

[SPBHM]

It's not useless when as was pointed out so many times that everyone likes to ignore, games can already load up past 3.5GB and don't have problems. Nvidia did not make adjustments for every game and engine in existence. Especially not some older ones I was messing with.

28GB/s is almost useless for such a fast GPU, maybe the games you see are only loading less important data on those 512MB (driver optimization working well) and that's why you don't see the difference, but at some point it could happen, because there is a massive drop in speed from the 3.5GB to the 512MB

 

[JoeRambo]

A lot of shadies left in Anandtech/Nvidia explanation.

I am not into conspiracy theories, but consider the following:

1) Some GTX970 supposedly have compability issues with mobo at launch - failing to initialize cards, sorted out by driver.
Quoting AT's own Ryan:
"Earliest build of NV's drivers were unable to initialize a GTX 970 on our testbed. Was a mobo compatibility issue that NV fixed"
What if those compatibility issues have nothing to do with mobo, but with some scheme of disabled units that initial driver did not support?
2) Some users seem to have a problem of getting card to allocate more than 3.5G of VRAM. What if this problem is caused by cards that have enough faulty hardware to completely block access to last 512MB and yet are still sold by Nvidia?

Given what intentional lie we were just fed, everything is out of the window, and thorough investigation needs to happen with exactly those cards that fail allocation. Now that tech community has some technical details, they can dug deeper to see what bandwith they are getting out of the last 512MB and if it isn't coming out of ram via PCIE ( dropping PCIE gen down to 1-2 will be handy here).

 

[Black Octagon]

Had they been honest about the ROP count and memory allocation up-front, I bet they could have easily spun this to say, "look! All this great performance for your dollar and the GPU doesn't even need flagship-level specs. Such is our mighty engineering-fu."

Bit late for that now, though.

 

[cmdrdredd]

28GB/s is almost useless for such a fast GPU, maybe the games you see are only loading less important data on those 512MB (driver optimization working well) and that's why you don't see the difference, but at some point it could happen, because there is a massive drop in speed from the 3.5GB to the 512MB

There is no driver optimization on games that are 5 years old I'm pretty sure.

 

[SlowSpyder]

Tsk tsk. I suspect that somebody did screw up in the marketing department. But somebody or department within Nvidia needs to catch that even if it goes out the door and correct it after the fact. The internet will figure this stuff out given enough time.

Yea, I absolutely believe it could have been an honest mistake. But in the ~five months since launch reviews and follow ups that have been plastered everywhere, no one at Nvidia that was in the know of the tech specs took notice that every review and tech site had the wrong information on the GTX970?

 

[HurleyBird]

LOL GeForce forums removed the angry thread from their 900 series listings, but surprisingly it still exists if you have the link.

https://forums.geforce.com/default/topic/803518/geforce-900-series/gtx-970-3-5gb-vram-issue/115/

That's pretty low.

Yea, I absolutely believe it could have been an honest mistake. But in the ~five months since launch reviews and follow ups that have been plastered everywhere, no one at Nvidia that was in the know of the tech specs took notice that every review and tech site had the wrong information on the GTX970?

There are only two plausible explanations: One is that Nvidia chose not to come forward with the correct specs after the fact assuming that nobody would catch on. The other is that while many people at Nvidia did see the mistake, Nvidia's corporate culture actively discourages reporting such issues and doing so can be damaging to an employee's career. If this culture does exist, it's probably a post-bumpgate phenomenon.

 

[dacostafilipe]

There is no driver optimization on games that are 5 years old I'm pretty sure.

It has nothing to do with games, it's related to memory allocation by the driver. It just tries to stay below 3.5 and/or swaps less used memory out.

 

[Sunaiac]

I don't see what all the fuss is about.

Before you had 290 performances for 33% more $.
Now you have 290 performances as long as you don't need memory, Rage32 after, for 33% more.

