Nvidia CEO on GTX 970 controversy: ‘We won’t let this happen again’

Jen-Hsun Huang, Nvidia’s CEO, has spoken out on the recent controversy about the GeForce GTX 970’s memory configuration.

In a blog post on Nvidia’s website, Huang emphasizes that the GTX 970’s lopsided memory configuration is by designβ€”and he says it’s a good thing. “We wanted GTX 970 to have 4GB of memory, as games are using more memory than ever,” he explains.

However, Huang concedes Nvidia should have been more transparent about the issue, and he strikes an apologetic tone at the end of the post:

The 4GB of memory on GTX 970 is used and useful to achieve the performance you are enjoying. And as ever, our engineers will continue to enhance game performance that you can regularly download using GeForce Experience.
This new feature of Maxwell should have been clearly detailed from the beginning.

We won’t let this happen again. We’ll do a better job next time.

Huang’s words may quell some of the discontent among GTX 970 owners. However, Nvidia might also have to make its case in court. As we reported yesterday, the company faces a class-action lawsuit for allegedly misleading users about the GTX 970’s performance.

Comments closed
    • ronch
    • 5 years ago

    Good thing Jensen and the Gang are owing up to it. At least they’re trying to be respectable.

    I can’t say the same thing for how AMD is poking fun at Nvidia’s misfortune and making a scene about it, giving away several R9 290x cards over at their YouTube Gaming Evolved segment. Really hate their Fixer ad campaign too. WTF, AMD. Grow up.

      • sschaem
      • 5 years ago

      But they are not… (Did you read nvidia CEO statement?)

      nvidia is simply saying “Sorry for us not telling you earlier about our cool design for the GTX 970, this design was done only for your benefit”

      Its a statement he had to do, but that admit nothing because any acknowledgment of wrong doing would be used against them in the pending lawsuit.

      So :

      Nothing about the bandwidth loss (196 VS the rated 224GB)
      Nothing about the L2 cache shrinking 1/8 its size
      Nothing about the ROPS miss represented to look like a GTX 980
      ….

      And of course all the drawbacks of having only 3.5GB directly addressable by a 224bit bus, not a full 256bit bus.

      Fact is *none* of that miss representation was done for end user benefits.

        • ronch
        • 5 years ago

        Well, when was the last time AMD admitted their mistakes?

        • derFunkenstein
        • 5 years ago

        I have no doubt the design is intentional and the performance is what it is, so I don’t have a problem with what he said.

    • jts888
    • 5 years ago

    Is there any way to browse users’ prior article (not forum) posts?

    It would be interesting to see which individuals’ activity revolves around supporting or criticizing particular companies.
    Whether paid or not, the appearance of pro/anti NV shilling seems almost overwhelming in these threads. πŸ™

    • wingless
    • 5 years ago

    Everybody is all up in arms about ONLY having 3.5GB of VRAM and here I am with only 2GB on a highend card for less than a year ago.

      • willmore
      • 5 years ago

      Was it advertised as having more than 2GB when you bought it?

    • tanker27
    • 5 years ago

    Gimme a couple of free games and I’ll call it even. πŸ˜›

    I have a 970, its an awesome card. Does it play games I play with satisfactorily mind blowing FPS with all they eye candy turned on…….YES. No complaints from me.

    • mkk
    • 5 years ago

    Well they certainly did _let_ this happen, so they’re correct on that.

    • Ninjitsu
    • 5 years ago

    Wow, I joke about 3.5GB too, but let’s be real here: The 970 has all 4GB available to it. I think some here are making too big of a deal about this.

    Okay, the last 512MB is 3% slower, due to the unique nature of the memory controller. Let’s ask ourselves 3 questions:

    1. Under what circumstances are all 4GB required?
    2. When #1 is true, does the card encounter catastrophic performance issues?
    3. Would the card have been better of, from a performance perspective, to not use the last 512MB?

    If you look deep within your hearts and answer these 3 questions, you will realise that the only real mistake Nvidia made* was not communicating architectural differences to review sites and the enthusiast community.

    Everyone else doesn’t even understand what it means, and would have just ignored it even IF it was mentioned.

    *Whether intentional or unintentional, is up for debate, but it seems clear enough the origin for the error/lie/whatever is the marketing department.

      • sschaem
      • 5 years ago

      this was covered ad nauseum

      The last 512MB can only be accessed via a 32bit controler, running at 1/8 the rated speed.

      Accessing this part of the memory also trash the L2 cache used for the first partition.
      (Some people believe this is the reason for the 970 micro stutter in stress conditions)

      But yes, this was a marketing ploy to inflate numbers. rops, l2 cache, bandwidth, memory size.
      (I’m saying this because I believe people at nvidia are very smart…)

        • Chrispy_
        • 5 years ago

        I doubt it was a marketing ploy. The transistor count of the buddy interface that allows access to all 4GB even whilst an defective L2 cache block is disabled is non-trivial – they needed to put four of these interfaces in even though only one would be used in the 970. die area is money, to these people.

        Secondly, it would be a nightmare to manufacture 970 cards with 3.5GB, because any one of the eight memory channels could be the faulty/disabled/harvested one, meaning that each partner would need to manufacture 8 board configurations for each of the eight 970 GPU variants making sure that the right GPU goes on the right board.

        The money they’d save on the 512MB GDDR5 modules would be vastly outweighed by the 8x increase in board manufacturing processes/assembly/inventory/logistics

        What nobody has done yet is find a way to completely disable the 0.5GB partition and test performance of a 4GB 970 alongside a “3.5GB 970”. I’m willing to bet the clothes I’m wearing that the 4GB card does better in games that need more than 3.5GB of VRAM, which is where all the teeth-gnashing nerdrage is being focussed.

          • MadManOriginal
          • 5 years ago

          I’d bet far more than that. It’s a 100% guaranteed certainty that local video memory, even over a 32-bit bus, is faster than PCIe transfers.

