Nvidia: the GeForce GTX 970 works exactly as intended

As we noted in a story this past weekend, some curious and technically minded owners of GeForce GTX 970 graphics cards have noticed unexpected behavior with regard to that card’s use of memory.

Users first noticed that the GTX 970 appears to allocate less of its available memory than the GeForce GTX 980 does, despite the fact that both cards have 4GB of onboard RAM. Specifically, the GTX 970 doesn’t seem to make use of its last 512MB of memory as often as the 980 does. Users then found, using directed tests built with GPU computing development tools, that the GTX 970 can access that last chunk of onboard RAM, but only at much lower transfer rates.

The question was: why? We know that the GeForce GTX 970 and 980 are based on the same GPU silicon, the GM204 chip, but they have different configurations. All of the GM204’s graphics processing units are enabled on the GTX 980, while three of the chip’s 16 shader multiprocessor (SM) units are disabled on the GTX 970. Since not every chip comes out of the fab perfect, chip companies often disable faulty portions of their chips and build cheaper products around them. Graphics processors tend to be massively parallel, so a GPU with a fraction of its units turned off could still power a viable and compelling product.

Could it be that the way Nvidia disabled portions of the GTX 970 was causing its unusual memory access behavior?

This story brewed for a while until Nvidia released a statement this past Saturday with a brief explanation of the issue. The statement read, in part:

The GeForce GTX 970 is equipped with 4GB of dedicated graphics memory.  However the 970 has a different configuration of SMs than the 980, and fewer crossbar resources to the memory system. To optimally manage memory traffic in this configuration, we segment graphics memory into a 3.5GB section and a 0.5GB section.  The GPU has higher priority access to the 3.5GB section.  When a game needs less than 3.5GB of video memory per draw command then it will only access the first partition, and 3rd party applications that measure memory usage will report 3.5GB of memory in use on GTX 970, but may report more for GTX 980 if there is more memory used by other commands.  When a game requires more than 3.5GB of memory then we use both segments.

The statement then went on to explain that the overall performance impact of accessing that last half-gigabyte section of RAM ought to be fairly minor. Nvidia provided a few examples comparing performance in RAM-constrained scenarios versus the GTX 980.

This revelation touched off a storm of discussion and speculation in the comments to our story and elsewhere, with folks wondering whether the GTX 970 was somehow broken or subject to a hardware bug.

To clear the air, Nvidia Senior VP of Hardware Engineering Jonah Alben spoke with us yesterday evening. Alben’s primary message to us was straightforward. He said the GTX 970 is “working exactly as we designed it.” He assured us the GTX 970 does have a full 4GB of memory and claimed “we do use it when we need to.”

Alben then explained that the GTX 970’s unusual memory access behavior is a consequence of a new feature built into Maxwell-class GPUs. That feature has to do with how Nvidia disables faulty portions of its chips when needed. Alben said the feature allows Nvidia to make a better product than it could have otherwise.

To help us understand this feature and how it impacts the GeForce GTX 970, Alben took us on a brief tour of the guts of the GM204 GPU. That tour relies pretty heavily on a simplified diagram of the chip that Alben provided us, which we’ve embedded below.

Across the top of the diagram are the shader multiprocessors or SMs, where most of the graphics computational work happens. The bottom half of the diagram shows the chip’s memory config. The GPU has four memory partitions, each with two 32-bit links to external DRAM chips. Those memory partitions are split into two chunks, each with its own section of L2 cache, memory controller and so on.

The middle of the diagram depicts the crossbar that facilitates communication between the shader and memory arrays. You can think of this crossbar as a switched fabric, much like an Ethernet switch, that allows any SM to talk with any L2 cache and memory controller.

Since the diagram depicts a GTX 970, three of the chip’s SMs have been grayed out, along with one of the two L2 cache sections in the memory partition on the far right. The GTX 980 has all of these units enabled.

In the prior generation of Kepler-derived GPUs, Alben explained, any chips with faulty portions of L2 cache would need to have an entire memory partition disabled. For example, the GeForce GTX 660 Ti is based on a GK104 chip with several SMs and an entire memory partition inactive, so it has an aggregate 192-bit connection to memory, down 64 bits from the full chip’s capabilities.

Nvidia’s engineers built a new feature into Maxwell that allows the company to make fuller use of a less-than-perfect chip. In the event that a memory partition has a bad section of L2 cache, the firm can disable the bad section of cache. The remaining L2 cache in the memory partition can then service both memory controllers in the partition thanks to a “buddy interface” between the L2 and the memory controllers. That “buddy interface” is shown as active, in a dark, horizontal arrow, in the bottom right memory partition on the diagram. In the other three memory partitions, this arrow is grayed out because the “buddy” interface is not used.

Thanks to this provision, Nvidia is able to equip the GeForce GTX 970 with a full 256-bit memory interface and still ship it at an attractive price in high volumes. Nvidia still ships some chips with both L2s in a memory partition disabled, like the GeForce GTX 970M for laptops, but Alben said “we have much fewer of those now.” So Nvidia is keeping more hardware functional on more chips thanks to this optimization. Still, this GPU configuration has some consequences we didn’t understand entirely when the card was first released.

For one, the GTX 970 lacks some of the cache capacity and ROP throughput we initially believed it had. Each L2 cache section on the GM204 has an associated ROP partition responsible for blending fragments into pixels and helping with multisampled antialiasing. With one of its L2s disabled, the GTX 970 has only 56 pixels per clock of ROP throughput, not the 64 pixels per clock of ROP throughput specified in the card’s initial specs sheets. (In an even crazier reality, that limit isn’t even the primary fill rate constraint in this product, since the GTX 970’s shader arrays can only send 52 pixels per clock onto the crossbar.) Also, the GTX 970’s total L2 cache capacity is 1792KB, not 2048KB, since one 256KB cache section is disabled.

Alben frankly admitted to us that Nvidia “screwed up” in communicating the GTX 970’s specifications in the reviewer’s guide it supplied to the press. He said that the team responsible for this document wasn’t aware that the GTX 970 takes advantage of Maxwell’s ability to disable selectively a portion of L2 cache.

More intriguing is the question of how well the GeForce GTX 970 makes use of its available memory. That question has two related dimensions, bandwidth and capacity.

To understand the bandwidth issue, first notice a constraint imposed by the disabling of an L2 cache section in the diagram above: the crossbar link into that section of L2 is also disabled. Although the GTX 970’s path to memory remains 256 bits wide in total, the width of the crossbar link is reduced. The two memory controllers in that partition must share a single crossbar link, backed by a single L2 cache. Access to the half-gigabyte of DRAM behind at least one of those two memory controllers will be slower at peak than elsewhere in the system. That 512MB of memory could still be potentially useful, but it’s a bit problematic. Nvidia must work around this limitation.

Now, consider how a full-fledged GM204 like the GTX 980 actually takes advantage of all of the memory bandwidth available to it. If it were to store its data contiguously in one big blob, the GM204 could only ever read or write data at the speed of a single DRAM chip or crossbar link. Instead, to achieve its full bandwidth potential, the GPU must distribute the data it stores across multiple DRAMs, so it can read or write to them all simultaneously. Alben’s diagram above indicates the GM204 has a 1KB stride. In other words, it stores 1KB of data in the first DRAM, then stores another 1KB in the next DRAM, and so on across the array. On the GTX 980, the chip strides across eight DRAMs and then wraps back around. Operations that read or write data sequentially should take advantage of all eight memory channels and achieve something close to the GPU’s peak transfer rate.

For the GTX 970, the Maxwell team had to get creative in order to prevent the slower half-gig of DRAM from becoming a problem. Their answer was to split the GTX 970’s memory into two segments: a large, fast 3.5GB segment and a smaller, slower 512MB segment. These two segments are handled very differently. The 3.5GB segment includes seven memory controllers, and the GPU strides across all seven of them equally. When the GPU is accessing this segment of memory, it should achieve 7/8th of its peak potential bandwidth, not far at all from what one would see with a fully enabled GM204. However, transfer rates for that last 512MB of memory will be much slower, at 1/8th the card’s total potential.

So, as Alben explains it, Nvidia’s hardware config is intended to behave like some GTX 970 owners have measured: fast in the first 3.5GB memory and slower after that.

With a conventional GPU config like the GTX 980, Alben notes, Nvidia reports two memory segments to the operating system: the GPU’s own RAM and the additional system memory available over PCI Express. With the GTX 970, Nvidia reports two separate segments of GPU memory, first the faster 3.5GB chunk and then the slower 512MB one, along with some “hints” telling the OS to prefer the larger, faster segment where possible. As a result, the OS should treat the GTX 970’s memory hierarchically with the correct preference: first the faster segment, then the slower segment. Should the application’s memory needs exceed 4GB in total, it will spill into PCIe memory, just as it would on the GTX 980.

Incidentally, Alben told us that this arrangement helps explain the behavior some folks have noted where the GTX 980 appears to use more of its total memory capacity than the GTX 970 does. Some of the data stored in video RAM during normal operation falls into a gray area: it’s been used at some point, but hasn’t recently been accessed and may not be used again. That data’s presence in RAM isn’t strictly necessary for the current work being done, but it could prove useful at some point in the future. Rather than evict this data from memory, the GPU will keep it there if that space in RAM isn’t otherwise needed. On the GTX 980, with a single memory segment, this “cold” data won’t be ejected until memory use approaches the 4GB limit. On the GTX 970, this same cold data is ejected at 3.5GB. Thus, when doing the same work, the GTX 970 may use less of its total RAM capacity as a matter of course than the GTX 980 does. Again, according to Alben, this behavior is part of Nvidia’s design. The GTX 970 still has access to the full 4GB of RAM if it’s strictly needed.

At this point, I had a simple question for Alben: would it have been better just to make the GTX 970 a 3.5GB card? After all, having that slow half-gig of RAM onboard seems a little iffy, doesn’t it? His response: I don’t think so, because that half-gig of memory is useful. You’re better off spilling into the final 512MB of onboard memory than spilling over to PCI Express, which is even slower still.

Also, Alben noted, with “good heursitics,” Nvidia is able to put data that’s not likely to be used as often into this half-gig segment. In other words, Nvidia’s driver developers may already be optimizing the way their software stores data in the GTX 970’s upper memory segment.

