How much video memory is enough?

One question we haven't answered decisively in our recent series of graphics card reviews is: how much video memory is enough? More pressingly given the 4GB limit for Radeon R9 Fury cards: how much is too little? Will a 4GB video card run into performance problems in current games, and if so, when?

In some ways, this question is harder to answer than one might expect. Some enthusiasts have taken to using monitoring tools in order to see how much video memory is in use while gaming, and that would seem to be a sensible route to understanding these matters. Trouble is, most of the available tools track video memory allocation at the operating system level, and that's not necessarily a good indicator of what's going on beneath the covers. In reality, the GPU driver decides how video memory is used in Direct3D games.

We might be able to approach this problem better by using vendor-specific development tools from AMD and Nvidia—and we may yet do so—but we can always fall back on the simplest thing: testing the hardware to see how it performs. We now have a number of video cards based on similar GPU architectures with different amounts of VRAM, from 4GB through 12GB. Why not run a quick test in order to get a sense of how different GPU memory configurations hold up under pressure?

My weapon of choice for this mission was a single game, Shadow of Mordor, which I chose for several reasons. For one, it's pretty widely regarded as one of the most VRAM-hungry games around right now. I installed the free HD assets pack available for it and cranked up all of the image quality settings in order to consume as much video memory as possible. Mordor has a built-in benchmark that allowed me to test at multiple resolutions in repeatable fashion with ease. The results won't be as fine-grained as those from our frame-time-based game tests, but a big drop in the FPS average should still serve as a clear indicator of a memory capacity problem.

Crucially, Mordor also has a nifty feature that will let us push these video cards to their breaking points. The game's settings allow one to choose a much higher virtual resolution than the native resolution of the attached display. The game renders everything at this higher virtual resolution and then downsamples the output to the display's native res, much like Nvidia's DSR and AMD's VSR features. Downsampling is basically just a form of full-scene anti-aliasing, and it can produce some dramatic improvements in image quality.

Using Mordor's settings menus, I was able to test at 2560×1440, 3840×2160 (aka 4K) and the higher virtual resolutions of 5760×3240 and 7680×4320. That last one is a staggering 33 megapixels, well beyond the pixel count of even a triple-4K monitor setup. I figured pushing that far should be enough to tease out any memory capacity limitations.

My first two victims were the Radeon R9 290X 4GB and the Radeon R9 390X 8GB. Both cards are based on the same AMD Hawaii GPU, and they have similar clock frequencies. The 390X has a 20MHz faster base clock and a tweaked PowerTune algorithm that could give it somewhat higher clock speeds in regular operation. It also has a somewhat higher memory clock. These differences are relatively modest in the grand scheme, and they shouldn't be a problem for our purposes. What we're looking for is relative performance scaling. Where does the 4GB card's performance fail to scale up as well as the 8GB card's?

The 290X's 4GB of memory doesn't put it at a relative disadvantage at 4K, but the cracks start to show at 5760×3240, where the gap between the two cards grows to four FPS. At 7680×4320, the 4GB card is clearly struggling, and the deficit widens to eight FPS. So we can see the impact of the 390X's added VRAM if we push hard enough.

From a purely practical standpoint, these performance differences don't really matter much. With FPS averages of 16 and 20 FPS, respectively, neither the 290X nor the 390X produces playable frame rates at 5760×3240, and the highest resolution is a slideshow on both cards.

What about the Radeon R9 Fury X, with its faster Fiji GPU paired with only 4GB of HBM-type VRAM?

The Fury X handles 3840×2160 without issue, but its performance drops off enough at 5760×3240 that it's slightly slower than the 390X. The Fury X falls further behind the 390X at 33 megapixels, despite the fact that the Fury X has substantially more memory bandwidth thanks to HBM. Almost surely, the Fury X is bumping up against a memory capacity limitation at the two higher resolutions.

What about the GeForce side of things, you ask? Here it all is in one graph, from the GTX 970 to the Titan X 12GB.