    • kamikaziechameleon
    • 5 years ago

    Tort Reform!

    • Luckyo
    • 5 years ago

    Comments here and comments at the actual blog are hilariously different. The contrast is exceptionally stark with this story.

    I guess it’s to be given that sites like these, that are designed to impact purchase choices are going to be well covered by astroturfers, whereas Nvidia’s own blog which is aimed more at people who already made a purchase is going to be far less “directed” in terms of what gets most visibility.

    I guess that’s one more reason to remind myself that overwhelming amount of commentary and voting on these tech sites is paid astroturfing, not actual people.

      • LoneWolf15
      • 5 years ago

      Citation needed.

        • Luckyo
        • 5 years ago

        Click the link in the actual story and compare.

      • JustAnEngineer
      • 5 years ago

      The compensated NVidia shills register here only to post in the front page comments to attack AMD and praise NVidia. The NVidia fanboys may also participate in the forums, but the shills don’t bother.

        • sweatshopking
        • 5 years ago

        LUCKILY I’M IN BOTH!! <3

        • Klimax
        • 5 years ago

        Sounds like preparations for dismissal of any opinion and facts which are not suitable for out worldview. Something like poising the well…

        And acting like AMD wasn’t caught as well is funny too.

    • Krogoth
    • 5 years ago

    What Jen really meant is that “We are going to fire/demote those in marketing who dropped the ball on this.”

    • Laykun
    • 5 years ago

    Personally I believe the real victims in this case are the retails, suppliers and manufacturers. One retailer I bought the 970s from managed to get a full refund from EVGA, however another retailer I bought a 970 from was refused money from their supplier (Gigabyte brand) so the retailer had to front up with the cash by law.

    You can say there’s no performance penality bla bla bla, fine believe what you want, but a false listing of technical specs compromises the supply chain due to the fact that people can invoke consumers rights laws, and any one, even if they don’t care about the fault can return the card now on a whim stating false adverstising. The card has become a total liability for anyone making, supplying and selling them. So because nvidia screwed up a lot of your local retailers have to suffer the consequences, particularly if nvidia doesn’t admit to the fault and doesn’t provide some form of reparations to the parties in their supply chain. In essence, the financial risk isn’t where it’s supposed to be, nvidia makes the hardware, nvidia did the marketing, nvidia made the screw up, yet it’s retailers, suppliers and manufacturers taking the financial risk.

    If you should be angry about anything, it should be about nvidia not supporting their supply partners all the way down to the store front.

    • oldog
    • 5 years ago

    But does it matter?

    [url<]http://www.gamespot.com/videos/reality-check-can-you-tell-the-difference-between-/2300-6423491/[/url<]

      • moose17145
      • 5 years ago

      If you take a game designed first and foremost (or in Star Citizens case, Soley) for the PC, and compare it against a console game, then yes, you can easily tell the difference.

      But if you take a console game that was simply ported to PC…. then no… you cannot tell as much of a difference other than (usually…. but not always) smoother frame rates.

      Just saying… there is not a chance any of the consoles would have a snowballs chance in hell at playing my heavily modded skyrim at acceptable frame rates. And modded skyrim is something crazy graphically.

    • Antias
    • 5 years ago

    Sigh, I really don’t car about any of this.
    I upgraded from 560Ti SLI to a single Gigabyte 970 OC…
    And i’m way happy at the huge performance gains i’ve recieved.
    The minimal impact I “may” have from the memory nerph is so not worth the stress and anxiety thats going on about this…
    chill people… πŸ™‚

    • Peter.Parker
    • 5 years ago

    “This won’t happen again” … My wife never buys this excuse when I say it.

      • HisDivineOrder
      • 5 years ago

      That’s because she’s the one who winds up in the neglige reading a magazine, bored and feeling a little less attractive each time you say it.

    • Vrock
    • 5 years ago

    Nerd rage about an affordable product that performs well is nerd rage, because the specs are technically accurate instead of *technically* accurate. Nerds who were quite happy until this “story” broke. And this is partly why I have largely abandoned PC gaming, and the PC hobby in general. Because you guys suck. A lot. You’re also extremely predictable, so you’ll focus your nerd rage on the red thumbs down icon up there, and feel a deep sense of satisfaction in doing so. Because again, you suck. You suck so much, you bite the hand that feeds your suckage.

      • ALiLPinkMonster
      • 5 years ago

      Everybody sucks. Doesn’t matter where you go.

        • MadManOriginal
        • 5 years ago

        Truth.

        Exhibit A: YOUR MOM.

        • Vrock
        • 5 years ago

        True, but some people suck harder and more frequently than others.

      • crabjokeman
      • 5 years ago

      Your pathetic trolling isn’t even worthy of a downvote. I’ll save those for ssk, thank you very much.

        • Vrock
        • 5 years ago

        The fact that you think I’m trolling, instead of lamenting a hobby and community I used to love, merely reinforces my point.

        I remember when graphics card companies ACTUALLY CHEATED to win at benchmarks. That was worthy of outrage, and that outrage did a lot of good. This fabricated “scandal” isn’t even CLOSE. In fact, it’s so lame and contrived, you people have even confused the lawyers.

          • l33t-g4m3r
          • 5 years ago

          Do you HONESTLY think that NV is no longer cheating in benchmarks, when TWIMTBP tessellated cement blocks, physx, and GameWorks still exist? NV merely went the efficient route and had developers (and reviewers) incorporate cheats instead of including them in the driver.

          NV made a 660 and sold it as a 670 with the 970. It’s real price point is $50+ less than where it is now, and AMD’s uncrippled competing card is still cheaper. I don’t call that “affordable”, I call it a rip-off.

          You know what pisses me off about the PC community? Apologist Fanboys, and console converts who complain about the PC’s complexity. Guess which part you fall in. NV screwed up, don’t make excuses for them. That’s what their PR dept is there for.