One of the big questions about all of this madness is what happens when the GTX 970 does have to spill into its last half-gig of memory. How much of a performance slowdown does one actually see compared to a GTX 980 based on a full-fledged GM204? So far, the firm has offered some benchmarks showing performance deltas from about one to three percent. In other words, when stepping up into a scenario where substantially more than 3.5GB of VRAM is used, the GTX 970 suffers very slightly more than the GTX 980 does.

Alben told us Nvidia continues to look into possible situations where the performance drop-offs are larger on the GTX 970, and he suggested that in those cases, the company will “see if we can improve the heuristics.” In short, Nvidia is taking responsiblity for managing the GTX 970’s funny VRAM config, and it’s possible that any problems that users turn up could be worked around via a driver update. In the end, that means the GTX 970’s performance may be a little more fragile than the 980’s, but Nvidia has a pretty good track record of staying on top of these kinds of things. This isn’t nearly the chore that, say, maintaining SLI profiles must be.

As our talk ended, Alben reiterated that the Maxwell team is pleased with the GTX 970 in its current form. “We’re proud of what we built, think it’s a great product. We thought this feature would make it a better product, and we think we achieved that goal. We want to make sure people understand it well.” 

Comments closed
    • Elash
    • 5 years ago

    I want to buy a MSI 970 GTX next week to be able to play Witcher 3 on High/Ultra at 1920×1080 and then i saw this problem with memory issue and i have this question; there is gonna be a problem running Witcher3, Arkham Knight,Mad Max on High/Ultra at 1920×1080 or this issue is gonna pop only if its used 2k/4k resolutions ?

      • NovusBogus
      • 5 years ago

      I’ve been poring over these old articles lately for the same reason; Geralt Time approaches and we’re gonna need a bigger boat. My general conclusions are as follows:

      -For Witcher 3 and Arkham, flagship titles that are shipping with the card, the probability of not getting awesome performance at any reasonable resolution is effectively zero. Nvidia would not allow this to happen, period.

      -For most 1080p games, even 1200 (my display though I usually run lower rez for UI reasons) and probably 1440, this is a non-issue as you won’t be using anywhere near 3GB of video memory. People getting this are almost exclusively running 4K or triple-1080p with maximum MSAA (notorious VRAM hog).

      -For future big-ticket games like Star Citizen that might want >3GB for giggles, it could be rough at first but the driver team should sort out any issues.

      -That leaves older and/or obscure games that use >3GB due to either high display resolution or excessive texture data. High res shouldn’t be an issue simply because GPU throughput will be the limiting factor; a 970 isn’t made for 4K gaming and it has nothing to do with memory bandwidth. Big textures (heavily modded Skyrim is the canonical example) are more concerning, this is the one area where I’d say a typical user might have to worry.

      Basically what we have here is AMD style die-harvesting, only instead of outright disabling the extra hardware they left some of it in play and we’re supposed to trust that the drivers will do what they’re supposed to and put everything in the right bucket. I don’t particularly like this, paranoiac that I am, but I also see little evidence that it’s an actual problem at normal settings.

    • investigator
    • 5 years ago

    I’m investigating a class action against NVidia for the GeForce. Please contact me to find out what you can do if you have one of these and now realize there is an issue. JLD@kbklawyers.com.

    • sschaem
    • 5 years ago

    Not to beat a dead horse…

    But I I recall in the past reviews would sometime include image quality difference,
    specially with new architectures. To check for driver speed tricks, but also HW quality.

    Now Dx11 is supposed to reduce this greatly, but maybe its time revisiting soon?

    Like what is the visual quality, not just performance, difference between a 2GB and 4GB card
    at 1080p and up with upcoming next gen console ‘ports’.

      • Meadows
      • 5 years ago

      Why would that be an issue? Image quality tests were dropped some time after it was noticed that the strict compliance required by Dx10 pretty much eliminated the differences between vendors. There’s been little point in testing it since.

        • VincentHanna
        • 5 years ago

        I’d be interested in re-confirming that after DX12 launches.

    • Tirk
    • 5 years ago

    [url<]http://a.disquscdn.com/get?url=http%3A%2F%2Fs30.postimg.org%2Fbgazih5e9%2Fimage.jpg&key=77yM5ONh0ydG7oeXGdK_3g&w=800&h=412[/url<] I thought this was hilarious when I saw it. Please take this as a joke with no seriousness behind it, but its uncanny how the tech culture can get so polarized over company's products.

    • LoneWolf15
    • 5 years ago

    Ever have someone lie to you, and say “I was worried you’d be upset if I told the truth?”

    When that happens, the answer is usually the same. Sure, maybe I’d have been upset –but I’d have respected the truth a whole lot more than a lie. Maybe the lie itself didn’t affect the bottom line, but if affects my trust of the person telling it.

    That’s kind of how I feel about this. Maybe nVidia’s design has all of the advertised performance, but not giving full disclosure, or being up-front (they claim it’s a mistake, but hey, Maxwell’s been out four months now) affects my ability to trust them, and their next product, and the product after that.

    Maybe we should treat hardware like we treat games from Ubisoft or EA, and wait six months from release, to be really sure there won’t be any surprises.

    • Wild Thing
    • 5 years ago

    The Green team is reeling…
    [url<]http://gamenab.net/2015/01/26/truth-about-the-g-sync-marketing-module-nvidia-using-vesa-adaptive-sync-technology-freesync/[/url<]

    • Steele
    • 5 years ago

    As someone who just bought a 970, I’m feeling a bit concerned.

    The content of the article? Meh. I can live with that. The card works great, and until I have the opportunity to really push the thing to its limits (GTA5?), I doubt I’ll ever see a problem.

    However, some of the comments in this thread have me worried. From what I’ve gathered, what i REALLY have is a 980 with damaged components? I’m imagining that there’s a slight chance that something got overlooked and I’m due for a catastrophic failure any day now.

    This comes from not understanding the nature of this “damage”. Like, if I were to pry the heatsink off, would I see physical damage to some of the parts, or is it simply that some of those components, for whatever reason, just don’t hold as much charge or data as they’re supposed to?

    How do I know that -my- 970 doesn’t have more deactivated portions than the next person? Is there an upper limit to the number of defective parts before the manufacturer would just junk the card? The article seems to indicate three… what if my card only had one damaged part, could I get the other two to work?

    Over all, if this card was supposed to be a 980, but failed a few tests, how can I be assured that its failures won’t continue to mount?

      • the
      • 5 years ago

      This is a quiet normal thing to do in the industry. The functional units disabled did not past testing due to manufacturing defects. When dealing with lithography that produces patterns in silicon whose width can be measured in [i<]atoms[/i<], it is common to have something be a little off. For example, a particle of dust falls on to the silicon wafer during the lithography process and prevents a circuit from being completed in the design. Instead of throwing the whole chip away, manufacturers will disable a functional unit where that defect resides. There is no risk that an individual unit won't fail later on for the same reasons a unit doesn't pass testing in manufacturing. Products have a minimum specification. The GTX 970 will always have 13 of its 16 SMX clusters enabled. Even if only one of the SMX clusters doesn't past testing, two additional ones will be chosen to be disable to fit the GTX 970's specification. There are no known means of enabling these disabled units once purchased in this case. However, even if you were able to enable these units, there is no means of actually knowing if they were disabled due to a genuine manufacturing defect or if they were cut to make the chip fall into the GTX 970s specification. (Certain products have been able to have their functional units enabled. Certain Radeon 9500 Pros could be modded into Radeon 9700 Pros. More recently Athlon X3 could have one CPU core enabled to become Athlon X4s. As noted above, these mods didn't always work as there still could be manufacturing defects.) As for when this practice started, Intel has been doing it since the 486 from decades past: the original 486SX chips were actually 486DX chips with a defective floating point unit. More recent examples are nVidia's GTX 480 which was such a large and complex chip, it never shipped in a 100% enabled version (its successor, the GTX 580 however did). The infamous Cell chip in the Playstation 3 is manufactured with eight SPE units but only 7 are enabled due to so few chips being perfect.

        • Steele
        • 5 years ago

        Ah, thanks for that! The idea that this sort of thing is standard practice is disconcerting; however, I can see that it would be prohibitively expensive and wasteful otherwise.

        I’ll keep this all in mind for my next upgrade in hopefully no less than four or five years from now!

          • plonk420
          • 5 years ago

          they do this with CPUs, too

          tho mostly it has to do with the power and speeds the processor can maintain. a slower CPU of a series is a “less pure” set of hardware than the high end (say, i7-4770K or faster)

          AMD did this with the X3 and some X2s of the Phenom II series.

      • MadManOriginal
      • 5 years ago

      Don’t take this wrong but…if you think you can see deactivated parts by looking at a chip, don’t worry about it and just enjoy playing games.

      • Krogoth
      • 5 years ago

      970 will work fine for the most part. The problem only arises if you attempt to throw it demanding applications that start to access the effect memory blocked (3.5-4.0GiB) range in the form of shuddering.

      970 were 980 chips (GM206) that have a problem with a few SMM blocks but most of the silicon was still viable. It was sold as a lower-binned GPU at a price point that reflected this. This is a standard practice through-out the semiconductor industry.

      The main beef is that Nvidia didn’t accurately reflect the actual impact and stats of disabled parts. It looks like even tried to *hide* it, but were caught and forced to give an explanation.

    • Wildchild
    • 5 years ago

    I’m kinda surprised by some of these comments. The simple fact is this – Nvidia deliberately lied to their customers and then tried to say it’s a “feature.” How incredibly insulting and pompous.

    If you aren’t bothered by this in some way, then you need to take your fanboy goggles off.

      • morphine
      • 5 years ago

      And yet, the cards perform just fine.

        • Krogoth
        • 5 years ago

        For most stuff, yes that’s quite true. However, when you are trying to push 970 to its limits then the minor issues from this problem start to appear (AFR-like micro-shuddering).

        What is bothering 970 users is how Nvidia’s marketing handled this. They chose to omit the problem despite knowing about it from the start and pretend it never happened until third-parties found it. Nvidia was forced to explain it. It is not the first nor will be the last time a company try using this “marketing tactic”.

          • morphine
          • 5 years ago

          While I agree that Nv’s marketing could have handled this better…

          There’s really no “push the 970 to its limits”. In any video card, you never get to actually make specific use of all of its VRAM because you need some left for framebuffers, AA buffers, Windows’ DWM, etc. That’s what irks me – most of this brouhaha was raised by people who fail to understand that basic concept, and now we get some misdirected anger and fanboyism claiming that the 970 has 3.5GB of VRAM after all. Sheesh.