Hmph. There's essentially no difference between the performance of the GTX 980 Ti 6GB and the Titan X 12GB, even at the very highest resolution we can test. Looks like 6GB is sufficient for this work. Heck, look closer, and the GTX 980's performance scales very similarly even though it only has 4GB of VRAM.

The only GeForce card whose performance doesn't follow the trend is the GTX 970, whose memory capacity and bandwidth are both, well, kind of weird due to a 3.5GB/0.5GB split in which the 0.5GB partition is much slower to access. We covered the details of this peculiar setup here. The GTX 970 appears to suffer a larger-than-expected performance drop-off at 5860×3240, likely due to its funky VRAM setup.

Now that we've seen the results from both camps, have a look at this match-up between the R9 Fury X and a couple of GeForces.

For whatever reason, a 4GB memory capacity limit appears to create more problems for the Fury X than it does for the GTX 980. As a result, the GTX 980 matches the performance of the much pricier Fury X at 5760×3240 and outdoes it at 33 megapixels.

We've seen this kind of thing before—in the only results from our Radeon R9 Fury review that showed a definitive difference between the 4GB and 8GB Radeons. The Radeons with 4GB had some frame time hiccups in Far Cry 4 at 4K that the 8GB models avoided:

As you can see, the 8GB Radeons avoid these frame-time spikes above 50 ms. So do all of the GeForces. Even the GeForce GTX 780 Ti with 3GB manages to sidestep this problem.

Why do the 4GB Radeons suffer when GeForce cards with 4GB don't? The answer probably comes down to the way GPU memory is managed in the graphics driver software, by and large. Quite possibly, AMD could improve the performance of the 4GB Radeons in both Mordor and Far Cry 4 with a change to the way it manages video memory.

There is one other factor to consider. Have a look at the results of this bandwidth test from our Fury X review. This test runs two ways: using a black texture that's easily compressible, and using a randomly colored texture that can't be compressed. The delta between these two scores tells us how effective the GPU's color compression scheme is.

As you can see, the color compression in Nvidia's Maxwell chips looks to be quite a bit more effective than the compression in Fury X. The Fury X still has a tremendous amount of memory bandwidth, of course, but we're more concerned about capacity. Assuming these GPUs store compressed data in a packed format that saves capacity as well as bandwidth, it's possible the Maxwell GPUs could be getting more out of each megabyte by using stronger compression.

So that's interesting.

Of course, much of what we've just demonstrated about memory capacity constraints is kind of academic for reasons we've noted. On a practical level, these results match what we saw in our initial reviews of the R9 Fury and Fury X: at resolutions of 4K and below, cards with 4GB of video memory can generally get by just fine, even with relatively high image quality settings. Similarly, the GeForce GTX 970 seems to handle 4K gaming quite well in spite of its funky partitioned memory. Meanwhile, at higher resolutions, no current single-GPU graphics card is fast enough for fluid gaming, no matter how much memory it might have. Even with 12GB, the Titan X averages less than 30 FPS in Shadow of Mordor at 5760×3240.

We'll have to see how this memory capacity story plays out over time. The 4GB Radeon Fury cards appear to be close enough to the edge—with a measurable problem in Far Cry 4 at 4K—to cause some worry about slightly more difficult cases we haven't tested, like 5K monitors, for example, or triple-4K setups. Multi-GPU schemes also impose some memory capacity overhead that could cause problems in places where single-GPU Radeons might not struggle. The biggest concern, though, is future games that simply require more memory due to the use of higher-quality textures and other assets. AMD has a bit of a challenge to manage, and it will likely need to tune its driver software carefully during the Fury's lifetime in order to prevent occasional issues. Here's hoping that work is effective.

Comments closed
    • VincentHanna
    • 4 years ago

    I don’t particularly want to buy a card with less than 8GB of gddr5 vram, as that is the “level” that console developers will be looking to fill.

    That said, the single thing that uses up the most vram in modern games is prerendered textures, and the quality of that relates directly to one’s video settings. My move to dual GTX 1280s in SLI in a few years will also likely coincide with an upgrade to 4k for that reason.