          Quite frankly, I’m tired of NV crippling all their mid-range products and selling them at higher prices than AMD’s uncrippled cards. Thus, I’m hoping NV will learn, and their next gen mid-range cards will not have crippled memory, or anything else. Consumers want their 970 to be the successor to the 670, not the 660. Just because the 660’s successor is named a 970, and performs “good enough” does not excuse the fact that it is in reality the new 660, and NV mislead it’s customers into thinking otherwise.

            • MadManOriginal
            • 5 years ago

            re: pricing…please check the launch price and what cards were going for at that time.

          • spugm1r3
          • 5 years ago

          I have to say, generally speaking, I agree with all of your points. I do get some of the outrage, from an engineering point of view. But the level of outrage would only be acceptable if everyone angry about their affordable (I paid $500+ for a 780, so…), high performing, lower power consuming GPU, also had degrees in electrical engineering.

        • LoneWolf15
        • 5 years ago

        SSK is actually funny when it trolls. This is just flamebait.

          • crabjokeman
          • 5 years ago

          ssk is rarely funny, but even a blind squirrel finds a nut sometimes. At least ssk is trying to be funny; this ass is just trying to be an ass here.

      • anotherengineer
      • 5 years ago

      “affordable” is a relative term

      $445 + $9(shipping) + 13% tax = $513.02

      [url<]http://www.newegg.ca/Product/Product.aspx?Item=N82E16814487088&cm_re=gtx970-_-14-487-088-_-Product[/url<]

        • Vrock
        • 5 years ago

        You’re in Canada. There’s 30% of your problem right there.

      • Krogoth
      • 5 years ago

      The publicly released specs were incorrect at 970’s launch and Nvidia failed for whatever reason to properly disclose the correct specs. They are suffering the consequences for this as a result.

      I do agree that whole affair is a minor thing in the grand scheme of things, but this isn’t nerd rage. It is consumers who are acting on their own interest. This is no different from a good/service that provided incorrect/contradictory information versus what was advertised and there are no disclaimers/warnings. The provider/manufacturer is suffering the consequences for it.

      Elitism plagues every hobby out there and I hate to break it to you, but your original post come off you being among that crowd.

      • fhohj
      • 5 years ago

      Johnny Cage Wins.

      Fatality.

      I bet you’d have no complaints buying a new car with a temp spare tire on it. Also, the GTX 970 is not in the “affordable” range, to the point where things should become “accept crap with your crap card”, [i<]assuming[/i<] such a stance should ever hold true. As a result, users do not expect major shortcomings on something like a 970. It's not as if it "4GB!" was on the box but reviewers were made aware of this situation. nVidia didn't tell anybody. And as an aside, they made mistake not doing so. If they had, it would have been another reason to talk up nVidia's skill because as no reviewers knew about this, they would have been caught off guard, without methods to find holes in it, and would have been largely impressed by it. Time will tell. It's all down to whether or not nvidia's driver heuristic thing will forever be up to the task, the maintenance of which forever holding a priority. New games make all cards slow eventually. It remains to be seen whether this gimpy knee on this card will cause it to falter even harder.

      • moose17145
      • 5 years ago

      If we all suck so much, and you hate us so much, and you have truly abandoned PC Gaming and PCs as a hobby… then why are you here?

      • Dr_b_
      • 5 years ago

      Vrock,

      Weren’t some people having real stuttering issues with these cards?

      If you are gaming at hi-res and max settings, invest in the top of the line card. Otherwise turn down settings a bit.

      What is the substance of the lawsuit, the card actually has 4GB RAM, going to be tough to win that one right?

    • LoneWolf15
    • 5 years ago

    You know, if it weren’t for the fact that I owned an NV40 card with PureVideo that didn’t work, and before that, an nForce4 mainboard with a broken hardware firewall, I might take the “never again” a bit more seriously.

    • DeadOfKnight
    • 5 years ago

    The bandwidth is 7/8 the size advertised. For the majority of users that’s a bigger sin than having a smaller frame buffer since we will rarely go above 3.5 GB at 1080p.

    Basically they wanted to put 4GB on all of their boxes to compete for market share with AMD. And that’s all they are apologizing for, which is missing the point.

      • Goofus Maximus
      • 5 years ago

      That’s a little bit overstated, since the bandwidth restriction only kicks in when you stress the graphics memory to the absolute limit so that it actually uses that last .5 GB of memory, as stated by the crew of the Tech Report themselves. The bandwidth is 8/8 the size advertised in most situations.

        • DeadOfKnight
        • 5 years ago

        When is the bandwidth 8/8? From what I understand, it’s always 7/8 and when that last .5 GB needs to be accessed it has that last 1/8 bandwidth allocated to it. It’s basically partitioned off and unused until you go over 3.5 GB and then at that point it causes slowdowns because it runs at 1/7 the speed of the rest of the card. Nvidia’s argument is this is better than running over memory, which may be true but I don’t think it’s going to happen that often. The false advertisement is not that we don’t have 4GB, but that under normal circumstances we don’t get 256 bandwidth. And the lower performance of that last .5 is just icing on the cake.

        And I’m not arguing that they made a bad choice. It makes sense for them to cut it down the way that they did rather than just not use that .5 GB. However, this card is falsely advertised, and not because it’s really 3.5 GB. That’s insignificant in comparison to claiming 256 bandwidth.

          • Waco
          • 5 years ago

          The bandwidth, in aggregate, is what they claimed. Also…if it doesn’t hurt performance appreciably (which it doesn’t)…DOES IT MATTER?

            • Krogoth
            • 5 years ago

            It is because it is not contiguous and it does impact performance in the form of micro-shuddering.

            Technically it is throughput of 196GiB/s for the 0-3.5GiB space and 28GiB/s for 3.5GiB-4.0GiB space.

            • EndlessWaves
            • 5 years ago

            How much of that is simply because it’s not being used appropriately though? Is the problem that games need more than 7/8 of the memory running at high speed or just that games are placing the data that needs high speed in the wrong bit of the memory?