            • Krogoth
            • 5 years ago

            The problem is that games that demand 3GiB of VRAM don’t know about the said problems are completely depended on the drivers to map the texture data. Ideally, 970 should only be mapping frame-buffer to the affected portion of the memory where the performance penalty would negatable. This isn’t always the case and there are already been reports from 970 users of “AFR-like shuddering” when the card gets pushed hard a.k.a using more than 3GiB of VRAM. It is no doubt due to the texture data being mapped into the bad sections.

            1-4% performance drop that Nvidia published doesn’t tell the whole story like how SLI/CF avg FPS figures don’t tell the whole story.

            • morphine
            • 5 years ago

            If that comes to be a significant issue (which I doubt but allow for the possibility), it’s not guaranteed but very likely that the driver team will fix it relatively easily.

            • Krogoth
            • 5 years ago

            So, the 970 ends-up being like SLI where its performance is tied to the drivers and games in question.

            • Voldenuit
            • 5 years ago

            [quote<]So, the 970 ends-up being like SLI where its performance is tied to the drivers and games in question.[/quote<] So, just like every other video card in the past decade (if not longer), then?

            • Krogoth
            • 5 years ago

            It is far more dependent in this instance just like SLI/CF profiles.

          • Tirk
          • 5 years ago

          If 970 owners are getting AFR-like micro-stuttering while using this card because of this issue, why are people shrugging it off as no big deal? We have a whole new metric in frame time because a 1-3% drop in fps numbers means nothing if there is obvious frame time issues. I can somewhat shrug off the issue if its only producible by reviewers but if actual owners are experiencing an issue with frame times when using the last .5 gib of vram which is advertised as part of the card’s spec then it doesn’t seem to be a non issue.

          Why not just show fps numbers if frame stuttering shouldn’t be a concern anymore? We have the tools now, a more in-depth look into how much micro-stuttering occurs when using the last .5 gb of the 970’s vram would probably be useful for prospective buyers of the product. Does the stuttering increase when put in sli?

            • Krogoth
            • 5 years ago

            Interesting enough, the shuddering issues with 970 under high-VRAM usage is why enthusiast took a closer and they were able to discover this little problem.

      • Klimax
      • 5 years ago

      Lied? Nope. The only place with incorrect information is reviewers guide, nowhere else. (usually other places have only high level sparse information) And you are twisting things beyond any recognition including qutoes where they have no business.

      I would strongly suggest you shouldn’t talk about “fanboy goggles”, hypocrisy is ugly thing…

      ETA: Reminder: If this were to be accepted as lying then AMD is in very big troubles… (just whole bulldozer fiasco would be giant hole there and quite few statements like “there is no such thing as next version of DirectX” would be perfect candidates as well)

        • Wildchild
        • 5 years ago

        Everything that has been said has been said already. The only thing I’ll rebuttal is

        [quote<] I would strongly suggest you shouldn't talk about "fanboy goggles", hypocrisy is ugly thing...[/quote<] I own a GTX 780, but thanks for the assumption. Just because I call out a companies BS doesn't mean I blindly follow its direct competition. Brand loyalty is bad, mmkay?

          • Klimax
          • 5 years ago

          You’re right. So far nothing more then a mistake in a pdf…

        • sschaem
        • 5 years ago

        Have you even bothered to check nvidia web site?

        The 970 and 980 are advertised as having the exact same memory performance and bus.
        256bit and 224GB rated.

        But no. The 970 does not have a 4GB 256bit bus, but a 3.5GB 224bit + a separate .5GB on 32bit.

        Both cannot be concurrently used and in turn CANNOT be labeled a 224GB bandwidth card like the 980.

        The 970 always access memory using 224bit.

        And this info was known to nvidia from day one , an errata that driver writer had to follow.
        Hence the first notice of the 3.5GB limit.

          • Klimax
          • 5 years ago

          Memory bus is all same. What is stated? theoretical bandwidth. It is correct. Simple multiplication. (And frankly, it is not really useful unless you know technical details anyway)

          Alias definition of that never changed… and thus specification as written is correct. (Silicon is there!)

          ETA: Nothing more to that…
          ETA2: Entire post of yours is not really relevant. And anybody buying purely on high level spec is frankly idiot.

    • tbone8ty
    • 5 years ago

    oh my god this is the end of Nvidia. such a horrible company. they cant do anything right. there drivers are horrible. I’ve had Nvidia all my life. Never thought id see the day but, now I’m switching to AMD.

    it goes something like that right? did i do it correctly fanboy master?

    • Ninjitsu
    • 5 years ago

    Lol at least it’s official word that this is happening by design, and not because “oh it’s a bug and you can’t simulate everything!”.

    Interesting information, though. Maybe they drop the prices a bit as a reconciliatory gesture. 😛

      • BlackDove
      • 5 years ago

      I think the 970 came out at a pretty cheap price to begin with.

      I also think its really stupid to waste engineering resources to do this, and if the bandwidth of the last .5GB is much slower they should have just made the 970 a 3.5GB card. That is, if the extra stuff there was just to keep the “256bit bus” and “4GB” specs on paper.

      Seems like a waste since the 970 is lower performance than a 780ti which is just fine with 3GB. I dont think having the extra .5GB is going to let you play games on ultra at 4K.

      However, i think not being able to get the right color space over HDMI at TV resolutions is a bigger deal than this. People have been complaining for years about that and Nvidia has done nothing.

      That affects many people and in a much worse way.

        • Ninjitsu
        • 5 years ago

        Much slower performance = 3% when the frame rate is already below 60 fps. Translation: the actual difference is not more than 3 fps, in what already are fringe cases.

        Then, 4GB is more forward looking, considering that games seem to assume 4GB VRAM these days, with the new consoles.

        The only issue in all of this is the lack of communication on Nvidia’s part.

        I have an Nvidia card, and the colour space is fine on my monitor. The bug you describe hasn’t got anything to do with TV resolutions, just with the TV itself (EDIT: In the sense that, the TV part is the problem, not the resolutions). Anyway, while a valid complaint, it’s unrelated to this issue.

          • BlackDove
          • 5 years ago

          No it has to do with the HDMI and Display Port standards calling for RGB 16-235 for “TV resolutions” like 1280×720, 1920×1080 and probably 3840×2160.

          A computer monitor needs an RGB 0-255 signal. If you have one of those resolutions and are using HDMI or Display Port, you have the issue.

          Like i said, this issue affects a bunch of people who dont even know they have it and no amount of calibration will fix it.

          Theres a driver hack or setting the monitor to RGB 16-235 as a “fix”.

            • Ninjitsu
            • 5 years ago

            Yeah I’m aware of the issue.The actual problem is with the driver detecting the monitor as an HDTV.

            I’ve verified and checked, I get the full colour range, and my monitor is detected as a monitor, not an HDTV.

            (If it means anything, I do image processing stuff too, so I’d notice if RGB values were a narrower range).

            EDIT: Again, I’m not sure why you’re bringing this up in the context of the 970’s memory distribution and performance.

            • BlackDove
            • 5 years ago

            Because its a bigger issue and what kind of monitor do you use with a 1920×1080 resolution and HDMI or Display Port but no DVI thats good enough for image processing?

            • Ninjitsu
            • 5 years ago

            Bigger only because *this* issue itself doesn’t really affect many people at all. You should be bringing it up whenever a new driver is launched, or submit a bug report to Nvidia.

            I use a Dell S2230L, over HDMI.

    • Dysthymia
    • 5 years ago

    I lost a little bit of respect for Nvidia here for misrepresenting the 970’s capabilities, but I understand how being 100% honest from the start could have put people off with the unusual memory configuration (which probably would have impacted sales).

    I should have known that for a card segmented one tier down from the top, ALL specs being the same except one was too good to be true. In any case, I still think I made the right call with a 970. Though honestly as much as I game these days and as much trouble as I’m having trying to fold, my next card might be AMD.

      • derFunkenstein
      • 5 years ago

      On the contrary, I think if they were a little more up-front about it, people would see “wow it’s handicapped to an extent and it still performs like this, this is really great”. The fact is, if AMD wasn’t running a Crazy Eddie’s Radeon Clearance Sale, it’d sell even better than it is.

        • sweatshopking
        • 5 years ago

        yes, but thank goodness amd IS running crazy eddies.

    • MadManOriginal
    • 5 years ago

    The NVidia hate is still strong, even after a lengthy explanation. I didn’t know so many computer nerds were losers who identify with the less successful company (AMD).

      • Platedslicer
      • 5 years ago

      Yes, questioning the top dog is surely a sign of a longstanding, unresolved loser complex that automatically invalidates any point a person might make.

      Edit: sarcasm warning for the sarcasm impaired

      • Krogoth
      • 5 years ago

      It is not Nvidia hate. It is being annoyed by shady marketing tactics and die-hard fanboys/shameless shills try to defend to the death.

      FYI, I have bought/used cards from AMD/ATI, Nvidia, 3Dfx and Matrox. If AMD/ATI, Intel or somebody else tried pulling similar non-sense I would also throw them into the hot seat.

      • chuckula
      • 5 years ago

      It is interesting that Fangate — where there actually were measurable differences between reviewers’ AMD cards and retail AMD cards — was loudly shouted down as ‘no big deal’ while a numeric error in technical documentation from Nvidia that has literally zero effect on the performance of the GTX-970 is now apocalyptic.

        • Krogoth
        • 5 years ago

        It has an impact on performance when the 970 utilizes 3GiB or more VRAM in the form of micro-shuddering. GPU has to play hot potato with the good and bad memory segments.

        • Platedslicer
        • 5 years ago

        “Fangate” wasn’t shouted down, except in the minds of particularly paranoid nVidia fanboys. Despite the best efforts of AMD fanboys, no amount of shouting could drown out the noise of the crappy reference coolers.

        AMD paid for their mistake with a bad rep for otherwise excellent cards. A lot of review sites refrained from recommending them for being so hot and loud – including Anandtech, which was running a sponsored “AMD Center” page back then. I imagine that must have strained the relationship.

