NVIDIA to Tune GTX 970 Resource Allocation with Driver Update

[Kyuuba]

Can software repair a broken hardware?

 

[xfia]

broken? maybe.. if each segment cant simultaneously be accessed then I dont see optimization going very far.

 

[TRWOV]

When you say Nvidia should compensate GTX 970 remember that others should do it too.

FX 8xxx CPUs have 8 integer cores and AMD always showed how the FPU resources were shared. The 2B transistor thing was corrected by themselves a mere day after reviews hit, not 4 months after. And the PS4 has 8GB of GDDR5, having a chunk reserved for the operating system is expected nowadays (XBM used 80 of the 256MB the PS3 had), we are not in the SNES era.

broken? maybe.. if each segment cant simultaneously be accessed then I dont see optimization going very far.

They could use the 0.5GB partition for window composing and such, kinda like having a secondary 512MB GPU for the OS or something.

nVidia says that the last 512MB chunk is 4 times faster than system RAM over PCIe... but you have to take into account that if said 512MB were accessible over the same link you wouldn't need to use system RAM in the first place making the comparison moot.

 

[mouacyk]

...
When you say Nvidia should compensate GTX 970 remember that others should do it too.

Each company's/product's customers are the most credible critics of that company or product. Otherwise, the company or product has no incentive to budge.

Can software repair a broken hardware?

This is exactly a bridge-out problem. If one of two bridges went out, you need to re-route traffic from one bridge to the other. The only way you can maintain the same throughput (people per crossing) is either double the speed on the bridge or shrink the people to half size but moving at the same speed. NVidia's software solution does not seem to accomplish either of those, because it's a hardware/physical problem. At best, they can smartly use the extra 0.5GB as a Level3 Cache to avoid hampering the performance of the primary 3.5GB.

 

[Kyuuba]

This is exactly a bridge-out problem. If one of two bridges went out, you need to re-route traffic from one bridge to the other. The only way you can maintain the same throughput (people per crossing) is either double the speed on the bridge or shrink the people to half size but moving at the same speed. NVidia's software solution does not seem to accomplish either of those, because it's a hardware/physical problem. At best, they can smartly use the extra 0.5GB as a Level3 Cache to avoid hampering the performance of the primary 3.5GB.

Sad news then.

 

[Yorgos]

Another idiot with MS Paint skills.
A CMT core/thread (call it whatever you want) has fully functional cores/threads.
The pipeline is complete for each one, as opposed to SMT.

2 billion transistors vs 1.2 billion. REALLY, is this even an argument. So you wake up one day and decide to buy 2 billion transistors? It's not like they told you that there are 2 billion transistors but the .8 are overbaked and the 1/3 of the CPU is not working.

 

[Naito]

Yeh but most reviews missed this phenomenon, and If you're in the category of user with high res display who needs 4GBs of memory its not good. Those users may have chosen a 290X instead or something else. Nvidia deserves heat from this because they were dishonest about specs.....and they waited until after Christmas sales until the tech media reported on it before they would admit it.

Nvidia advertised the 970 as "having the same memory subsystem as the 980" when it clearly doesn't. There has been a thread on Nvidia forums for 3 months since the 970 came out about stuttering over 3.5Gb, I don't believe Nvidia just figured this out now.

That's a fair point. Nvidia should not have mislead customers, however 4GB is just another number. Just because you have 4GB of VRAM, doesn't mean that it is suddenly capable of running high resolution/high quality games, many other factors come into play. For example, a 4GB GTX 670 doesn't perform to a much higher degree than the standard 2GB SKU at say 1440p, just because it has double the VRAM that's because there are other limiting factors. This also holds true for the GTX 780/Ti; a full 1GB less VRAM than the GTX 970/980, but still perform to a similar level even at 4K.

The way I see it, is that there is two issues here:

1. Nvidia mislead customers. If you purchased the GTX 970 purely on the basis that it has the same memory subsystem as the GTX 980 then sure, by all means be pissed off however, the GTX 970 still performs where it did in the initial reviews and still competes with the competitors products. Its ability to run high resolution/high quality content hasn't changed, it just has optimization issues in very particular scenarios.