    • oldDummy
    • 4 years ago

    Coulda, Shoulda, Woulda.
    The privileged followers of this site make suggestions which sound like demands. They then get upset when their suggestions are not treated with…well…privilege.
    While knowledgeable/ well-informed; core groupies of this site are harpies.
    A major downfall of subscription services, in general: Customer is always right.

    ps- nice article.

      • Meadows
      • 4 years ago

      If you’re referring to me: I haven’t made any demands, nor was I upset.

    • DaveJustDave
    • 4 years ago

    I’m just wondering.. Would it kill them to make video cards with upgradeable VRAM? They could introduce a whole new standard and find ways to get money from those who don’t want to upgrade every cycle.

    I remember my ATI VGA wonder.. Came with 256kb of memory.. And I had to buy 8 chips to upgrade it to the 512kb and thus unlocking the elusive 640×480 super VGA resolution.

      • Meadows
      • 4 years ago

      That would only be useful if VRAM usage grew significantly faster than GPU usage, which used to be the case back in the 90’s when screen resolutions kept getting higher and higher. Which is why you remember some of the videocards from that era sporting the feature.

      Nowadays, VRAM and GPU usage increase mostly in tandem, so upgrading just the VRAM of your existing GPU probably won’t do much for its longevity.

      • robliz2Q
      • 4 years ago

      That’s odd, as I had cards with 256 or 512 KiB VRAM, which ran games fine at1600x1200.
      Wasn’t so long ago 32bit address space, 2GiB system RAM & 512 GiB VRAM was common.

        • PancersCloud
        • 4 years ago

        You were running games at 1600×1200 in 1990? …and you’re from the future when gfx cards will have half a TB of VRAM? O.o

    • ramon zarat
    • 4 years ago

    Samsung just announced 48GB HBM for GPU with 16GB version available in 2016 and 48GB in 2018… [url<]http://www.kitguru.net/components/graphic-cards/anton-shilov/samsung-expects-graphics-cards-with-6144-bit-bus-48gb-of-hbm-memory-onboard/[/url<] So if you plan to buy a new GPU, now might not be the best time. Very soon, 8GB will be the bare minimum. When the 48GB is available, the 16GB will become the bare minimum. So even top of the line 4GB cards bought today will be obsolete (not slightly less performing, entirely obsolete due to the enormous RAM gap) within the next 24 months, 36 tops! How much memory is enough? 48GB seems good for the mid term future.

      • Ninjitsu
      • 4 years ago

      No, that’s not how it works.

    • vision33r
    • 4 years ago

    Xbox One/PS4 are good examples where proper optimization on fixed hardware set can enable games to run at very rich levels on really small video ram.

    Games that received proper driver optimization will perform better with limited VRAM. Having more VRAM just gives you more buffer for poorly optimized games. I think 2GB is perfectly fine for most 1080p games, with 3GB better for 2k displays and up to 4k. With the right driver optimization it can run more efficient than wastefully consume video memory. With DX Texture Compression we can get 10:1 compression ratios without much penalty these days.

    I think only game devs need cards that feature 6GB+ where they can freely blow up texture sizes and resolution for dev work and optimize later.

      • BobbinThreadbare
      • 4 years ago

      Xbone and PS4 have unified memory. A game can use as much of the 8 GB available* as they want as VRAM.

      *the xbone launched with 3gb reserved for system, but I think they’ve reduced it.

      • Ninjitsu
      • 4 years ago

      4GB is min for 1080p and up these days, if you want to turn on all the eye candy.

        • Meadows
        • 4 years ago

        Is it? We weren’t even testing for that. </GlaDOS>

        • Krogoth
        • 4 years ago

        That’s not remotely accurate.

        4GiB of VRAM is only [b<]needed[/b<] when you are gaming at 2560x1440 or higher with all of the works.

    • ronch
    • 4 years ago

    Back in 1999 when I got my Voodoo3 3000 with 16MB, a friend asked me how much VRAM it had, and when I said it had 16MB he was like, “Whaaaatt!!! 16MB!!!!?? Wow!!!”. Haha.