    • Klimax
    • 5 years ago

    This type of issue can highlight problems with some websites. Take article by Techreport or PCPer and you see neutral reporting on things. Just technical matter and little to no commentary/blog-like content. In contrast you get Anadtech, were you get no less then two “damage controls” and other blog-like crap interspersed with technical content starching whole thing to four page sub-average article. (DC usually is used as poisoning the well. Rarely is used correctly…)

    ==

    Also for future reference I found some interesting things when reading original response by NVidia. Some interesting papers or articles…

    [url<]http://citeseerx.ist.psu.edu/viewdoc/download?doi=...[/url<] "Butterfly networks offer minimal hop count for a given router radix while having no path diversity and requiring very long wires. A crossbar interconnect can be seen as a 1-stage butterfly and scales quadratically in area as the number of ports increase." [url<]http://web.eecs.umich.edu/~twenisch/papers/ispass1...[/url<] GT200: [url<]http://www.realworldtech.com/gt200/10/[/url<] "Loads are then issued across a whole warp and sent over the intra-chip crossbar bus to the GDDR3 memory controller. Store instructions are handled in a similar manner, first addresses are calculated and then the stores are sent across the intra-chip crossbar to the ROP units and then to the GDDR3 memory controller."

    • jessterman21
    • 5 years ago

    [quote<]One of the big questions about all of this madness[/quote<] Madness... ? THIS IS NVIDIA!!!

    • chuckula
    • 5 years ago

    In other news, it turns out that Intel has been LYING the whole time!
    The 4770K? Not really a 4-core part. It’ just 2 cores with some ginned up overclocking and faked interfaces to fool you into THINKING it’s 4 cores. All the benchmarks were perfectly valid though….

    AMD: VICTORY!! INTEL WAS LYING THE WHOLE TIME!!

    AMD (about 5 seconds later): Uh wait.. you mean we were really losing to a 2 core processor this whole time…. uh… Intel sucks nevermind.

    • Kougar
    • 5 years ago

    Glad to read the official explanation and the reasoning behind it. It makes a lot more sense once it is explained out.

    It sounds like next time the product team that works with reviewers will be briefed on such functionality and the documentation will be more carefully screened as well which is always a good thing…

      • JustAnEngineer
      • 5 years ago

      NVidia’s Evil Marketing Geniuses were entirely too busy trying to educate reviewers into reporting that the most important measurement for a graphics card had suddenly changed to “performance per watt” instead of “performance per dollar.”

        • auxy
        • 5 years ago

        I feel like this may have been in jest, but this really is a serious frustration of mine (and many others.) I don’t CARE about the performance per watt of a desktop GPU; it’s NOT a concern of mine and frankly lauding a company over and over when they release a product for the third time with basically the same GPU performance (GTX 690, GTX 780Ti, GTX 980) just because it runs cooler and uses less power is irritating to say the least. Where are the gains? ( ;∀;) If you improve efficiency, that’s not an excuse to reduce complexity and maintain performance…

        • derFunkenstein
        • 5 years ago

        What do you expect? They’re trying to break into mobile in a big way using Maxwell.

          • JustAnEngineer
          • 5 years ago

          I expect that the evil marketing geniuses asked the engineering types where their new chip was better than the competition’s and whatever that answer was suddenly became [b<]the most[/b<] important thing to be communicated to reviewers and NVidia focus group evangelists.

            • Ryu Connor
            • 5 years ago

            I got 2006 on the line for you JAE.

            She’s here to remind you that laptops outsell desktops and that said gap has continued to grow over the last nine years. He says he won’t bore you with the smartphone and tablet talk either.

            Aren’t you the guy who loves the buggy whip speech Danny DeVito gives?

            • JustAnEngineer
            • 5 years ago

            I’m going to agree with Auxy on this one. There are two very obvious flaws in your argument.

            1) A high-end mobile gaming GPU in a gaming laptop doesn’t run from a battery for more than 20-30 minutes. You lug your gaming laptop from outlet to outlet and plug in before you start your game.

            2) These are reviews of hot-clocked STRIX / MATRIX / OMGWTFBBQ versions of desktop gaming graphics cards (which far too frequently get reported as if they represent the less expensive stock-clocked cards). These dual or triple slot PCIe graphics cards are installed in desktop PCs that have hundreds or even thousands of watts of power available and the cooling capabilities to match.

            If a reviewer were reviewing a mobile graphics solution that was normally run from a battery (Adreno/Imageon or Mali or another phone or tablet graphics solution), then yes, I would be very interested in power consumption.

            • Ryu Connor
            • 5 years ago

            What do your points have to do with anything?

            Are you seriously going to first argue that since current cards strain batteries they should stop trying to improve upon that?

            The second argument is just as facetious. Cause nobody likes tiny PCs? Oh, wait, you’re the guy always recommending mATX rigs.

            You got some real cognitive dissonance going on here.

            Do I need to have 2006 call you again? Do you love buggy whips or not? Cause right now, you definitely do.

    • homerdog
    • 5 years ago

    [quote<]With one of its L2s disabled, the GTX 970 has only 56 pixels per clock of ROP throughput, not the 64 pixels per clock of ROP throughput specified in the card's initial specs sheets. [b<](In an even crazier reality, that limit doesn't even matter in this product, since the GTX 970's shader arrays can only send 52 pixels per clock onto the crossbar.)[/b<][/quote<] Scott, I'm sure you know the bolded part isn't strictly true. Shadowmapping for example is ROP bound regardless of alu throughput.

      • Damage
      • 5 years ago

      Right. I’ve changed the text to be more specific: “that limit isn’t even the primary fill rate constraint in this product”. Thanks.

        • homerdog
        • 5 years ago

        You’re the best =)

    • Krogoth
    • 5 years ago

    Nvidia isn’t out of possible legal heat. Somebody in marketing is going have the “talk”.

    The difference between this and 660 debacle is that Nvidia’s marketing or even upper managment tried to brush it off. 660’s asynchronous memory issues were known from the beginning and were recorded in the white papers.

    It is akin to manufacturer releasing a product with a known issue into the marketplace without informing the customer until after the fact. It looks like they even tried to “hide” hoping nobody would notice or care. I doubt somebody in engineering/design is at fault here.

    It comes to no surprise that a rep from that department is saying that 970 is working as intended.

    This whole bloody thing could have been avoided if Nvidia have been forthcoming with this problem and marketed 970 as 4GiB* or 3.5GiB with a detailed explanation on why this is the case. 970 buyers (real victims of this) wouldn’t be raising hell over this.

    It is a good thing that curious enthusiast are able to find little problems like this.

    I doubt a recall will happen. It would more like a be a refund of some small amount to existing 970 users if a class action lawsuit does happen and Nvidia loses or makes a settlement.

    This whole thing reminds me of the whole Audigy 1/2 debacle with 24bit/96KHz on everything, but it actually only did it for output while inputs were downsampled to 16bit/44KHz. Creative didn’t bother mentioning this until after the fact.

      • torquer
      • 5 years ago

      If only you could turn being a professional contrarian into a marketable skill.

      This is the very picture of a tempest in a teacup.

        • Krogoth
        • 5 years ago

        The problem is how Nvidia marketing department’s execution on how to deal with this problem.

        They chose to “handwave it” and let it slide in hopes nobody would notice. They were caught red-handed and were forced to give an explanation.

        It is like Coca-Cola decide to quietly remove “50% sugar” out of their product with no notification to their customers or distributors. Customers later find out by chance through noticing that new batch “tasted” a little different and wanted to know why. Coca-Cola keep marketing that product contains the same amount of sugar as before.

        If Nvidia notify and told about everyone about the issue from the start. Nobody except the die-hard AMD/ATI fanboys would have blink an eye.

        If we are willing to accept this kind of sleazy marketing. It sets a precedent that allows Nvidia and other tech companies are allow similar crap to go through and they know that nobody will question or dare to challenge it.

        970 users are already speaking and are demanding refund/returns from their card vendors. It is only a matter of time before this goes up the chain and certain heads at Nvidia are going to roll.

          • BlackDove
          • 5 years ago

          The only way you can claim that Nvidia lied is if they lied about the memory bandwidth or cache size.

            • Firestarter
            • 5 years ago

            IMO they lied about the memory capacity because it’s clear that not all of it is as fast as it ought to be, based on the product description. Now if it were a tiny slowdown or tiny amount, nobody would care . However, roughly 1/8th (or is it actually 3/16th?) is not an insignificant amount, and the reduction in bandwidth is apparently big enough to warrant [i<]avoiding[/i<] it. If it's bad enough to avoid it by not including it in the main buffer, wouldn't that at least require a caveat in the product description, telling us that not all memory is created equal? Instead Nvidia's marketing didn't tell us because they thought it wouldn't matter and that we wouldn't find out. edit: even Nvidia's own product pages claim that the 970's bandwidth is as good as the 980's. Yeah you might say that the marketing department screwed up, but without a redaction that screw up definitely becomes false advertising.

            • Pwnstar
            • 5 years ago

            Capacity is rating of size and not speed. nVidia didn’t lie about the speed, they lied about the bus width and the cache.

            • puppetworx
            • 5 years ago

            [url=http://www.anandtech.com/show/8935/geforce-gtx-970-correcting-the-specs-exploring-memory-allocation/2<]Anandtech is claiming[/url<] that memory bandwidth is 196GB/s (best case) not the 224GB/s NVidia [url=http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-970/specifications<]claims[/url<]. Anandtech's claim is that the 0.5GB partition cannot be accessed concurrently with the 3.5GB partition. As far as I've seen Anandtech is the only site claiming this so far however. If true it's easy to see how wasting 7 times as many clocks accessing one partition for a given asset while the other far larger partition has to wait could cause issues. Edit: Logic.

            • Firestarter
            • 5 years ago

            Lying about the bus width IS lying about the speed. That said, the specs that they show on their website are still technically correct, even though the 224GB/s specified for the 970 is not actually the same as the 224GB/s specified for the 980, because of only 7 of out 8 controllers being active most of the time. It’s only when the graphics card is used in a way that Nvidia [i<]actively[/i<] tries to prevent that it is actually able of hitting that 224GB/s peak. IMO, to think that all this is just normal and justifiable is a bit of double-think. And yes, the sky is not falling in the least, the 970 is still a great card and I'd still buy one. What surprises me is that Nvidia think that they should stoop down to this level to sell their graphics cards. Aren't they good enough on their own merit? Couldn't they have just sold the 970 as having 3.5GB + some cache, with a 224bit bus? I mean, it's pretty clear for all of us that it would still be beating AMD's cards on the important performance metrics and the only reason AMD is competing is because they dropped their prices accordingly.