            Yes, it’s not ideal but there must be hundreds of such compromises in both AMD and nVidias chips.

            Also, doesn’t the GTX 970 implement the same memory saving colour compression we saw in the 285 that’s absent from the 290X? If so it may be that it’ll happily run the same settings as the 290X 4GB within it’s 3.5GB of fast memory.

            • DeadOfKnight
            • 5 years ago

            Personally I’m not really offended if the price is right, and it really is. It’s still a good card for the money. The lack of transparency is the issue, and their apology is lackluster. They’re missing the point. They think we’re just mad that it’s not a 4 GB card in the traditional sense, but really the problem is that the card is slower than implied under stress.

          • Goofus Maximus
          • 5 years ago

          The bandwidth is 8/8 whenever the card is not actively using that last .5gb of memory, which is actually most of the time, since you have to go CRAZY to stress the card to that limit. I know PCPerspective has tested this. If I could find it, I’d post the TR article, but here it is below.

          [url<]https://www.youtube.com/watch?v=b74MYv8ldXc#t=214[/url<] for a clear explanation of the issue. [url<]http://www.pcper.com/reviews/Graphics-Cards/Frame-Rating-GTX-970-Memory-Issues-Tested-SLI[/url<] Also, it's not "really 3.5 gb" at all. It has the full 4 gb. The final .5 runs slowly, due to the shared controller, but it's all there, all accessible, and all the game benchmarks run by third party reviewers like this site are the same before the controversy and after. Right now there just a feeding frenzy of negativity that is entirely too large for the relatively minor crime involved, and these "7/8" spurious factoidlets are being bandied about by the pitchfork-wielding masses who are all too into fanrage mode to think clearly. Sure Nvidia's marketing department pulled a big booboo in their PR for the card, but really, it's not really that big of an issue in reality. I'm launching this "things aren't so bad as they are being stated" post, with it's dig at the mob-frenzy overstating the severity of the problem, with the full expectation that it will be downvoted into the very depths of perdition. I don't care! Too me, the folks doing the downvoting are part of the problem of internet mob-action, and it will just give me the feeling of being Joan-of Arc; a martyr for the cause of "it's just not that serious folks!" πŸ™‚ (There! I just excessively stretched my own sense of self-importance as well! I'm really just another internet noise-maker, after all!)

            • jts888
            • 5 years ago

            The 7/8 thing comes from the fact that it’s 7/8ths the speed of the 980, which was advertised as having an identical memory system.

            So, 8/8th = 8*512MiB, 8*32b wide, 8*28GB/s, 8*256kiB L2, 8*16 TMU, 8*8 ROP.
            7/8 = 3.5GiB, 224b wide, 196GB/s, 1792kiB L2, 112 TMU, 56 ROP system the 970 really operates as.

            The 980 stripes texture, etc. reads across all 8 of its memory sub-buses, while the 970 stripes data only across the 7 corresponding to the non-disabled L2/TMU/ROP block.
            In theory, some slow work could be done in parallel on the slow 512MiB chunk, but Nvidia apparently couldn’t figure out how to do so without wasting shader resources, so the driver just does its absolute best to never put anything in the slow “last” 512 MiB segment.

            The number of people who actually looked at the memory specs and bought a 970 instead of a 980 because they wanted memory speed but didn’t care as much about shaders is probably minimal, since the only application I can think of would be doing stuff like playing older games such as Quake 3 at 8k resolutions or at 4k, 240+ Hz or something.

            That said, I’m OK with Nvidia being a whipping boy on this issue if it helps keep vendors more honest in the future.

      • gdonner
      • 5 years ago

      Well said! Business 101: Stop lying and trying to deceive your customers:

      [url<]http://www.gregdonner.org/thoughts/thoughts10.html#business101[/url<] Unfortunately, it happens all the time--even with their lies and deceit splattered all over forums, YouTube, etc.

    • brucethemoose
    • 5 years ago

    All this fuss, and 970 prices STILL haven’t dropped.

      • MadManOriginal
      • 5 years ago

      Which speaks directly about the rational, market view of the situation.

    • anotherengineer
    • 5 years ago

    Insert Jen-Hsun Huang facepalm here.

    • sschaem
    • 5 years ago

    Just a thought.

    I read many GTX 980/970 reviews when they cam out. maybe half a dozen.

    All reported the data from nvidia. the rop count, the L2 cache, the memory bandwidth.
    Usually in table form, making all those spec identical betwen the GTX970 and GTX980

    Making it clear that none of the area of the chip had been cut down.

    Now… How many nvidia engineer (considering its their little new born) went on the web to see how their work was reviewed by the tech community?

    I can only guess that all those specification errors didn’t escape them….

    Did they report it to marketing and got shut down?
    Did they keep quiet?
    Did they know it was intentional?

    My point I guess is, its bound that many engineers at nvidia looked at those reviews, but for one reason or another the spec where never corrected.

      • jessterman21
      • 5 years ago

      Marketing people do not like being corrected. Especially after a public release.

        • bill94el
        • 5 years ago

        You got dat right πŸ™‚

      • MadManOriginal
      • 5 years ago

      Funny, I read the performance tests in reviews first or when skimming. The specs and technical stuff is just fun nerdy info.

    • south side sammy
    • 5 years ago

    I bet they’re sitting around a table right now discussing who’s going to be sacrificed for rehab and who’s going into detox. that should make things okay with the public…….. it usually does right?

    • sschaem
    • 5 years ago

    “This new feature of Maxwell should have been clearly detailed from the beginning”

    Along with the misleading bandwidth rating of 224GB, when its 196GB
    Along with the L2 cache size that is suddently 1/8 smaller
    Along with the ROP count that is 1/8 smaller
    Along with the 3.5GB partition only running on a 224bit bus

    And the fact that carefull driver optimization are required , forever, to have the 32bit bus not slowing down games (thats why now the card was limiting itself to 3.5GB)
    This slow as molasses “sideband” memory (that pollute the main pool cache when accessed, is going to create a mess. No game designer is going to optimize for it. so its all going to be driver black magic.