    • BitBlaster
    • 4 years ago

    Damage (Scott),

    thank you for all the hard work you do keeping us informed. A sugestion for TR’s motto “The Tech Report: Come for the knowledge, STAY for the knowledge”

    A small correction in the second paragraph of the third graph about the GTX 970:

    [quote<]5860x3240[/quote<] the resolution should be "5760x3240". I'm curious about the choice of graphs. Why choose multiple line graphs with only 4 data points? Why not a single bar graph with all cards at each resolution?

    • itachi
    • 4 years ago

    How about a heavily modded Skyrim in 1440p or even 4k 🙂

    • brucek2
    • 4 years ago

    The last few sentences are correct: the fear factor on this topic is about future games, not current ones.

    I expect a graphics card I buy in Year 0 to be gradually and gracefully obsoleted in Years 2-4, in the form of lower fps and/or lowered graphics settings. That’s what happens as a GPU becomes less powerful relative to the design point of current games.

    The fear with VRAM is that instead of a graceful decline, it might be a cliff. Publishers and hardware makers may choose a new standard, implement it in a given year, and all of a sudden my card that’s still fine in terms of horsepower is rendered useless solely for VRAM reasons. Super frustrating if its a $400 GPU broken because of a lack of $40 of VRAM.

    Not sure if there’s a way to “test” for this, but maybe a helpful continuing series would be an ongoing discussion with publishers aimed at getting the best possible understanding of their internal VRAM usage map. I’d like to know several years out if possible.

      • itachi
      • 4 years ago

      Nailed it, that’s why I mentioned at this point and time, releasing the Furys with “only” was a bad move. (it got me 5 downvotes..ugh) also with how fast things are going these days, I mean my hd5870 lasted me forever 5 years or so, and ONLY started really struggling in games like metro, bf3-4, crysis 3 of course my dual core @ 4.4 was limiting me aswell but you see what I mean, in 3-5 years of time I wonder what the VRAM requirement be.. probably much more than 4 ! the VRAM requirement is growing exponentially.

      Besides when one spend 700$ (or more) on a graphic card, one expect futureproofness, and I’m talking real future proofness not just 1.5 years.

      To me 980ti vs Fury .. the choice is easily made : VRAM, VRAM and VRAM. 980ti wins, even if it’s performance was to what the Fury is to the 980Ti ;).

        • Ninjitsu
        • 4 years ago

        I doubt console focused games will cross 8GB – maybe not more than 4GB, unless of course the dev does more work for the PC port (higher quality artwork, larger textures, more objects, fancier effects, etc).

        PC exclusives may, of course – but devs are unlikely to do this unless they see enough people with cards that can support it.

      • Meadows
      • 4 years ago

      This is a pretty good point. There should be a recurring test of these things every year using whatever games are current at the time.

    • yeeeeman
    • 4 years ago

    512MB on my 9600GT are enough, for sure. I think the enough question is ireleveant. Why? Because each and every graphics card (at least on desktop side of things) has enough memory to not run into a memory starvation issue before a gpu performance one. I mean, when you’ll pass those 4gb on the Fury X, you already are at sub 20FPS level, where it doesn’t really matter anymore. So the issue is already solved even before starting to benchmark.

    • madgun
    • 4 years ago

    Thanks Scott for bringing to light the truth.

    Mind you red team nuts will come at you saying, wait for the new drivers by the end of the year to improve performance. Then wait for Shadow of Mordor to get DX 12.0. Then wait for Fury 2X to clean sweep the competition.

    For them everything’s a conspiracy hatched against AMD and all their supposed performance improvements are far away in the future.