            • Pwnstar
            • 5 years ago

            You’re right. I think I was mostly replying to you calling it “capacity”. They put 4GB on the board, 500MB of it is very slow.

          • Meadows
          • 5 years ago

          Coke-Cola? lol.

      • derFunkenstein
      • 5 years ago

      Just what this conversation needed: Kroger making legal claims. Self parody get!

        • Krogoth
        • 5 years ago

        I’m not making legal claims. I’m just saying that whole thing opens up Nvidia to potential legal action. It just a matter of somebody tries to push a case for it.

        What is certain about this whole debacle is that certain heads will roll and suddenly new disclaimers will be introduced into EULA and product warranty with future products.

        There have been cases over sillier non-sense where the plaintiff won.

          • VincentHanna
          • 5 years ago

          Totally agree with you bro, this is [b<]at least[/b<] as serious as the 1991 man who sued Anheiser-Busch for false advertising when drinking their beer failed to get him laid and he suffered emotional distress.

            • derFunkenstein
            • 5 years ago

            lololololol

            • Krogoth
            • 5 years ago

            Nah, this case is more like a car manufacturer claims and marketed Sports Car X having 300HPs but due to some technical issue that was quietly omitted the output of the engine was actually 280HP. A few months later, some independent dyno tests on Sports Car X reveals this and car enthusiast wanted to know why and while exploring this they discovered the technical issue.

            Technically, Sports Car X works fine in almost all cases and excepted usage patterns, but when you want to push the car to its limits the technical problems reduces its potential.

            • VincentHanna
            • 5 years ago

            bad analogies are bad

            • Krogoth
            • 5 years ago

            It is a rather fitting analogy.

            Product A works fine for the most part but it doesn’t work up to advertised rating/mertic due to a technical issue that the manufacturer quietly omitted.

      • glacius555
      • 5 years ago

      Yeah, more like $10 in 10 years, much like with Pentiums.

    • RdVi
    • 5 years ago

    They talk about the issue like it’s a feature. It definitely is when compared to older parts, it means the consumer gets a faster GPU than they would otherwise. Surely though, the consumer is also paying more for this faster part? So, when they fail to advertise the ‘feature’ properly to the consumer, it potentially becomes a feature that benefits Nvidia and hurts the consumer.

    Sure, consumers should research performance in the way of reviews instead of relying on advertised specs, but I don’t think it’s that straightforward in this case. The issue will likely have an effect on the standing of the 970’s relative performance in newer titles compared to the 980 and the competition, even with more normal framebuffers. With consoles packing 8GB of RAM and currently using 5GB for games, surely more and more titles are going to push past 3.5GB, making the performance buckle disproportionately to today where only the super GPU limiting high resolution settings use over 3.5GB VRAM.

      • Ninjitsu
      • 5 years ago

      The consumer is actually paying (much) less for a faster part, when compared to older parts.

    • sschaem
    • 5 years ago

    Recap of my last posts, and the final take away:

    The GTX970 address memory using a main 224bit bus capable of addressing 3.5GB of memory
    but also come with a 32bit ‘sideband’ bus capable of addressing a separate .5GB

    And this explain the 3 issue people have been noticing:

    1) Why the GTX970 raw memory bandwith is lower then the 980
    (even so on spec nvidia label the 980 & 970 with the same 224GB rated memory bus)

    2) Why the driver limit’ed’ allocation to 3.5GB , even so the card is 4GB equipped

    3) Why bandwidth drop off when accessing the last .5GB of the cards memory

      • Voldenuit
      • 5 years ago

      [quote<]The GTX970 address memory using a main 224bit bus capable of addressing 3.5GB of memory but also come with a 32bit 'sideband' bus capable of addressing a separate .5GB[/quote<] Pretty much how I read it, and seems to jive with the Nai numbers (150 GB/s in first 3.5 GB is almost eactly 7x the 22 GB/s that the last 512 MB gets). Since the eighth channel has to share L2 with the seventh channel, I wonder (from a purely academic standpoint) whether there are going to be memory contention issues when a program is accessing data from both the first 3.5 GB segment and the last 512 MB. Does Nai's benchmark test addresses sequentially? It may not notice any contention issues if it does. I would still like to see TR do some in-depth testing of this, even though I'm not in the "DIE NGREEDIA" camp on this issue.

    • kishish
    • 5 years ago

    So the 970 actually has less L2? Probably more of a future performance hit not having as much L2 as having the last 512MB run at 28GB/s.

      • jts888
      • 5 years ago

      Nvidia correctly reported the 970s as having reduced CUDA core and texture unit counts (from 2048/128 down to 1664/104), but they did not disclose the drops in ROPs (64 -> 56), effective memory size, bus size and bandwidth (4GiB -> 3.5GiB, 256b -> 224b, 224 GB/s->196 GB/s), or L2 cache size (2048kiB -> 1792kiB).

      The ROP, L2 size, and RAM bandwidth hits all are part of why 970s are worse cards than 980s for any workload but were not disclosed.

      The figures in themselves will matter to almost nobody, but the lingering issue was how much Nvidia knowingly lied on seeming insignificant details for the purposes of better marketing a binned product.

        • Chrispy_
        • 5 years ago

        The doubt, cynicism and distrust is strong in you, young padawan.

        I would have agreed with you in the past but I genuinely don’t think that marketing goons have a clue. Look at AMD – their marketing department can’t even find their own ass.

        [b<]The real heart of the matter is this: The 970 is a slightly reduced card for a very reduced price. Regardless of what configuration it has, the configuration reviewers received is the configuration you can buy today.[/b<] Nvidia's marketing department could claim it has a 4096-bit, chicken-and-dumpling BaconTech Engine with MagicBeans and WonderZ, but the real-world performance in tested games at tested resolutions is the reason people will buy the card, BaconTech or not....

          • rechicero
          • 5 years ago

          That’s not true. A lot of people doesn’t read enthusiast sites and, for them, the numbers are the tipping points. And 4 is more than less than 4. And, as a lot of people already said, this could be an issue in future games, so enthusiast would like to know too. In such a balanced race as the GPU business, that small lie can mean a lot of sells.

        • Klimax
        • 5 years ago

        A lie? If this is a lie , then I would like to know how would label AMD on multiple occasions…

        Reminder: It has to be proven it was intentional for it to be a lie. (Otherwise other companies are in deep crap…)

          • VincentHanna
          • 5 years ago

          It doesn’t have to be intentional, per se. The fact that the “bug” was well known, explored, tested, and patched internally (and apparently baked into the architecture from the word go) is enough.

          If it hadn’t turned out to be such a minor issue (–2-4% additional performance lost), I could understand people being very upset.

            • Klimax
            • 5 years ago

            The only problem was error in spec sheet in guides. Nothing and nowhere else. The only thing you could even attempt to liable as a lie, is wrong ROP/L2 spec, but that is just all.

            So what is it “enough”? Enough for what?

          • Pwnstar
          • 5 years ago

          You can’t excuse bad behavior by pointing to more bad behavior.

          • Tirk
          • 5 years ago

          Hehe if we are to excuse Sally for telling a lie because Susan told more lies then we are going down a very slippery slope are we not? Let alone, creating a very obvious logical fallacy.

      • the
      • 5 years ago

      The reduction in L2 cache will have an impact in performance [i<]today[/i<]. As a cache, it can be used immediately by any game that needs more than 1792 KB of graphics memory. You'd have to go back a decade or two to find games where this wouldn't be an issue. It is one of the reason why the GTX 980 is faster in current titles. The long term issue is that last 512 MB chunk of memory because few games can utilize more than 3 GB of memory. When games really start increasing assets, the difference between the GTX 970 and GTX 980 will widen by a few percent.

        • Klimax
        • 5 years ago

        Games don’t depend on L2 (not visible), drivers do. And to my knowledge this large L2 was never seen before Maxwell.

        As for rest, we will see. Will depend how well driver team will do their job.

          • the
          • 5 years ago

          While true that games don’t directly see the L2, as a cache they will use it . The drivers do have access to the L2 cache, the main purpose is for configuration and not active management of the cache’s contents. It is far superior in terms of performance to let the hardware decide what gets stored in the L2.

          While true to date that this is nVidia’s largest on-die GPU cache, it isn’t their first nor are previous incarnations far behind. nVidia has used L2 caches before: the GTX 780 Ti had 1536 KB and the GTX 680 had 512 KB. If I were a guessing man, I’d say that the GM200 would have 3 MB of L2 cache.

    • f0d
    • 5 years ago

    when will we see benchmarks of how this configuration actually affects performance?

      • Klimax
      • 5 years ago

      Likely already in progress. (IIRC PCPer is doing their own and I would be surprised if TechReport didn’t too)

      Pity I don’t have both 980 and 970 (currently no Maxwell here), it’d be interesting exercise.

    • sschaem
    • 5 years ago

    “Exibit A”

    [url<]http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-980/specifications[/url<] [url<]http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-970/specifications[/url<] nvidia is selling both card having the exact same memory bandwidth. From what we see, the 970 bandwidth is actually 20% lower at best, and up to 7x slower in the last 512MB edit: nvidia diagram makes sense, the GTX970 is really a 3.5GB 224bit cards + .5GB available on a 32bit bus

      • Klimax
      • 5 years ago

      Because those figures are always about theoretical memory bandwidth (simple multiplications). Never promised anything else…

    • tanker27
    • 5 years ago

    So I am not up to speed as I used to be on hardware, but this article prompts me to ask:

    “So the 970 is a less than perfect 980? Like its damage someway when it comes off the silicon?” “It’s not intentionally cast that way?”

    (please forgive the ignorance)

      • sschaem
      • 5 years ago

      Yes and no. The 970 uses the same silicon as the 980.

      But by design, the part that fail to be full 980 can still become 970.

      The part disabled to become a 970 c lass card is affecting the memory controller function.

      edit: nvidia latest diagram show that 2 of the 32bit memory controller are ‘switched’ VS operating in parallel.
      For this to work the memory is never accessed at a full 256bit, but at 224bits
      One of the MC is left unused.