    But I guess with the lawsuit, the CEO cant admit to all this publicly…

    “It was all done for your own good, no ours”

      • Laykun
      • 5 years ago

      Game developers don’t have a choice about what goes where in memory, it’s all up to the driver to determine that. So even if developers wanted to optimise for it they really couldn’t beyond some ugly ass hack.

    • NeelyCam
    • 5 years ago

    [url=http://www.pcmag.com/article2/0,2817,2477327,00.asp<]HI GUIZE![/url<]

      • Srsly_Bro
      • 5 years ago

      LOL. Thanks for the post NellyCam

    • Srsly_Bro
    • 5 years ago

    Any accounting people in here?

    FASB ASC 450-20(e), Loss Contingencies, comes to mind.

    • Andrew Lauritzen
    • 5 years ago

    Seems like a completely reasonable response to me. It’s a perfectly sensible hardware feature, and they acknowledged that they should have been clearer about it up front. You can of course construct conspiracy theories about them hiding it intentionally but that’s ultimately a waste of time.

    • moose17145
    • 5 years ago

    I really think this is getting blown way out of proportion.

    I kinda feel bad for the engineers though.

    They probably worked really hard to get the 970’s lopsided memory to work well (and honestly, from what I can see in benchmarks, it DOES work well), and thought they were making a better product by creating this quarky work-around to get it a full 4GB of memory in the event it would be needed.

    And then the software guys who worked really hard to get the drivers working well with the hardware. Like they said… they were trying to make it a better product by doing it this way, and I do believe them on that front.

    I honestly thought this approach to dealing with a slightly crippled 980 chip to get as much performance out of it as possible was really neat and innovative.

    I still think the 970 is a really good card for anyone looking for a card in that performance / price bracket.

    I just hope that it does not dissuade the engineers from trying to do stuff like this in the future (being innovative with chips that are not quite perfect for flagship status). I don’t blame them for this debacle. I blame marketing and other entities involved.

    And I say this as someone who has for a long time now Preferred ATI/AMD cards and who is quite happy with my current R9 290. So I wouldn’t say my opinion on this matter is biased in NV’s favor.

      • HisDivineOrder
      • 5 years ago

      It’s funny. Thinking like that, you have these incredible GPU engineers that enabled a less cut-down version of Maxwell GPU’s to be produced so we could get more GPU for our buck. Remember when the world was on fire for a slightly cut down version of Maxwell for $330? No one could believe it. The performance didn’t undercut that, either. It hit the mark everyone expected for that level of being cut down AND the price was lower than anyone thought possible.

      Then you had the software team that helped make it possible. It’s no easy feat.

      All undone because a media relations team screwed up back in November and put nVidia in a bind where they could walk back the technical specs they released (and take flak then and lose sales during the holiday period from just internet complaining) or wait to walk them back later in the post-holiday world.

      They chose the latter and for that they do deserve their pound of flesh extracted, but the tragedy really is the performance was and is still present. So much so, no one noticed a performance deficit until they started obsessively watching their GPU memory usage.

      If only that marketing team had just given the right specs from the get-go, all of this could have been avoided. That screwup put nVidia management in the choice between admit the problem right at the start or later down the line.

      And choosing later down the line was dishonest. The performance is still great, though, so the blow is softened by the fact that nVidia didn’t ADVERTISE their card with ads that said, “64 ROP’s! 4 GIGABYTES OF VRAM 100% OF WHICH IS RUNNING AT FULL BANDWIDTH!”

      So false advertising? Eh… no. Just false relaying of info to review sites. Review sites that didn’t do their due diligence to test the cards and make sure all specs lined up exactly as described, testing the memory utilization to be sure everything was in order. Not their fault, but I wonder how many times must corporations prove they are entities built from the ground up to make money at all costs before people start NOT TRUSTING anything they say.

      nVidia, AMD, Intel, Facebook, Google, Microsoft, or even–dareIsayit–Comcast or Time Warner. They cannot be trusted. Test everything. Take nothing for granted. Life will be better and easier for everyone if you do.

      • l33t-g4m3r
      • 5 years ago

      There’s nothing wrong with the 970 per se. It was just misrepresented, and AMD’s uncrippled competing card is cheaper. I’d have no problem with the 970 if they were forthright about it’s capabilities and it sold at a price point more in line with those capabilities.

    • Welch
    • 5 years ago

    Yep… so…

    He is basically saying you’ve been enjoying the performance of the 970 this entire time, we didn’t feel the need to explain the “design” in Maxwell that allows us to use 2 different speeds of RAM and we are sorry we didn’t explain it to you better. In the future we will make sure to explain it properly and properly charge you for it most likely.

    • hansmuff
    • 5 years ago

    “new feature of Maxwell”

    I thought Kepler already was doing the very same thing on the 660Ti, with slow performance on the last 512MB because of asymmetrical configuration on one memory controller..

    • Prestige Worldwide
    • 5 years ago

    WHAT ABOUT 64 VS 56 ROPS?

    IT AIN’T ALL ABOUT THE VRAM!

      • HisDivineOrder
      • 5 years ago

      Well, at least you know he didn’t apologize for that one.

      Just like how they never advertised it having 64 ROP’s, he isn’t acknowledging that they ever said anything about it ever.

      Well, except to review sites that demanded MORE technical data.

      I’d imagine this will be a great reason for nVidia to pull back on releasing technical data in the future. They can sidestep the entire issue by not discussing the technical design of the GPU and let people guess and be amazed by the performance.

      If they’d done that here, no one would be whining and everyone would just be enjoying their Geforce 970’s in peace.

      Sometimes you have to be cautious about what you complain about. Lest they take away all sources of your problem and make you lament the loss later.