    • kn00tcn
    • 4 years ago

    this is quite a logical fallacy, you need to do one final test with the same maximum settings EXCEPT textures to confirm that it’s not an architecture limitation

    also, what kind of poorly done game engine is not streaming data? think about how grand theft auto creates an area of pedestrians/cars/objects localized around the player, this is what textures should be like (especially on consoles)

    as an anecdote, i ran battlefield hardline beta on a 660, afterburner said only 1.5gb vram was used at all times even on ultra settings, implying that the game is dynamically allocating (edit: 2gb 660 is similar/worse than a 970 in regards to partitioned vram)

    here is a better example: [url<]https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_Titan_X/33.html[/url<] notice how call of duty is 'using' 7.3gb ram, but we all know it's not struggling on 3gb cards, so it's an intelligently designed engine that is scalable

    • HisDivineOrder
    • 4 years ago

    Interesting stuff.

    If one wanted to really close the book on the issue of VRAM, though, I’d preferred to see a more comprehensive article that includes the 2GB and 3GB cards that are more prevalent in the market, the 1080p resolution for another point on your trend charts, and more games.

    I’m not sure we can really make assumptions about trends based on the results of just one game’s usage of VRAM. For example, if I just went by my results with The Witcher 3, I’d say that Ultra in games at 1600p doesn’t need more than 2GB cards in SLI. I know this isn’t the case, but one game outlier works well with SLI for me. I wouldn’t make assumptions about all games based on how that one game does. In particular, HardOCP is pretty adamant that Dying Light is THE game to go to for testing limited VRAM. Odd you didn’t look at that if you really wanted to have the final word on the subject. There are so few games that really push VRAM atm, it’s actually surprising you didn’t just hit them all (ie., Dying Light, Shadow of Mordor, Watch_Dogs, Far Cry 4 (which you mentioned in passing with your sole reminder of how your site favors frame times instead of avg fps), and GTA V).

    I think my first reaction, though, was surprise that you didn’t really go much into frame latency with regards to these different resolutions and different GPU’s at varying amounts of VRAM. Because it seems like when VRAM runs out, you’d be more likely to see variance in frame delay.

    Still, I appreciate what you did here. Thanks.

    • sleepygeepy
    • 4 years ago

    Just wanted to thank you Scott for putting up a great article as always.

    Very interesting and informative read considering a lot of the guys (in my local forum) keep on pining about 4GB versus 6GB versus 8GB and “future proofing”. It looks like 4GB is still plenty even for 4K with the exception of some games I suppose.

    If you might have some time to re-visit this article, I would be greatly interested in how dual GPU cards stack up to higher resolutions like 1440p and 4K.

    We are getting good prices now on GTX 970 and GTX 980 videocards and many are considering adding a second one for 2-way SLI. Was wondering how will 4GB memory shape up with a second GPU adding extra performance 🙂

      • K-L-Waster
      • 4 years ago

      Exactly.

      Another way of looking at it: a 980TI costs pretty much the same amount as 2 970s. Would buying a 970 and seeing if it performs well enough for you and keeping the possibility of buying a second one later be viable? Or, would you be better off to just save up for the 980 TI right from the get go.

      • hansmuff
      • 4 years ago

      I went down a similar path back when the 660Ti was the bee’s knees in terms of power vs cost. It even had a “performance hole” in the 1.5-2.0GB memory range, cough cough 970.

      In any event, I bought a 660Ti in 2012 and decided that I can get pretty insane 1080 performance out of 660Ti SLI. And boy, I did.

      But what I have learned along the way is this: graphics performance in games evolves in multiple directions, and some directions you can guess, some you can’t.

      For instance, 2GB on the 660Ti wasn’t really much of an issue. I still game at 1080, well 1200, but there are very rare instances where I have to reduce texture quality because there isn’t enough VRAM.

      But then NV gameworks comes along and with it some new shiny things. DOF, HBAO+ etc. And suddenly my 660Ti combo chokes on things that are certainly optional, but well, nice. I can play Dying Light with absolutely fantastic performance, well, if I am OK with turning off some of those shinies.
      And this feature obsolescence gets worse as engines incorporate more and more of the shinies.

      Then another angle; SLI. I will say, it works AMAZINGLY well and I had maybe one or two games really give me problems, that’s nothing. But lo and behold along come games like Rage, and Wolfenstein: TNO and it’s successor, using an engine that by design can not and will never utilize SLI. More of those games may be coming and in those, it’s single card performance. Not really great.