      In effect the GTX970 is partition this way:

      Partition A : a 224bit bus accessing 3.5GB of memory
      Partition B: a 32bit bus accessing .5GB of memory

      This examplain multiple things:

      a) Why the GTX970 raw memory bandwisth is lower then the 980, even so on spec nvidia tells us both are 224GB bandwith cards

      b) Why the driver limit’ed’ allocation to 3.5GB , even so the card is 4GB equipped

      C) Why bandwidth drop off when accessing the last .5GB of the cards memory

      • Melvar
      • 5 years ago

      Yes. For some time now chipmakers have been selectively disabling bad sections of silicon in otherwise working chips and selling them as lower end parts rather than just discarding them.

      • Chrispy_
      • 5 years ago

      Yeah, it’s just like when AMD were selling Athlon X3 processors. They were actually all quad-core silicon but if a defect meant that only one of the four cores was faulty, AMD just disabled it and sold it as a triple-core processor.

      Likewise, AMD and Nvidia have been disabling defective chunks of their graphics chips and selling them as lesser products for years, so the GTX970 is a defective GTX980 and and another example is AMD’s HD7950 was a defective HD7970.

      This process is called die-harvesting – whenever you see that phrase it means selling defective silicon with the broken part(s) disabled.

      • tanker27
      • 5 years ago

      well I’ll be danged, learn something new everyday.

      Less than perfect silicon, hold on to it, turn off the bad parts and make a whole new item = PROFIT!

      • psuedonymous
      • 5 years ago

      Yes. this is known as ‘binning’, and is standar4d practice. Every time you see multiple different products based on the same ‘core’, it’s due to binning. Intel’s Core i3/i5/i7 range? Binning. 280 and 280X? Binning. 260 and 260X? Binning. Same die size, same core technology, different component counts? Binning.

      The faulty/defect parts are the parts that are disabled. That’s the entire point.

    • chuckula
    • 5 years ago

    So how many FPS slower will games run today vs. yesterday due to this announcement?

      • Terra_Nocuus
      • 5 years ago

      ALL OF THEM /ssk

    • deruberhanyok
    • 5 years ago

    I’m wondering how much of an issue this will be once the 8gb chips start shipping in volume?

    [url<]https://techreport.com/news/27676/samsung-starts-making-8gb-gddr5-memory-chips[/url<] Seems to me that it would enable easy doubling of capacity on these cards, so we could have a 960 with 4GB of memory and a 980 with 8GB of memory, and maybe at that point NVIDIA could just say "eh, the 970 only has 7GB of memory". Sort of like if they'd cut it down by a full 25% and gave it 3GB of memory on a 192-bit interface, except in reverse. If that makes sense. It also makes me wonder why they didn't just say it has 3.5GB of memory in the first place.

      • BlackDove
      • 5 years ago

      These GPUs are barely more powerful than disabled GK110. Why would anyone waste money on an 8GB 970?

        • aspect
        • 5 years ago

        So people can use the highest res textures on games where the textures don’t even look that great like Shadows of Mordor.

          • BlackDove
          • 5 years ago

          Wouldnt you be lagging already, before you hit the 3.5GB limit?

            • auxy
            • 5 years ago

            No. Loading more textures doesn’t make a game slow down.

            Does installing more data on your hard disk make the existing data slow down? It’s not a perfect analogy, but the point stands. On a normal GPU, the entire VRAM is accessed at essentially the same speed, so having more assets from which to construct the scene is not a detriment to performance.

            • NTMBK
            • 5 years ago

            Higher resolution textures reduce texture-fetch coherency and reduce the effectiveness of the texture cache, as well as increasing total bandwidth usage. They certainly do slow the game down.

            • auxy
            • 5 years ago

            You are making the invalid assumption that higher resolution textures equate to an inability to store all necessary texture data for a scene in local memory.

            Total bandwidth usage (as you should well know) is overwhelmingly linked to buffer size discounting external factors such as framerate limits.

    • Chrispy_
    • 5 years ago

    So, 4GB is better than 3.5GB because the gimped 1/8th speed 512MB is still faster to access than system RAM over the PCIe bus.

    When you buy any die-harvested part (CPU or GPU, AMD, Nvidia or Intel alike) you know you’re losing parts of the die in exchange for a lower-cost product. and in the case of the 970 they never tried to hide the fact that it only had 7/8 of the L2 cache and that there would obviously be some memory speed loss. If you [i<]assumed[/i<] that the memory bandwidth would be completely unaffected by the die-harvesting then disappointment is your own fault. It is still 4GB, it does still have a 256-bit memory controller and whether it uses that as effectively as the 980 is irrelevant. People wanting 980 performance should have ponied up for a 980 in the first place! As much as I dislike Nvidia's business practices, treatment of the Linux community and developer bribery to push their own bespoke agenda and APIs, I don't think they deserve bad press for this. It's a die-harvested product that is incomplete compared to the fully-enabled part. Any assumed performance based on spec-sheet figures alone is misplaced hope - which is why we have websites like TR to give us real-world performance values across a wide range of different scenarios.

      • geekl33tgamer
      • 5 years ago

      Finally someone talking sense on this whole thing, that has been completely over exaggerated on some other sites.

        • jts888
        • 5 years ago

        The performance impact is likely minimal in most realistic circumstances, but the issue is that the 970 GPUs were specifically designed to be able to technically access the last half gigabyte of RAM but should basically never do so in real world scenarios, and the fact that half the technical stats Nvidia’s marketing have put out are essentially lies.

        They aren’t 256b/4 GiB/224 GBps/2048kiB cache/64 ROP.
        They’re 224b/3.5 GiB/196 GBps/1792kiB cache/56 ROP.

        The fact that the L2/MC buddy link was specifically designed to allow the last bit of RAM to be “used” means that Nvidia intended this product to be misrepresented before it was ever built, so the numerous incorrect specifications might not deserve the benefit of a doubt.

          • Voldenuit
          • 5 years ago

          [quote<]The fact that the L2/MC buddy link was specifically designed to allow the last bit of RAM to be "used" means that Nvidia intended this product to be misrepresented before it was ever built, so the numerous incorrect specifications might not deserve the benefit of a doubt.[/quote<] Being a large corporation with many independent groups, it's just as likely the hardware engineers and scientists came up with the architecture, the process and chip engineers came up with the block layout, some optimization group came up with the idea of disabling parts of SMs and the L2, and finally the driver team optimized what they felt was the best way to use the resources available to a GPU with asymmetric memory access. Some of my working group divisions might be a bit off since I'm not a GPU designer nor work in computer hardware, but there doesn't have to be a conspiracy. I will agree that the marketing department deserves a rap on the knuckles for misrepresenting the raw memory speed of the 970 GTX, but that could as easily be due to miscommunication or failing to understand the technical nuances of the card.

            • puppetworx
            • 5 years ago

            [quote<]I will agree that the marketing department deserves a rap on the knuckles for misrepresenting the raw memory speed of the 970 GTX, but that could as easily be due to miscommunication or failing to understand the technical nuances of the card.[/quote<] And none of the engineers saw any of the PR, packaging, reviews, etc. from the last three months and emailed the PR department for them to correct it? There was no-one like [url=http://blogs.nvidia.com/blog/author/tom-petersen/<]Tom Petersen[/url<](not that I'm blaming him or calling him out, just using him as an example), with a strong grasp of the technology, in the PR department? More likely the PR deparment took a calculated risk on a small lie because it looked better.

            • Pwnstar
            • 5 years ago

            Yeah, I have a hard time believing not one of their engineers bought a 970 and noticed the spec difference and then didn’t say anything during the last 4 months.

          • Chrispy_
          • 5 years ago

          IMO, the buddy link was a conscious decision to spend transistor budget on minimising waste. Salvaging that 512MB of GDDR5 is better than losing it, even if it’s slower than the main 3.5GB chunk.

          If they’d not included the buddy links they could have made the whole die slightly smaller and fabricated more dies from a single wafer for more profits. Instead, they spent the transistor budget on ensuring the 970 (and potential 960Ti) don’t suffer as much from the die harvest as they could have done.

          • Ninjitsu
          • 5 years ago

          [quote<] The fact that the L2/MC buddy link was specifically designed to allow the last bit of RAM to be "used" means that Nvidia intended this product to be misrepresented before it was ever built, so the numerous incorrect specifications might not deserve the benefit of a doubt. [/quote<] No, not [i<]this[/i<] product and not to be [i<]misrepresented[/i<]. It's built into the architecture, and from a hardware design standpoint it's quite smart and quite brilliant. Marketing ended up misrepresenting stuff, inadvertently or not.

      • VincentHanna
      • 5 years ago

      what’s wrong with Nvidia’s treatment of the linux community?

        • cobalt
        • 5 years ago

        I assume that’s a reference to the proprietary nature of their drivers. (For me, given that their Linux drivers are as good as the Windows ones, I’ll take the opposite stance and claim that’s pretty good treatment of the Linux community.)

        That said, I usually agree with Crispy, so maybe I’m misinterpreting that statement.

          • Chrispy_
          • 5 years ago

          I’m probably out of date; I remember this from a couple of years ago:

          [quote<]Linus Torvalds, the man behind the Linux kernel, has called Nvidia 'the single worst company we have ever dealt with,' raising eyebrows - and his middle finger - during a presentation at the Aalto Centre for Entrepreneurship in Finland.[/quote<] There were loads of videos, articles and a huge cheer from the Linux fanbase so I'm assuming it wasn't just Linus' one-man grudge. Google "Linus slams Nvidia" or something!

            • Ninjitsu
            • 5 years ago

            IIRC he didn’t like the fact that they weren’t open source. But Nvidia does put out good drivers for Linux like they do for Windows, at least in the last few years, from whatever I’ve read over time.

            • Concupiscence
            • 5 years ago

            Outside of some reported edge cases like laptop chips, Nvidia’s Linux drivers have been solid for a [b<]very[/b<] long time. They gave me no trouble at all on a Frankenputer running Slackware back in 2003, and I'm pretty sure they were workable without many issues when I started college in the year 2000.

      • jihadjoe
      • 5 years ago

      What I wonder about is why the way memory is connected changed. The GTX670 and 760 seemingly had full access to the 256-bit BUS, despite having disabled SMs, so why not with the 970?

        • Voldenuit
        • 5 years ago

        It’s not the SMs that are affecting the memory access, it’s the disabled chunk of L2 into which the eighth memory controller feeds.