        • renz496
        • 5 years ago

        [quote<]I'd imagine this will be a great reason for nVidia to pull back on releasing technical data in the future. They can sidestep the entire issue by not discussing the technical design of the GPU and let people guess and be amazed by the performance.[/quote<] this will be interesting if nvidia really do that. so i want to know what TR takes if nvidia really pull such a move?

    • AmazighQ
    • 5 years ago

    That is the same kind of apology my 3 year old son makes after he pissed next to the toilet again.
    You know what ever he is going to say. The next day there is going to be a big chance he will do the same.

      • HisDivineOrder
      • 5 years ago

      Working as intended?

      • crabjokeman
      • 5 years ago

      Start saving for his MBA now.

    • deruberhanyok
    • 5 years ago

    It’s weird, because when they start using those 8Gb GDDR5 chips you guys reported on:

    [url<]https://techreport.com/news/27676/samsung-starts-making-8gb-gddr5-memory-chips[/url<] They can just make a 7GB version, and no one would care. They'd say, oh, 970 has 7GB of memory because it's got some stuff cut out, and the 980 has 8GB, same way we've seen other cards have 3GB of memory vs 4GB when the memory bus width is cut down. Maybe they just really didn't want to advertise it as a 3.5GB card? Which is weird, because we've had 1.2GB cards, and 1.5GB cards, and 1.75GB cards, so having that .5 on the end wouldn't have been unusual. This whole thing has just been surreal. I returned my 970 because of the coil whine, but if it hadn't been doing that, I'd still be perfectly happy with it.

      • tay
      • 5 years ago

      It was pointed out to me, that each memory controller unit on the chip is 64 bits wide. Since each memory chip on the the card is 32 bits wide (32*8 = 256 bits) they would need to disable 2 memory chips and make it a 3 GB card. I’m willing to bet good money that a 3 GB / 196 bit card will be a fair amount slower than the ~ 224 bits and ~ 4 GB that the 970 ended up with.

      I’d buy a 970 if it was $50-80 cheaper.

        • mczak
        • 5 years ago

        It is however not very obvious what exactly ties 2 32bit channels together in a 64bit memory controller. Since apparently both the L2 cache and ROP are effectively per 32bit partition.
        I think a design which would have allowed for just 7 32bit channels would have been a better idea, which then would make for a “true” 224bit / 3.5GB card. As is the additional 0.5GB of memory seems to cause more trouble than it’s worth due to the problematic management. It appears to me such a design would have been trivial but it’s possible there are indeed technical reasons (and not just marketing ones for being able to sell cards with slightly more memory) preventing this however.

          • the
          • 5 years ago

          It likely has to do with how timings are handled at a low level. Things like skew are calibrated for the full 64 bit width. Various buffers of handling skew are also done on a per controller basis.

          Now having said that, nVidia could have done 8 memory controllers instead of 4 but this would have used additional die space to handle additional calibration logic and additional buffers. It is an engineering trade off.

      • geekl33tgamer
      • 5 years ago

      Its all about keeping up with the Jones’s, so to speak. Its closest performance rival has 4GB, so being sold with “less” wouldn’t be as appealing come retail time?

      • sschaem
      • 5 years ago

      3.5GB because its a perceived value compared to the competition.

      That extra .5GB add value to the card making the gap larger.
      Also making a lot of the spec the same as the GTX 980.
      Same ROPS, same L2 cache, same memory bus spec (256bit / 224GB) add perceived value.

      Its sad, because nvidia didn’t need this marketing scheme at all to have a successful product.

      • September
      • 5 years ago

      Well 3.5 rounds up to 4 anyway so who cares?

    • MetricT
    • 5 years ago

    As a GTX 970 owner, I’d like to tell these lawyers to go pleasure themselves with a saguaro cactus. The 970 is a tremendous value. and I’m thrilled with mine.

    All this does is increase the odds that the 970+1 won’t be nearly as attractive as the 980+1 because nVidia will have to increase the product differentiation to keep the lawyers happy.

      • Flying Fox
      • 5 years ago

      But when you are qualified for that $10 settlement, will you sign up to be in the class?

        • Deanjo
        • 5 years ago

        He’s metric and the class action is only applicable in the States. NO SOUP FOR HIM!

          • Srsly_Bro
          • 5 years ago

          LOL epic post. idc what SSK says about you, you’re alright.

        • solo_clipper
        • 5 years ago

        Of course he will, may take a while though. I just signed up for this. [url<]https://www.themoneyismine.ca/[/url<]

    • Silus
    • 5 years ago

    GTX 970 performance before this was known : 70 fps in Game X

    GTX 970 performance after this was known : 70 fps in Game X

    Sure, they made a mistake in the marketting documents and should own up to it, but they did not mislead anyone in terms of the card’s performance.

      • HisDivineOrder
      • 5 years ago

      Shhhhh. People who don’t own these cards need something to scream about. People who did/do own these cards either returned/sold them as soon as they found out the dealio or they learned to live in a post-not64ROP’s/3.5gigsofVRAMwithBenefits world.

      There should be a poll done to find out how many people complaining bitterly about this issue own a non-nVidia card in their PC. I imagine more than a few Negative Nancies are Gaming Evolved.

        • Tirk
        • 5 years ago

        So I take it likewise that you would assume Nvidia owners berate AMD issues equally as well?

          • HisDivineOrder
          • 5 years ago

          I take it that a lot of the 970 owners who are verified 970 owners seem to be more concerned with the performance living up to the benchmarks than metrics that were merely there to justify why the performance was at a certain level.

          Put another way, if I bought a $330 970, I’d want the performance I saw in benchmarks. If I didn’t get that, then I’d be upset. If they promised me it had 2048 ROP’s and 20TB of RAM and it performed well enough I thought it had that, but then one day they told me it had half the ROP’s and half the VRAM, but I’d thought all along until that very moment it performed up to my expectations for the cost I paid, then… I’d argue I wouldn’t care.