      There is the frame delay issue inherent in CF/SLI. It’s not terrible and it’s something I can mostly live with just fine, but it is there.

      DirectX 12 might jumble things up somewhat, in a way that the frame delays ease up and hey, multi-GPU may become more effective and usable and you can use different cards altogether. But DX12 games are ways away, and engines aren’t going to be DX12 optimized for a while.

      All that it to say, I believe SLI is pretty great technology if you have a need for it. But it’s absolutely terrible future-proofing. I sold my 660Tis and put in a single 970. A lot of stuff runs nicer now.

    • vargis14
    • 4 years ago

    Thank you for another fantastic read. Curious did you disable NV boost for your tests? I did not read every line sorry:)

    I have a request since a lot of people cannot afford 980s and such could you test 2gb vs 4gb with say a GTX 680 or GTX 770 since I know a lot of people are still sporting 2mb cards and since I do not have a 2mb 770 to test myself and compare it to my 4gb 770 also I only have a 3440-1440 monitor.

    But if you want to mail me out and lend me a GTX 770 I would be glad to write up some archaic and probably hard to understand review of 2mb vs 4mb VRAM comparison on the 680/770 since I can clock my 770’s down to 680 levels. I would also disable boost on both cards so we do not have any clock disrepencies

      • Meadows
      • 4 years ago

      “2mb”? Seriously? Once would’ve been a typo, but you did it several times.

      • Damage
      • 4 years ago

      I didn’t disable GPU Boost or PowerTune. Curious as to why anyone would try to do so. Don’t think it’s possible, though.

    • K-L-Waster
    • 4 years ago

    Another piece of info that would be good to know is, how does this test play out on the new Damagebox with 2x GTX 970s? Does having an SLI setup scale the same way, better, or worse?

      • Westbrook348
      • 4 years ago

      I would love to see more analysis of SLI 970 too especially vs a single 980Ti. I like that we keep getting 295×2 dual GPU numbers, but I can’t figure out why SLI 970 hasn’t been analyzed at all using Scott’s frame time algorithms. I know he’s limited on time, but he did write 12 whole pages reviewing the Titan X. I’ve gotta think SLI 970s are as popular as a $1000 GPU.

      It’s Scott’s site and I respect him immensely, but I agree with people clammoring for less focus on 4K and $1000 GPUs and more focus on setups the majority of people are actually using. Either that or take a look at tech that hasn’t been examined closely for a little while. For instance, frame time scaling in SLI has been pretty much ignored since the 2011 “Inside the Second” article and 2013 update. The frequent 295X2 testing has given us good data from the AMD side, but otherwise I was only able to find the 2013 7990 article. These were interesting tests that I wish we could see occasional updates about, even though dual GPU setups aren’t very common.

      I also wish we could see some updated info about 3D vision, especially since we have VR around the corner and the ROG Swift enables stereoscopic 3D at 1440p. The last article on 3D gaming was April 2012 by some guy named Cyril, and since then a bunch of excellent 3D games have been released especially Tomb Raider, AC4 Black Flag, Deus Ex 3, Diablo 3, Witcher 3. I’m currently playing GTA in 3D and oh. my. god. It’s not even the same game as the 2D version. The 3D Vision community is still about 50k ppl who visit and use Helixmods for improved compatibility with Nvidia drivers. So hopefully we can get an 2015 update on 3D gaming, even if Scott just takes a look at Tomb Raider (I think it might blow him away, but then again he’s used Oculus Rift – I haven’t been so fortunate).

    • BryanC
    • 4 years ago

    As MathMan said earlier, Damage is wrong here. GPU color compression is bandwidth only. It does not save capacity.

    This is because during rendering, the GPU has to write new images to memory. Before the images are rendered, it’s impossible to know how well the compression will perform. Some outputs are very compressible, while others are not compressible at all. If the GPU waited to allocate memory for the result until the output was completely rendered and compressed, it would have to dynamically allocate memory for every output. This would be expensive and add overhead, but even worse – the gains would be inconsistent, making performance a strong function of the compressibility of the images being rendered.