        Nvidia is saying that under Fermi or Kepler, they would have to had disabled the the entire L2 partition and both the memory controllers that feed into it, leading to a 192 bit memory path with 3 GB of RAM. Instead, with Maxwell, it looks like they are able to provide a 224-bit + 32 bit memory path with 3.5GB + 0.5 GB of RAM.

        So they’re right, it [i<]s[/i<] a bfeature, not a bug. A buggy feature, perhaps, and not perfect, but it still sounds better than a 192-bit x70 card would have been.

      • Pholostan
      • 5 years ago

      [quote<]in the case of the 970 they never tried to hide the fact that it only had 7/8 of the L2 cache[/quote<] Didn't they try to do exactly that? Hide the fact? The spec was 2MB of L2 and 64 ROPs up until a couple of days ago. Now it is corrected. I don't think most consumers know what a die-harvested part is. We know, but we are very few compared to all who buy higher end graphics cards. If we don't call Nvidia on this, who will? What stops them form doing even worse next time? What stops AMD from doing something similar? Intel? Etc. The market needs informed consumers, lying to them should be black mark on their company reputation.

    • tviceman
    • 5 years ago

    In other news, the GTX 970 performs exactly the same today as it did last week, and the week before that, and the week before that…..

      • derFunkenstein
      • 5 years ago

      Yup. This stuff is interesting from an academic point of view, but people are getting way bent out of shape over what I think is effectively nothing. This “handicap” is baked into the performance reviews, so it’s not like they’re taking more performance away.

        • the
        • 5 years ago

        The issue is more about future titles than anything here on the market today. With console packing 8 GB of memory (granted, not all of that is going toward GPU assets), having a fast 3.5 GB pool may not be enough going forward. That 512 MB pool of memory can be the inflection point between smoothness and choppiness at higher resolutions.

          • derFunkenstein
          • 5 years ago

          But once it goes over 3.5GB for an app with exclusive access, it will use that last little bit – albeit at lower speeds. Since I don’t have a GTX 970 myself, I’m kinda waiting on the best investigative websites (TR and PCPer) to try to force the 970 to go over that 3.5GB mark. Based on what nVidia’s statement said on Saturday, it *will* eventually use that RAM if the app forces the card’s hand, and it’s (allegedly) not at a huge disadvantage when doing so – that the soft 3.5GB limit is to keep performance optimal. I’d love to see someone test that theory and really show us if there’s a difference. My own GPU (GTX 760) could stand to be replaced with my 1440p display so I’ll eagerly anticipate the results.

          • Ninjitsu
          • 5 years ago

          It’ll make a difference of 1-3%. And from the benchmarks Nvidia’s given us, that’s a 1-2 fps difference (10% of 30 fps is 3 fps).

          Not sure it’s really worth so much hullabaloo.

          Damn, I always wanted to use “hullabaloo” in a sentence.

            • VincentHanna
            • 5 years ago

            who plays games at 30FPS?

            • Ninjitsu
            • 5 years ago

            Whoever plays recent games at 4K with a 970.

            • Voldenuit
            • 5 years ago

            Yep. Early adopter tax.

            Take home message seems to be if you want 4k, wait for the Freesync(maybe Gsync) IPS panels, and the next generation of GPUs.

            The good news is that 2.5K panels are really cheap now, and would make a good secondary monitor down the road. Also, all the ‘upper midrange’ cards have more than enough performance for this and are reasonably priced.

            If you’re worried about future-proofing, well, that’s never been a particularly strong suit of PC gaming. I remember when my 486 DX2 and Tseng ET VESA cards were the bee’s knees…

        • Pwnstar
        • 5 years ago

        I think it is because people don’t like being lied to.

          • derFunkenstein
          • 5 years ago

          Lied to about what? how it performs? Are you calling TR and other HW review sites liars?

            • Pwnstar
            • 5 years ago

            nVidia lied about the specifications of the 970.

            • derFunkenstein
            • 5 years ago

            They certainly didn’t lie about how it performs, though. All that matters to me is price and performance and to a small extent power consumption because I don’t want a tiny furnace in the summer. The 970 is actually a little bit of a whiff right now because the vanilla R9 290 is a faster card for considerably less money. But it performs how it performs. It’s like Scott was saying on the podcast – it performed at this level when they did this review, people did some digging, nVidia made their statements, and at the end of the day it still performs how it performed. And people were apparently happy about it because they’ve apparently sold 2million GM204-based cards based not on specs, but how it performs.

            • VincentHanna
            • 5 years ago

            No, the specs are actually accurate.

            “technically, if you kindof, sortof squint and cock your head to the left while humming the devil went down to georgia, it looks like an argument could be made that there is a technical omission from the specs.”

            but that isn’t the same as the specs being wrong.

            They aren’t.

        • Tirk
        • 5 years ago

        More in depth performance reviews might uncover some performance issues not always picked up before the issue was known.

        A great example of that was the introduction of the frame time metric along with fps numbers. An issue was discovered about discrepancies in frame times, so further analysis uncovered better tests to show what the consumer would be experiencing. Now the frame time metric is just as an important metric as fps on GPUs and because of that discovery Nvidia and AMD have focused on their GPUs improving that metric, much to the happiness of their consumers.

        Its somewhat short sighted to assume all the performance issues are appropriately “baked” in to current tests if those tests really did not account for the design issues of the card.

      • sschaem
      • 5 years ago

      Yes, its not like bump gate all over again, where the 970 & 980 will stop working … thank god.

      But the reason this is talked about is that this was a know design limitation from day one, and all point to nvidia choosing to keep that a secret to gain marketing advantages.

      Better to sell a 4GB card then a 3.5GB card…

    • sschaem
    • 5 years ago

    Clarification needed:

    The CUDA memory / L2 cache test app show the drop off happen at 3.25 GB not 3.5GB
    (The secondary issue reported, only related, was that nvidia driver max at 3.5GB. we now know for good reasons)

    512MB would indicate that 1/8 of the memory is affected,
    but the benchmark seem to show 1/5.3 is (a 13/16 ratio)

    Now , the diagram in the article actually show that 2 of the 8 ram chip are placed on the ‘crippled’ channel…

    BTW, the benchmark numbers also make sense. ~7x slower access, meaning the last chip is only on a 32bit bus.

    So in effect the GTX 970 is a 256bit 3.5GB card .. but also as 512MB of ‘sideband’ memory at 32bit

    nvidia really should have just sold it as a 3.5GB cards and avoided this mess.

      • Meadows
      • 5 years ago

      Prove your drop-off numbers.

        • sschaem
        • 5 years ago

        I haven’t seen number posted by nvidia, wish I had.
        So ‘my’ numbers are just what have been posted comparing the same benchmark executed on the 970 and 980.

      • tsk
      • 5 years ago

      Those numbers stem from people running nai’s benchmark incorrectly, I.e not in headless mode.
      Which of course is a problem when people just download a benchmark, run it and raise their pitchfork about numbers while really having no idea what is actually going on.

        • sschaem
        • 5 years ago

        That benchmark highlight two issues.

        note: that a non headless card should report about 1GB less in bandwidth from the continual ramdac DMA of the frame buffer.

        Here is what we see:

        First, even so the 970 and 980 both sport a 256bit buds with 7GHZ effective memory,
        the 970 seem to be about 13/16 slower in the <3.2GB space. 150 vs 178

        Yep, that 13/16 number again…

        Assuming the 980 was headless, but the 970 was not, this would account for a 1GB difference.
        Here we see 28GB.

        please note that both card are rated to have exactly EQUAL available bandwith of 224GB:
        (nvidia is telling us that both card have the same bus and bandwidth.)

        [url<]http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-980/specifications[/url<] [url<]http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-970/specifications[/url<] But thats not the point of contention (even so this should be looked into) The issue is what happen at the 3.25 mark. The GTX 970 performance drop from 150GB to 22GB This not a headless issue. Do you have link to the nai benchmark results under well defined configuration ? So far the data we have point to a 13/16 memory performance drop , even so nvidia say the 980 and 970 have equal bandwidth. AND that only the first 13/16 of the memory is accessed via the full 256bit bus architecture. edit: scanning, I see many review mention this explicitly. "The GTX 970 features the same 224GB/s memory bandwidth as it makes use of the same GDDR5 1750MHz (7.0Gbps) memory and 256-bit wide memory bus."

          • willmore
          • 5 years ago

          From reading this article and some over at PCper, the conclusion I come to is that what you’re seeing is caused by the last 1GiB being accessed over *one* crossbar port rather than 2–like all other memory banks.

          So, if you treat it as a 7 bank card (3.5GiB) then you’re going to see full memory bandwidth, but if you start to access that last bank, both the last and the second to last (which are both accessed over the same crossbar port and L2 cache) banks now have to compete for bandwidth on that 7th crossbar port.

          One might be able to infer more if specific tests were run. Do you have links to (proper headless) runs of this benchmark?

      • jihadjoe
      • 5 years ago

      I’m not too bummed about the 4GB thing, because the card does actually have 4GB. “256bit” IMO is the bigger issue, because the card never truly has 256bit performance.

      But hey, 224+32 = 256!

      Reminds me of the time Atari sold the Jaguar as a 64-bit console because it had a 32-bit CPU and a 32-bit DSP.

    • NTMBK
    • 5 years ago

    [quote<]working exactly as we designed it[/quote<] And our design has a bug.

      • Pwnstar
      • 5 years ago

      No, they did this on purpose. Bugs are unintentional.

    • puppetworx
    • 5 years ago

    [quote<]We thought this feature would make it a better product, and we think we achieved that goal. We want to make sure people understand it well.[/quote<] Such a great "[i<]feature[/i<]" you forgot to explain or even tell us anything about it at launch. I couldn't help have myself a sensible chuckle when I read that.

    • puppetworx
    • 5 years ago

    There’s a broken link, here’s the right one for those interested until it gets fixed: [url=https://techreport.com/blog/27143/here-another-reason-the-geforce-gtx-970-is-slower-than-the-gtx-980<]https://techreport.com/blog/27143/here-another-reason-the-geforce-gtx-970-is-slower-than-the-gtx-980[/url<] BTW I'd totally forgotten about that article, was nteresting to re-read with this in mind.

      • Damage
      • 5 years ago

      Doh. Link fixed.