          And I don’t think people who actually own a 970 are by and large going to care. So I wouldn’t say that nVidia users are more likely to defend it. I’m saying 970 owners probably are, though.

            • Tirk
            • 5 years ago

            At what point then do you consider the dissenting opinion to your own from a 970 owner?

            You made the assumption that just because someone disagrees with you its because they don’t own said hardware. That’s a very convenient way of discounting other potential 970 owners opinion, which makes a false affirmation of your own opinion based on nothing but your imagination.

            If you’re so confident 970 owners don’t care then why even respond the way you did, they would also apparently disregard any additional information that might counter their perception of the performance they expect from their card.

          • MadManOriginal
          • 5 years ago

          Berating AMD issues is like kicking a sick, elderly dog. It just feels wrong.

            • Tirk
            • 5 years ago

            Lol you said it not me. But I’m glad you feel its wrong to kick a sick, elderly dog.

      • sschaem
      • 5 years ago

      224GB memory bus…. full 4GB of memory.

      Thats not misleading ? come on.

        • Waco
        • 5 years ago

        Both of those specs are true, I’m not seeing your point.

          • sschaem
          • 5 years ago

          ?

          224GB, no its 196GB *OR* 28GB . Also to use the 512MB on the 32bit bus require some driver magic as this is not exposed in any way to developers.

          Test apps have shown that the GTX 970 could not populate its full 4GB, only later where driver ‘updated’ to use that gimped .5GB.

          Yes, you have a card with 4GB worth of memory, but 12% of it is gimped and was non functional until nvidia got exposed and updated their drivers.

          And keeping all that info a secret is not misleading? (specially when the spec where all inflated and wrong, and where later corrected)

          And you dont think nvidia showing the GTX 980 and GTX 970 memory subsystem side by side with the exact same spec is misleading when in fact the GTX970 is over 12% slower then rated in nvidia spec?

            • Waco
            • 5 years ago

            No, it’s 224 GB/s in aggregate. Asymmetrical memory access is not a new thing in GPUs…

            • jihadjoe
            • 5 years ago

            > 224GB, no its 196GB *OR* 28GB . Also to use the 512MB on the 32bit bus require some driver magic as this is not exposed in any way to developers.

            You could say the segmentation is actually being caused by the “driver magic”, and without it the GTX970 would appear to have 4GB of homogenous memory, except one 512MB block would transfer slower than the other 7 because of the missing cache line.

            Either way the days of game developers coding direct to metal are long behind us. Accessing ANY of the VRAM on Windows pretty much requires the intervention of some driver magic.

        • Klimax
        • 5 years ago

        Does it have 4GB. Yes, it does. Does it have full memory controller? Yes, it does. (No block of it was disabled) Post fully and completely wrong.

          • sschaem
          • 5 years ago

          Did you check nvidia provided block diagram? its public info.

          the chip cant communicate with the low level MC block directly.
          it does so via the L2 cache connected to the crossbar…
          and so, do you notice that the GTX 970 doesnt have the require L2 cache block connected to the last MC block ?

            • Klimax
            • 5 years ago

            And? Still has full controller and memory. Internal implementation is for those two things in specs irrelevant. And we had number of asymmetrical cards from NVidia anyway. (They just oversubscribed directly memory controller, here it moved to L2/ROP) So even there we see nothing unusual.

            It in no way changed anything tangible, just few docs…

      • MrDweezil
      • 5 years ago

      Seriously, who cares? I skip over everything in a video card review except the price and the graphs anyway. It could have 2mb of memory and the graphics chip from a SNES and as long as it performs the way they claimed I’ll be perfectly happy.

        • VincentHanna
        • 5 years ago

        Question for all the nerds out there:

        At what number of [b<](PETA?)[/b<]hertz would you have to overclock a SNES chip so that it could compete with a 970 in raw calcs per second?

      • Krogoth
      • 5 years ago

      False analogy

      This problem only affects that 970 when it forced to use beyond the 3.5GiB range. A more correct analogy would be like this.

      980 performance when game is consuming ~3.7GiB of VRAM = ~40-50FPS

      970 performance when game is consuming ~3.7GiB of VRAM = ~35-45FPS with constant bouts of [b<]micro-shuddering[/b<] 290X performance when game is consuming ~3.7GiB of VRAM = ~35-45FPS It does impact performance under such conditions and 970 buyers didn't know about this issue at launch. They are kinda annoyed about it.

        • cobalt
        • 5 years ago

        [quote<] 970 performance when game is consuming ~3.7GiB of VRAM = ~35-45FPS with constant bouts of micro-shuddering[/quote<] Has this actually been shown to happen, or are you just speculating?

          • Krogoth
          • 5 years ago

          Yes, it has been shown to happen to a number of 970 users and a few them happen to have a 980 on hand to make a comparison.

          In fact, it the reason how the whole 970 issue got discovered. Some 970 users got curious wanted to test their hardware to its limits to see what it can do. They noticed that 970 was shuddering for some strange and decided to take a closer look. They noticed that problem begins to happen after VRAM usage goes beyond 3.5GiB and they also noticed 970 becomes reluctant on using that memory space. They also noticed a frequent dippings in the frame-timing and GPU utilization charts.

          In light of this, some CUDA developer on a forum decided to use a simple GPU memory testing application for a spin on a headless setup between the 980 and 970. They noticed that same thing. The results were post on a number of enthusiast sites and this is how this whole thing became viral.

            • cobalt
            • 5 years ago

            This should be measurable — the in-game micro-stuttering — simply by using metrics like TR’s 99th percentile frame times, or the overally cumulative distribution function. The synthetic CUDA tests will obviously show that bandwidth drop, but the NVIDIA response was that in games, the driver should avoid pathological cases that the synthetic tests are designed to trigger.

            So it’s not that I don’t believe those forum posts you’re referring to, but I’d like to see some better quantified cases where it does happen.