    TL;DR: It’s not practical to save on memory size with color compression. It does, however, save a huge amount of bandwidth, so it’s worth doing.

    • Laykun
    • 4 years ago

    Excellent article. I’d live to see a yearly or even 6 monthly report on the state of the gaming landscape in terms of memory capacity to performance. I can spout on and on about how the 4GB on the Fury X won’t be enough in the future but the proof will really be in the pudding. It will be very interesting to see how consoles influence VRAM usage in PC games as time moves on.

      • Ninjitsu
      • 4 years ago

      Frankly I don’t think it matters – and I’ve said this repeatedly – today’s GPUs aren’t really good for 4K in the way they are for 1080p or 1440p. By the time more than 1% gamers have 4K, you’ll have GPUs truly capable of driving that resolution and 8GB HBM to match.

        • Laykun
        • 4 years ago

        I think it’s pretty well understood by now that this site isn’t for your ‘average’ user. Even if the articles don’t provide useful information to the 99% it’s not a 99% that actually ever consider even reading TR, where as the people who do visit the site, at least I’d imagine, would be interested in the results of the article, even if it’s only at an academic level.

          • Den
          • 4 years ago

          The Steam survey doesn’t even have data on how many people have a 980ti, much less dual 980s or better which are what is really needed for playing modern AAA games at 4K. Even here, I’m sure video cards that perform in the 960-980 range are much more common than anything better. The 980 can barely average about 60FPS on nearly all games at 1080p (nothing can handle the new Batman game at 1080p60fps as far as I know). So 1080p is probably more relevant than 4K here. And learning about how much of a limit 2gb is at lower resolutions is both a relevant and academic issue – much more relevant than looking at games running at 8fps at 8K.

          That said, I’m perfectly fine with the testing done. I’m not looking at getting a new card anytime soon so relevancy is irrelevant to me atm – I want to know what pushes the 980ti to its limits of VRAM (which unfortunately seems to be nothing). The other thing I’d be interested in seeing how VRAM usage at 1080p has increased in the last 2-3 years.

          • Ninjitsu
          • 4 years ago

          Oh I don’t know about that.
          [url<]https://techreport.com/news/27055/the-tr-hardware-survey-2014-what-inside-your-main-desktop-pc[/url<]

      • jihadjoe
      • 4 years ago

      IMO by the time the 4GB or 6GB isn’t enough to run games smoothly, the Fury X GPU itself would also be a bottleneck.

      8GB on cards now is just like those GeForce MXs with 256MB back in the day. The GPU is nowhere near powerful enough to actually make good use of all that memory.

    • Generic
    • 4 years ago

    My desire to compliment Damage on a job well done is overshadowed by my disappointment that AC Unity was not included on the graphs for our amusement. I mean if you would’ve just put a label on an arbitrary parallel line in your graphs without any explanation…

    …it would’ve made my day is all I’m sayin’

    • NoOne ButMe
    • 4 years ago

    Or, shockingly, if you are looking at 60+ FPS Max settings a Fury X will be fine on the 1440p resolution you need to drop to in most games. 4GB isn’t an issue for almost any single card at sane settings. Obviously there will be some rare case where this is not true. Same holds for the 970/980/290/290x/etc.

    Or, the fact that Nvidia “miss-communicated” about the 970 is far worse than the actual results of what they did 🙂

    • ImSpartacus
    • 4 years ago

    I just wanted to say that I really liked this article.

    The quality of article is a lot higher when you’re not handcuffed by the need to churn out pages of graphs and tables. Just did exactly what you needed to demonstrate a point and nothing more.

    • Krogoth
    • 4 years ago

    Ironically, it is because of gaming consoles is probably the biggest reason why 4GiB or more of VRAM isn’t a necessity for majority of users and games. It also helps that requires an insane amount of manhours to create the content that utillize that much VRAM.

    4K Gaming and AA/AF is going to be the killer app for 4GIB or more VRAM, but the fill-rate and shading power on modern GPUs isn’t there yet unless you like to game at 30-40FPS.