It just makes the perf/price ratio worst, isn't it something (post GTX500) nVidia works really hard on ?

 

[thilanliyan]

I find it hard to believe that nVidia didn't know about the true technical specs of the card, and missed the mistakes in initial reviews. They should have just called the card 3.5GB and be done with it...although "3.5GB" doesn't sound as good as "4GB"...

In the end, the performance impact isn't huge from what I gather, so not a big deal IMO.

 

[Erenhardt]

The performance still is what it is, it isn't like the card suddenly got slower with this new information.

Show a single review that tested the performance above 3.5GB. We knew how it runs below that. We expected it to run seamlessly up to 4GB. Thing is, its not. Even nvidia showed it have and performance impact of 4-6% on average (best case nvidia numbers)
To make it more clear, running 60 fps game, each second you drop 3 frames. But guess what, its AVERAGE. So, one second you will have all 60 frames, and another you will get only 54. How is that not a stuttering problem.

 

[cmdrdredd]

It has nothing to do with games, it's related to memory allocation by the driver. It just tries to stay below 3.5 and/or swaps less used memory out.

So then why is everyone talking about "nvidia will have to keep up support for the latest games"? See I don't think people still know what's going on because in older games I can use more than 3.5GB at specific settings and I don't see anything that looks like I took a 100GB/s bandwidth hit.

Show a single review that tested the performance above 3.5GB. We knew how it runs below that. We expected it to run seamlessly up to 4GB. Thing is, its not. Even nvidia showed it have and performance impact of 4-6% on average (best case nvidia numbers)
To make it more clear, running 60 fps game, each second you drop 3 frames. But guess what, its AVERAGE. So, one second you will have all 60 frames, and another you will get only 54. How is that not a stuttering problem.

Show me a modern game that doesn't use more than 3.5GB at 4k resolution. Many sites did 4k benchmarks. Plus you talk about stuttering when you can't see the stuttering and they fixed much of the issue with dropped frames with a new driver release. Funnily enough the 980 also dropped frames.

 

[thilanliyan]

So then why is everyone talking about "nvidia will have to keep up support for the latest games"? See I don't think people still know what's going on because in older games I can use more than 3.5GB at specific settings and I don't see anything that looks like I took a 100GB/s bandwidth hit.

Maybe in the older games, the bottleneck is elsewhere?

 

[HurleyBird]

So then why is everyone talking about "nvidia will have to keep up support for the latest games"? See I don't think people still know what's going on because in older games I can use more than 3.5GB at specific settings and I don't see anything that looks like I took a 100GB/s bandwidth hit.

Yet many are reporting bad performance and/or stuttering on the GTX 970 playing modded Skyrim.

 

[3DVagabond]

Petition for a refund?

https://www.change.org/p/nvidia-refund-for-gtx-970

 

[cmdrdredd]

Maybe in the older games, the bottleneck is elsewhere?

What bottlekneck might that be? I was seeing how high I can go with DSR at max in game visual settings in games like Batman AC which is going on 5 years old. Some settings put me above 3.5GB and actually really close to the full 4GB but were still at frame rates that I would consider playable for that title. If it was an FPS I'd say no, it's too slow.

Yet many are reporting bad performance and/or stuttering on the GTX 970 playing modded Skyrim.

Depending on the settings, I bet most cards would stutter. Running modded skyrim (an engine prone to some stuttering anyway) with 8x MSAA at 1440p or higher will probably make a 980 stutter at times too.

 

[sontin]

I don't see what all the fuss is about.

Before you had 290 performances for 33% more $.
Now you have 290 performances as long as you don't need memory, Rage32 after, for 33% more.

It just makes the perf/price ratio worst, isn't it something (post GTX500) nVidia works really hard on ?

Hm,
before the GTX970 you paid 20% more for a 290.
After the GTX970 you paid 38% less for a 290.