            • Krogoth
            • 5 years ago

            The problem was that sites that did the first wave of 970 reviews with time-framing mertic never push the 970 hard enough where it was forced to use 3.5GiB-.4.0GiB. This is how the problem got passed unnoticed until a few curious geeks on their hands on 970 and wanted to see what it could do.

            • cobalt
            • 5 years ago

            Just to be clear, I’m not arguing with you, I’d just like some hard numbers and some exploration of the boundaries — e.g., from a site that does these types of measurement all the time. (I haven’t noticed any problems with my 970, but I’m guessing I’m not pushing those boundaries, so I’d like to know where those boundaries are, and what the actual effect is once you hit them.)

        • Kougar
        • 5 years ago

        I disagree. 4K performance results from before the disclosure of the incorrect specs haven’t changed either.

        I have yet to see anyone conclusively show settings that prove there is a real performance hit beyond what was expected from a cut down 980, particularly when the 980/970 are within 5-10 FPS of each other.

          • Krogoth
          • 5 years ago

          It is because the problem doesn’t really impact avg FPS charts just like SLI/CF don’t show micro-shuddering despite the fact it generates a high avg FPS.

          If you threw the same cards under a frame-timing metric while using more then 3.5GiB of VRAM. The 970’s memory issue becomes painfully apparent. You can immediately tell a clear difference from a 980 and 970 with the same game and conditions and is not just from the avg FPS being lower.

          The cause is pretty simple. The 970 is forced to poll from the 512MiB partition but in order to access that part of it. It needs to take a massive detour on the core logic and this causes the rest of the silicon to stop for a tiny portion of a second in order to read/write on that segment.

          Ideally, this part of memory space should be reserved for frame buffer by the OS which doesn’t need shading power or memory bandwidth of rest of the chip. The problem is that you need the software and drivers to be coded for this when the current software ecology was build with the assumption that memory allocation space on GPU is continuous not split up into partitions. This makes 970 behave more like SLI/CF in this situation.

            • MadManOriginal
            • 5 years ago

            Yes, there is microstuttering in cases when this memory segment is used. But you continue to miss the point that at such settings, the performance of both the GTX 970 and GTX 980 is extremely low – 30 FPS or lower. Even if the GTX 980 is smoother, but still delivering such low framerates, no one is going to use either card at such low framerates.

            • Krogoth
            • 5 years ago

            It is a night and day difference at such framerates.

            Micro-stuttering has always been a big problem with AFR. You can mask most of the affects if the framerate is high enough (60FPS+).

      • Alexko
      • 5 years ago

      It’s not that simple. Yes, the reviews still hold, but users could expect graceful performance scaling for games that require up to 4GB of memory. In practice, that will only work up to 3.5GB.

      There weren’t many such games when the card was launched and reviewed, but they’re starting to show up and there will be more in the future, as Huang himself stated: “We wanted GTX 970 to have 4GB of memory, as games are using more memory than ever.”

      As for NVIDIA making a mistake, perhaps. But I’m not inclined to believe that no one in the company noticed it before it surfaced on Internet forums, so they were content to keep misleading people instead of coming clean.

        • MadManOriginal
        • 5 years ago

        When the last 512MB comes into play, the performance is low anyway due to other reasons – the GTX 980 also performs slowly in those cases.

        From a practical perspective, while there might be microstuttering, let’s be real – the 970 costs 60% of a 980 and so still represents a great value.

          • Alexko
          • 5 years ago

          The GTX 680 doesn’t suffer as much, and the difference can potentially be very perceptible if a game needs between 3.5GB and 4.0GB.

          Sure, the GTX 670 is still a good deal, just less so than NVIDIA’s specifications initially indicated. And that’s not nothing.

      • Firestarter
      • 5 years ago

      [quote<]they did not mislead anyone in terms of the card's performance[/quote<] That is not completely true. Yes, the performance in practice is just as it was at launch or even better and it sells mainly on that premise, but they also offer specifications of the card from which [i<]theoretical[/i<] performance can be calculated, as publications like TR often do. Based on those theoretical numbers, you can start making assumptions about the performance of the graphics cards relative to others based on the same architecture, which is exactly what TR did in their [url=https://techreport.com/review/27067/nvidia-geforce-gtx-980-and-970-graphics-cards-reviewed/3<]GTX970/980 review[/url<]. However, the GTX970 was so much slower than they expected given their identical RAM specs, that TR even did [url=https://techreport.com/blog/27143/here-another-reason-the-geforce-gtx-970-is-slower-than-the-gtx-980<]another article[/url<] on it that still failed to pick up on this. If a publication like TR fails to identify the cause of this discrepancy between theoretical performance and that of the practice, along with the other specialist hardware reviewers, how is a regular Joe Schmoe ever going to pick up on it? Yet it it Joe Schmoe, more than most of us, who looks at the packaging of this GTX970 while he's at Best Buy and sees that it has as much RAM as the GTX980 with the same specified speed, and concludes that it must be almost as fast as its way more expensive brother. I still think Nvidia made a fantastic product with 'Big Maxwell' and I'd still buy a GTX970 right now if I were in the market, but at the same time I think it was a stupid move to falsely advertise its memory characteristics, and I think it's absolutely idiotic that they haven't actually redacted that yet. That their official website STILL says that the card has 224GB/s of bandwidth is a farce.

    • Flying Fox
    • 5 years ago

    This mini mea culpa is paving the way of a quick settlement. They will shell out a few millions, possibly split over the next few quarters as one-time charges, knock a cent or two from their EPSes, and we all move on.

      • Firestarter
      • 5 years ago

      they can’t do one time charges without violating AMDs patent

    • maxxcool
    • 5 years ago

    Sooo I just read “It is as designed.. you like it. but it also wasn’t our fault.. But we won’t let it happen again.”

    ….. um ..

      • Ninjitsu
      • 5 years ago

      fault = lack of communication.

      • rhysl
      • 5 years ago

      “Aint no apology round here !!!”

Pin It on Pinterest

Share This