    • TwoEars
    • 5 years ago

    Nice write-up mister Wasson.

    • loophole
    • 5 years ago

    How interesting! Excellent write-up.

    A little typo that jumped out at me: in the paragraph starting “Incidentally, Alben told us…” the third last sentence should be “Thus, when doing the same work, the GTX 970 may use less of its total RAM capacity as a matter of course than the GTX [b<]980[/b<] does.

      • DPete27
      • 5 years ago

      I noticed that too.

      • Damage
      • 5 years ago

      Fixed.

      • Techgoudy
      • 5 years ago

      With how long the article was I’m surprised there weren’t more typos. Good job TR, keep up the awesome work!

    • SuperSpy
    • 5 years ago

    The reason they couldn’t call it a 3.5 GB card is purely marketing. The card’s RAM amount is just about the only video card statistic that can be dumbed down to a single number, so they couldn’t possibly sell cards with less memory than their competitor, even if it didn’t matter. See: all the crappy $50 cards with 4GB of VRAM just so the box art people have something to print in bold.

    On the flip side, it’s a bit shady to call it a 4GB card when the last 512MB is so horribly crippled.

      • cynan
      • 5 years ago

      Rather, that’s why they bothered going through the trouble of having shaders ever access the last 512mb at all, rather than just disabling it completely.

      I doubt that slower last 512mb of memory makes much of a performance difference in many real world scenarios. But imagine the alternative – the hoopla that would have ensued if this story had surfaced and people found out that their 4GB card was really truly only 3.5GB with and extra 512mb of dead weight..

        • Firestarter
        • 5 years ago

        or that their 3.5GB card actually has 4GB but can’t access it. Which is worse?

    • anotherengineer
    • 5 years ago

    And it’s currently on SALE in Canada!!!!!

    Only $440 + $10 shipping + 13% tax = $505.50

    [url<]http://www.ncix.com/detail/msi-geforce-gtx-970-twinfrozr-2a-102245-1445.htm[/url<] edit smaller sale, revised price increase.

      • sweatshopking
      • 5 years ago

      while the 290 is [url<]http://www.memoryexpress.com/Products/MX50295[/url<] 339$ CAD

      • Pwnstar
      • 5 years ago

      $485 is on “sale”?!?!?!?

        • sweatshopking
        • 5 years ago

        hence the strong advantage for the r9 290(x) in canada vs nvidia.

    • ultima_trev
    • 5 years ago

    So more or less this could be considered a GPU with a 224-bit memory bus and 56 ROPs. Would have made more sense to call it GTX 960 Ti. Wonder if this also affects GTX 980M, another hobbled GM204 with a “256-bit” memory bus.

      • ChangWang
      • 5 years ago

      I swear, you plucked that statement right out of my mind. And I think it leaves room for a 970 Ti if they bin some chips with 1664 shaders (or more) and a fully enabled crossbar, L2 and Mem bus.

      • GeForce6200
      • 5 years ago

      According to a test run by a member on OCN the 980M is not affected when using Nii’s test. The GTX980 has 12SMM to 970’s 13.
      [url<]http://www.overclock.net/t/1535502/gtx-970s-can-only-use-3-5gb-of-4gb-vram-issue/300[/url<]

    • sweatshopking
    • 5 years ago

    [quote<] Nvidia is able to equip the GeForce GTX 970 with a full 256-bit memory interface and still ship it at an attractive price in high volumes [/quote<] then why is the 970 100$ more than a 290, but the 290 has a 512 bit interface?

      • yokem55
      • 5 years ago

      Because Nvidia believes it is better for their business that they turn a profit on the products that they sell – [url<]https://techreport.com/news/27698/amd-slipped-back-into-the-red-last-quarter[/url<]

        • sweatshopking
        • 5 years ago

        I don’t care if nvidia turns a profit. I don’t own them. I want cheap/fast gpus.
        My issue is when the pretend they’re doing us a favor. “full 256 bit memory interface”. How is that full when the competition is 512? attractive price? it’s 100$ more (at least in canada) than a faster 290.
        The 970 is a fine card, and I’m fine with people buying it. I even LIKE nvidia, and my last gpu prior to my 290 was maxwell based (though prior to that was a reference 290x which was FAST, but LOUD. My new one I can’t hear). I just don’t love when they act like they’re doing us a favor.

          • yokem55
          • 5 years ago

          Memory busses don’t render frames, GPU’s do. If Nvidia can get good performance out of a narrower bus, what does it matter to me?

        • Pwnstar
        • 5 years ago

        AMD isn’t making a profit in general, but that doesn’t mean their GPUs aren’t profitable.

      • DancinJack
      • 5 years ago

      You mean after numerous price cuts by AMD because the can’t sell it? Yes, then you would be on point. Let’s at least TRY to remember the 290 launched at 399 USD.

        • auxy
        • 5 years ago

        I think it has more to do with AMD trying to capture market share. If you look at their financials the graphics division is what’s keeping AMD afloat, so I think the GPUs are selling through just fine.

        Also keep in mind that for a time AMD products were selling [i<]above[/i<] list price -- significantly so -- something that I have never seen NVIDIA products doing. (*‘∀‘)

          • DancinJack
          • 5 years ago

          We can at least correlate can’t sell it and market share. I’m just saying that he was comparing apples and oranges.

            • sweatshopking
            • 5 years ago

            no i’m not. They claimed attractive prices and FULL memory interfaces. It’s more expensive and has 50% of the memory interface. i’m not referring to whether it sells or not.

            • DancinJack
            • 5 years ago

            I’m confused. It is a full 256-bit interface and at launch, the 970 was cheaper than the 290. What are you so worked up about? In your mind FULL = 512 and only 512?

            • sweatshopking
            • 5 years ago

            Who cares what the prices were “at launch”. are you buying one 6 months ago? That information is entirely irrelevant today. My phone cost 500$ when it was new. now it’s a 200$ phone. such is the market of technology.

            I’m not “so worked up”. Were you thinking that it could possibly have a 256bit interface but only 213 bits actually worked? No, of course not.
            the definition of full within this context: “not lacking or omitting anything; complete.”
            They’re presenting the card as if 256 was the FULL POSSIBLE bit interface amount, when it clearly isn’t, and the competition is a full DOUBLE that amount. it’s just that they’re being disingenuous with the use of the term “full”. it’s not “full” anymore than me selling you a “full half a meter stick, with a full 50cm!” when the guy standing next to me has a full meter stick.
            Would you be cool with intel selling you a “full 2.5ghz cpu”? of course you wouldn’t since that doesn’t even make any sense in any rational way. Neither does this.
            for the record, the interface is largely irrelevant. nvidia tends to be lower, but has similar performance. I am fine with that. I just don’t like the disingenuous marketing.

            • Pwnstar
            • 5 years ago

            The 970 does not have a 256-bit bus. That was part of the nVidia lies.

            • VincentHanna
            • 5 years ago

            it DOES have a 256 bus, however it is broken up into 2 pieces, a 32 bit and a 224bit.

      • End User
      • 5 years ago

      PhysX

      😛

        • sweatshopking
        • 5 years ago

        lol +1

      • swaaye
      • 5 years ago

      It’s nothing new. AMD had to sell 512-bit HD 2900 at a low price because it sucked. NVIDIA had to sell 512-bit GTX 2xx at a fairly low price because HD 48xx was so competitive.

      970/980 do more with the same bandwidth than current AMD hardware so here we are again.

      • Meadows
      • 5 years ago

      Since when is 256-bit more attractive than 512-bit?

      All memory interfaces are beautiful.

        • sweatshopking
        • 5 years ago

        YOU SOUND LIKE A 960 KINDA GUY!

          • Meadows
          • 5 years ago

          Nah, I actually like them big.

        • the
        • 5 years ago

        I’m sorry, but a 256 bit interface isn’t wide enough to store my ego.

      • the
      • 5 years ago

      It stems from the idea that if nVidia disabled a full ROP/memory controller cluster, it’d have shipped with a 192 bit wide interface and 3 GB of memory.

      As for why nVidia can charge more: supply and demand. The R9 290 has been out well over a year where as the GTX 970 is only 3 months old. That time difference has allowed nVidia to roll out a new better/faster architecture. And people don’t like paying more for less (*glares menacingly at Comcast, Time Warner, and AT&T*) so AMD repositioned the price of the R9 290 to fall in line with the new competitor.

      There is a good chance that when AMD gets around to launching the R9 300 series that the current GTX 900 series will get some price cuts to account for the new competitors.

      Edit: Typo correction

        • sweatshopking
        • 5 years ago

        Your first sentence finally makes some sense. This is the response I was looking for.

      • derFunkenstein
      • 5 years ago

      Because it performs well and people are willing to pay it. I don’t really care about “specs” if the performance is where it should be at the price they ask.

      • Srsly_Bro
      • 5 years ago

      An attractive price for the company, not the consumers. 🙂

    • swaaye
    • 5 years ago

    It all sounds reasonable but yup they should have been more clear about it in the beginning. Somebody was bound to discover the quirk with focused corner-case tests, and now here we are with controversy.

    Considering that reviews with fancy modern testing methodology didn’t notice the behavior, even with super high resolutions, SLI testing, and the latest eyecandy games, it’s hard to complain too much about the technology.

      • Klimax
      • 5 years ago

      Actually in some cases (IIRC) there was quite a drop in performance in comparison to 980. One of linked Damage’s articles was investigating the anomaly.

      • VincentHanna
      • 5 years ago

      Actually this is good for Nvidia. As far as they are concerned, its free marketing. For normal people with realistic expectations who made informed decisions based on real world data, this is a non-issue, and anyone still harping on and on about it seems like a petty fanboy.

      Meanwhile, anyone who did happen to buy this card, good news, this is likely going to be one of the best supported driver generations since the 5xx.

      • Ninjitsu
      • 5 years ago

      Probably because no one was looking for it in the reviews.

    • chuckula
    • 5 years ago

    Can we officially call this giant conspiracy Memghazi now?

      • bthylafh
      • 5 years ago

      That probably sounded clever in your head.

        • chuckula
        • 5 years ago

        It sounds even more clever when incorporated in a techno mix.

      • exilon
      • 5 years ago

      It doesn’t flow as well as Bendghazi.

      • puppetworx
      • 5 years ago

      I prefer Ngreediagate.

Pin It on Pinterest

Share This