2. Perhaps the biggest problem here is not that it has 3.5GB primary partition and a 0.5GB secondary partition, but rather how it is utilized. From what I have read, only one partition can be accessed at a time in hardware and due to optimization issues it can, on occasion, cause stuttering. Not knowing anything about the inner workings and latencies associated with VRAM to system RAM vs VRAM partition one to VRAM partition two and vice versa, I'd hazard a guess and say while having the secondary partition as a cache of sorts, with the idea that it saves time accessing system RAM, may seem like a good idea, if it limits access to primary partition and still gives similar/more latency to VRAM to system RAM access, it may not be such a good idea after all (i.e. it causes more stuttering than it helps reduce it). Either way, Nvidia really need to work out the optimization issues.

 

[xorbe]

Software fix in nv control panel 3D hw settings: Gimped 0.5GB segment [ Enabled / Disabled ]

 

[RejZoR]

You forgot the "Auto" setting...

 

[Fluffmeister]

2 billion transistors vs 1.2 billion. REALLY, is this even an argument. So you wake up one day and decide to buy 2 billion transistors? It's not like they told you that there are 2 billion transistors but the .8 are overbaked and the 1/3 of the CPU is not working.

The number of transistors is of course irrelevant, but pople are questioning how a large company can have such miscommication betweens departments (surely not a revelation?), or if they even read reviews of their own products, AMD seems to be no different in this regard.

The 2B transistor thing was corrected by themselves a mere day after reviews hit, not 4 months after.

http://www.techpowerup.com/156123/a...million-less-transistors-than-it-thought.html

 

[FYFI13]

I hope Nvidia will slash the price of GTX 970 because of this issue and then I'll buy one for myself

 

[THU31]

This card needs 8 GiB of VRAM with eight 8-Gbit GDDR5 chips (instead of eight 4-Gbit ones). The price would not be much higher, but we would get 7 GiB of full bandwidth. That would be enough for pretty much anything until Pascal comes along.

 

[TRWOV]

The number of transistors is of course irrelevant, but pople are questioning how a large company can have such miscommication betweens departments (surely not a revelation?), or if they even read reviews of their own products, AMD seems to be no different in this regard.



http://www.techpowerup.com/156123/a...million-less-transistors-than-it-thought.html

Yeah, but the "loss" of .8B transistor didn't impact how the CPU performed and AMD sent a correction precisely after reviews hit, thus they must have read them. The memory partition on the 970 has the potential of lowering the card's performance on some setups and people didn't know in advance or in a timely manner.

Now, I know you haven't had any problems with yours on any games but there are people that has had them so the algorithm that nVidia used isn't 100% reliable. You can see lots of videos on YouTube showing random stuttering on the 970. Granted, several were uploaded this week so some could be fake but there are some that were uploaded weeks before the drama imploded:

Uploaded 29/12/2014

Uploaded 04/10/2014

Uploaded 07/01/2015

So the card has problems in some setups. I've often had to help people with a problem I can't reproduce, not every PC is the same even if you have two with the same components, OS, etc.

 

[daftshadow]

I hope Nvidia will slash the price of GTX 970 because of this issue and then I'll buy one for myself

Prices won't be slashed until the upcoming Holiday season of this year or whenever AMD launches its r300 series. My opinion, Nvidia will not cut prices on the GTX 970 because of this "issue". Although I would prefer some kind of compensation such as a trade up program for exchanging the 970 for the 980 at a reduced cost.

 

[Casecutter]

I think that is actually being addressed, is it not?

I've yet to hear Nvidia is going do any sort of restitution, for those who are seeing the problems.
I suppose if the driver can resolve the 0.5GB secondary partition running slower and more seamless in transitioning between the partitions then they'll have redemption.

So , you are of the opinion that is was a planned strategy from the get-go as well?

Well YES! Nvidia engineering said they fused off the L2. Just because the company says that marketing wast unaware doesn't mean it wasn't "planned".
"This team (PR) was unaware that with "Maxwell," you could segment components previously thought indivisible, or that you could "partial disable" components."
There big boys and darn well knew at Executive level meetings (marketing isn’t part of that?) what and how those chips were needing to be “segmented” to achieve the volumes asked for by marketing?

Well we have different five threads devoted to the subject here. I guess the next advertising/marketing misstep (assuming there is one by your reckoning) should be a doozy..

I don't start the treads. As above... this can’t be sloughed-off as just some "miss-step" by low level marketing nerds.

The issue certainly isn't, but the level of outrage being shown over a hardware component whose performance hasn't deviated one iota since it was launched and reviewed is certainly cause for humour.