:sneaky:

 

[SlowSpyder]

Depending on the settings, I bet most cards would stutter. Running modded skyrim (an engine prone to some stuttering anyway) with 8x MSAA at 1440p or higher will probably make a 980 stutter at times too.

But I think what people are concerned about is the stutter that might otherwise not be there on a 224GB/sec 4GB card. Assuming non-CPU limited situations, all cards will choke at some point, we don't get infinite FPS in any game. We understand that, there will always be a limiting factor in your frame rate. But, Nvidia wasn't up front with this potential limitation that can cause stuttering that was previously unknown. This card may stutter and have poor performance in situations that it wouldn't have been expected to, in addition to your normal limiting factors that affect all cards.

Someone should have gotten this information out to tech sites. It could have been spun as 'we were going to go 3.5GB on this, but our great engineering team and driver team found away to make the extra 512MB add value to your card... blah blah blah..." But in stead now they are coming off as if they were trying to hide this.

 

[rainy]

Hm,
before the GTX970 you paid 20% more for a 290.
After the GTX970 you paid 38% less for a 290.

:sneaky:

Sounds like some memory problems - and what happened after Radeon 4850/70 debut?
GTX 280 has dropped from 649 dollars to 499 and GTX 260 from 399 to 299.

 

[xthetenth]

2) Some users seem to have a problem of getting card to allocate more than 3.5G of VRAM. What if this problem is caused by cards that have enough faulty hardware to completely block access to last 512MB and yet are still sold by Nvidia?

I wouldn't be surprised if this happened in cases where stuff that didn't have to be stored in memory was causing a pathologic worst case scenario and 3.5 GB performs better than 3.5 XOR .5 GB.

 

[Mondozei]

I don't buy Nvidia's explanations. Throwing their marketing team under the bus is a true and tried tactic. Jen-Hsun is deeply invovled in the creation of all Nvidia GPUs, he's a very technical CEO and a founder of the company. He may not be as involved as Jonah Alben but the notion that he doesn't know a lot of details about his products is laughable. It fails the smell test.

When he sees the card being advertised as 4 GB of VRAM and doesn't react, c'mon people. Of course he knew.

It's a straight-up lie. The interesting part is if this would hold up in court. I'm guessing no, because you can't really prove that the higher-ups knew about this falsehood, even if it's extremely implausible that they didn't considering that the "4 GB of VRAM" was widely advertised everywhere. They couldn't have missed it. And since it's doubtful that a guy like Jen-Hsun would be ignorant of the most common basics of his own products... it leaves the inevitable conclusion that this was a deception from Nvidia.

It was, as Ryan pointed out, a really dumb decision that bit them in the ass. And I doubt they'll try something similar again. But it still sucks for those who bought it to have essentially been lied to and promised a card that isn't what they were told. And for those who dismisses this, this does have an effect in intensive VRAM games and it has introduced microstuttering for many, so it is an issue.

 

[R0H1T]

What I hate most is, Nvidia falsely advertised GTX 970s 3.5GB VRAM as 4 GB. Not to mention incorrect values for the L2 cache and ROPs listed under specifications, refusing to admit it for months just so their top executives could earn profit by unloading shares.

Last year, between April and most of November, each transaction of NVDA typically involved less than 3,000 shares.

Within the past two months, Nvidia's largest shareholders garnered a minimum of 4 million dollars.

David Shannon
01/23/2015 - Sold 10,000 Shares
01/21/2015 - Sold 21,400 Shares
01/02/2015 - Sold 8,600 Shares
Approximately $800,000

JONES HARVEY C
12/22/2014 - Sold 88,000 Shares
Approximately 2 Million Dollars

PURI AJAY K
12/03/2014 - Sold 24,250 Shares
11/28/2014 - Sold 34,174 Shares
Approximately 1.2 Million Dollars

You cannot deny that insider trading has occurred as each individual listed above dumped ~10%+ of their total shares.