Agreed, while owners who run into this might see the level at which some in community more "white-washing" legitimate concerns just because the "canned" reviews didn't expose it! (Nvidia darn well knew they wouldn't.) If it has no effect as you intend, why couldn't Nvidia not just put information (at release or since) of the use of "segmented" implementation of memory, which something they’re executed in the past. I can't believe there wasn’t many engineers and executives reading reviews who couldn’t recognize this “supposed marketing over-site”, but those employees/company couldn't come out right then, or months ago after they saw the rise of issues on their own forums!

 

[RealNeil]

NVIDIA Said: "and if current owners are not satisfied with their purchase, they should return it for a refund or exchange"

Yeah! All of you guys start returning your 970s and I'll buy them used or refurbished for a lot less!

 

[newtekie1]

I see a lot of people saying they've lost faith in nVidia and are buying AMD now because of this. But what gets me is this has no affect on the performance we saw in reviews. These are just specs on paper, yes they were reported wrong, but they are just the specs on paper. The performance does not change just because the specs on paper change. Just like changing the name doesn't make a difference. And some of you know I'm very much not against renaming cards. I didn't have a problem with it when nVidia did it with G92, and I didn't have a problem with it when AMD did it with Tahiti. Again, the performance for the money is what matters, and that remains unchanged.

But the question I have for everyone trying to say AMD is somehow above doing what nVidia did is, do you remember when AMD released a series of graphics cards and then 2 months later after all the reviews were done and published, released a driver that sneakily reduced performance to stop the cards from overheating and dying prematurely? Do you remember that?

So, which is worse, revising some specs on paper and leaving the performance the same, or purposely lowering the performance with a driver update and not telling anyone?

 

[HumanSmoke]

I see a lot of people saying they've lost faith in nVidia and are buying AMD now because of this.

A lot of that will be AMD shilling and general trolling. Even a cursory look at the forums here will show the most vociferous posters aren't using Nvidia, let alone the GTX 970. I also note that many of the disgruntled "users" on many sites are first-time posters (the same can be said for similar issues that befall AMD, Intel, and Apple). From our own membership of people that actually own the card, I don't see many in a hurry to return it - the larger sentiment seems to be a hope that prices of refurbed cards allows them to buy a second (or more). Of the people I know that actually own them, most are pretty happy - they still haven't got over the buzz of what is for them is an affordable, quiet, and impressive piece of kit (they all bought Gigabyte G1 gaming cards).
An indication might be seen with the poll TPU are conducting. The comments seem to indicate that while Nvidia did wrong, the hardware is OK - but the poll indicates an overwhelming difference of opinion.

But what gets me is this has no affect on the performance we saw in reviews.

Matters not a jot. People gotta get their hate on. You aren't living unless you embrace armchair/hashtag activism, and join the raging against everyone from Obama to why Japan has more flavours of Kit-Kats than your country.

 

[xorbe]

nVidia has altered the deal. Pray that they don't alter it any further.

 

[alwayssts]

I see a lot of people saying they've lost faith in nVidia and are buying AMD now because of this. But what gets me is this has no affect on the performance we saw in reviews. These are just specs on paper, yes they were reported wrong, but they are just the specs on paper. The performance does not change just because the specs on paper change. Just like changing the name doesn't make a difference. And some of you know I'm very much not against renaming cards. I didn't have a problem with it when nVidia did it with G92, and I didn't have a problem with it when AMD did it with Tahiti. Again, the performance for the money is what matters, and that remains unchanged.

But the question I have for everyone trying to say AMD is somehow above doing what nVidia did is, do you remember when AMD released a series of graphics cards and then 2 months later after all the reviews were done and published, released a driver that sneakily reduced performance to stop the cards from overheating and dying prematurely? Do you remember that?

So, which is worse, revising some specs on paper and leaving the performance the same, or purposely lowering the performance with a driver update and not telling anyone?