Source: http://www.nasdaq.com/symbol/nvda/insider-trades

You could say the same about nearly every review site out there, since this thing was reported on Nvidia forums the least they could do was investigate the issue before end users forced their hand. I'm not putting the blame solely on review sites but often times there's this conflict of interest you know & if it hadn't been for the news featuring on front pages of guru3d, techpowerup et al(forum members reporting it) you wouldn't have seen Ryan's review exploring the issue anytime soon.

I doubt this can turn into a class action lawsuit but withholding information or misrepresentation of facts ought to have serious repercussions not just from the end users (good luck getting the 970 owners ditching Nvidia anytime soon, they'll just get a 980 instead :awe but also from the relevant consumer protection agencies in various countries !

 

[SteveGrabowski]

What amends can Nvidia make? I want a true 4GB card and the only options for that are full refunds so I can buy a 290x instead or an upgrade to a GTX 980. Either are pipe dreams.

 

[Silverforce11]

Whatever the excuse is, they knew from the start. The maxwell drivers were designed to keep it at under 3.5gb in games they optimized for. No driver team is going to do that without being directed to do so. :/

Regarding the 3.5gb issue itself, does it have an effect in games? Yes & No as AT's article covered. It depends on the game & NV's heuristic/optimizations. When it fails is when games push the vram limit.

"While the frame rates were equivalently positioned at both 4K and 1080p, the frame times weren’t. The graph below shows the 1% frame times for Shadows of Mordor, meaning the worst 1% times (in milliseconds)."

http://www.extremetech.com/extreme/...idias-penultimate-gpu-have-a-memory-problem/2

Also occurs at 1080p where FPS is >60 fps. Poor frame latency = stutters.

http://www.neogaf.com/forum/showthread.php?t=977645&page=16

Also occurs in Skyrim with texture mods & Arma 3 per user reports, and its not at 4K or ridiculous settings.

Star Citizen is known to use ~3.5gb vram at normal resolutions. GTA V on the PC, if its anything like the 4th version which pushed vram limits to the max.. well, its a trend that newer games will push vram more. It's generally referred to as "progress".

Anyway this thread is going to be closed. Because stock trading or insider trading allegations don't belong on the tech forums.

 

[Abwx]

It's called legal money printing.

Thoses guys are not crazy, they have the right to sell stock options that have reached maturity, wich is surely the case of the sold stocks, if sold at peaks it point to inside knowledge used for the purpose but you can be sure taht they sold only part of what was allowed as a cover, legaly nothing can be made against them.

 

[Spanners]

What I hate most is, Nvidia falsely advertised GTX 970s 3.5GB VRAM as 4 GB. Not to mention incorrect values for the L2 cache and ROPs listed under specifications, refusing to admit it for months just so their top executives could earn profit by unloading shares.

Last year, between April and most of November, each transaction of NVDA typically involved less than 3,000 shares.

Within the past two months, Nvidia's largest shareholders garnered a minimum of 4 million dollars.

David Shannon
01/23/2015 - Sold 10,000 Shares
01/21/2015 - Sold 21,400 Shares
01/02/2015 - Sold 8,600 Shares
Approximately $800,000

JONES HARVEY C
12/22/2014 - Sold 88,000 Shares
Approximately 2 Million Dollars

PURI AJAY K
12/03/2014 - Sold 24,250 Shares
11/28/2014 - Sold 34,174 Shares
Approximately 1.2 Million Dollars

You cannot deny that insider trading has occurred as each individual listed above dumped ~10%+ of their total shares.

Source: http://www.nasdaq.com/symbol/nvda/insider-trades

*Reaches for his tinfoil*

You're going to need a lot more information than that to even begin to accuse those people of insider trading. Seeing as how this came to light independently of Nvidia it just doesn't makes sense. Lets dump our stocks because maybe some people are going to find out *rubs hands together* we only have 56 ROPS on the 970? I'm hesitant to say it but LOL.

I'm not sure this will have any measurable effect on their stock price anyway.