The problem that I think many non-owners can't seem to grasp (and as shown in the videos kindly posted above) is not that general performance hasn't changed, as you're right it doesn't change the performance shown at launch. The problem is the transition from the 3.5GB to .5GB segment causes stutter. This is very real and it is extremely annoying. This was not showcased/highlighted in many (any?) reviews, perhaps as they didn't think to look for it or saw any potential hiccups as some other personal anomaly. Maybe most tested it at resolutions that could be contained within 3.5GB (again, this is a great 1080p->1440p card as it is), or scenarios the core was bottlenecked before vram became the bottleneck. The fact remains, there are scenarios where the core can put up with gaming scenarios that would utilize that partition for a fluid experience (in essence I disagree with many that say it is moot because it can't). There are instances where the bottleneck is that .5GB, or rather switching to it causes stutter (ie resolutions/settings in Mordor that would otherwise run solidly above 30fps, I'm sure there are others) and that is a problem, especially because we were lied to about it's capabilities. Had we known about that, it may have caused some people to buy a 290x, as at higher resolutions (while otherwise a similar-performing core) the AMD cards will not have this problem. I have said it about 37 times in this thread: no 290(x) would fit in my case; the 970 is the best option for me regardless. That doesn't change the fact the stutter is annoying.

From my understanding, nvidia is doing it's best to shove everything typically contained in ram (driver/os stuff even used at idle) into that .5GB partition so games will generally not utilize it. That *should* help, granted I have no idea what is typical in that regard. As I sit idle at my desktop currently, the OS is using 1GB video ram (but that is also at 4k60). I have no idea how much of that generally is taken away from the OS when a game takes exclusive control of the gpu.

 

[Xzibit]

nVidia has altered the deal. Pray that they don't alter it any further.

There is no updated driver to come fix the day anymore.

PeterS@Nvidia has edited his original post.

Looks like Nvidia is erasing any knowledge of culpability for any potential legal action to come.

Once an engineer now customer care. Poor Peter is going to be tossed to the wolfs.

 

[alwayssts]

There is no updated driver to come fix the day anymore.

PeterS@Nvidia has edited his original post.

Looks like Nvidia is erasing any knowledge of culpability for any potential legal action to come.

Once an engineer now customer care. Poor Peter is going to be tossed to the wolfs.

Yeah, that's unfortunate. I truly wish large corporations realized a little a honesty/culpability can go a long way towards customer loyalty. Honest/thorough/grounded (less pr-speak, more actual no-bs dialogue) interaction between providers and customers is an important tool this day and age...point of the internet and all that. Can't say I'm surprised, nvidia has never been much for giving the personal leeway you see in the blunt expose'-style Richard Huddy/Eric Demers/Wavey Dave approach....Granted I have noticed even AMD gives Huddy a handler these days when doing interviews.

 

[HumanSmoke]

Can't say I'm surprised, nvidia has never been much for giving the personal leeway you see in the blunt expose'-style Richard Huddy/Eric Demers/Wavey Dave approach....Granted I have noticed even AMD gives Huddy a handler these days when doing interviews.

Probably a dying breed given the microscope companies are examined under these days. Is Dave still at AMD, I thought he gave that gig away (commenting on the GTX 970 issue would tend to support that). As for Huddy, I'm surprised AMD don't have a ball gag on hand whenever there is a microphone in close proximity - he does have a habit of putting his employer in difficult situations (whether AMD, Intel, or Nvidia).

 

[RejZoR]

I see a lot of people saying they've lost faith in nVidia and are buying AMD now because of this. But what gets me is this has no affect on the performance we saw in reviews. These are just specs on paper, yes they were reported wrong, but they are just the specs on paper. The performance does not change just because the specs on paper change. Just like changing the name doesn't make a difference. And some of you know I'm very much not against renaming cards. I didn't have a problem with it when nVidia did it with G92, and I didn't have a problem with it when AMD did it with Tahiti. Again, the performance for the money is what matters, and that remains unchanged.

But the question I have for everyone trying to say AMD is somehow above doing what nVidia did is, do you remember when AMD released a series of graphics cards and then 2 months later after all the reviews were done and published, released a driver that sneakily reduced performance to stop the cards from overheating and dying prematurely? Do you remember that?

So, which is worse, revising some specs on paper and leaving the performance the same, or purposely lowering the performance with a driver update and not telling anyone?

Now it doesn't. But do you know for a fact that all will work fine after 1 year when games that demand more memory start coming out? Do you think NVIDIA will care if it will perform like crap then? So, why exactly this defense mode from the users? If you actualyl bought the damn thing you should be even more outraged than we who were going to buy it and then this shit came up. Just a thought...

 

[Xzibit]

More confirmation no Driver Fix for 970 is coming...

Nvidia GeForce twitter response

We are always improving performance through drivers but there are no plans for an update specifically for the GTX 970

Since this has blown up they are going into shut-down mode. Any statement they make will have to be vented through there legal team and I'm sure they are telling them don't reference 970 as it can be seen and used as an admission of culpability.