    • itachi
    • 4 years ago

    Very nice article, just as I been saying, the 8gb Hawai remake is an even better move from AMD than Fury.. well I guess in 1080-1440p its not gonna be limited, but I’m talking on the long run..

      • itachi
      • 4 years ago

      -5 vote ?! you guys got buyer regret ? LOL

      • Den
      • 4 years ago

      In the long run, even the 4gb 290 and 290xs are still not going to be limited by VRAM. 3 years from now, you should be playing 720p on the 290x to get good enough frame rates on the latest games and still not needing that VRAM. Really the only case a Fury X could need more than 4GB of VRAM is if you use XF if there is a lot of VRAM overhead caused by the XF so you can play 4K on current games. In a few years, XF Fury X will only be good enough for 1440p for the most demanding games. XF/SLI is not really useful anyways unless you get it when the card is released and you need the power then (not 3 years from now).

      If your main concern is 3-5 years from now, you should buy a $200 card now and then buy another $200 card 3 years from now and you’ll save money and have the same performance you’d have getting a $500+ card now. You should only buy a $500+ card if you want its performance now. No matter what GPU you get, you should plan on upgrading within 3 years.

    • USAFTW
    • 4 years ago

    It’s relieving to know that even 4GB on a 980 isn’t low enough to cause memory spillover to system memory, that 12 gigs on the Titan X really looks redundant now.
    What will memory usage look like with DX12 is anybodys guess – I’m guessing if the end result looks like Arkham Knight we’ll know who to blame.

      • NoOne ButMe
      • 4 years ago

      Memory usage will be higher with lazy devs, should be lower for optimize crazed devs.

      To clarify, “lazy devs” is mostly a product of a company rushing developers or telling developers to put far more into a game than needed. Think of “lazy devs” as the people behind AC:unity, Battlefield 4, etc. and the “optimize crazed devs” as the people behind games like The Witcher 3.

      Thanks to Laykun for bringing up the issues with the shorthand I wrote.

        • Laykun
        • 4 years ago

        I love this misconception about game developers. There aren’t many ‘lazy’ developers out there, it’s more of a problem of time constraints put down by publishers/managers. You can optimize a game engine to the moon and back but one day it needs to actually ship with a product.

          • NoOne ButMe
          • 4 years ago

          yes, my wording was bad. You are entirely correct.
          Wording has a note explaining it better now.

    • Nevermind
    • 4 years ago

    “Even with 12GB, the Titan X averages less than 30 FPS in Shadow of Mordor at 5760×3240.”

    Is that good? Expected? Bad? I have no frame of reference for that resolution.

      • Meadows
      • 4 years ago

      It’s kind of bad. The game does fit into VRAM just fine but the GPU can’t squeeze out pixels any faster.

        • Nevermind
        • 4 years ago

        Unoptimized or bad port code or?

          • Meadows
          • 4 years ago

          Neither, the resolution is just ahead of its time.

            • Nevermind
            • 4 years ago

            Still, 30fps? 12 Gigs of ram? Where’s the limitation, bus bandwidth?

            • Meadows
            • 4 years ago

            A resolution that high will saturate fill rate, shader capacity, and bandwidth all at the same time, but probably runs out of bandwidth last of all, otherwise the Fury X would’ve had better results.

            • Den
            • 4 years ago

            Not necessarily. VRAM seems to be the issue for Fury X- its better than the 390x in most ways except VRAM but still performs as well as a 4GB 290x at 8K. Without clearing that bottleneck you can’t really tell what the next bottleneck would be. Considering how much VRAM is being used, it really could be memory bandwidth that is the limiting factor. The fact that the 390x performs better than the 980 at 8K more than it does at 5760×3240 could provide some clues about what is limiting them.

            Surprised there have been no mentions of 64 ROPs anywhere in the comments after all the rumors of the Fury X being limited by them. Given that the 980 still scales about as well as the 980ti I’m guessing 64 ROPs are still enough for 8K resolution.

            • Ninjitsu
            • 4 years ago

            I don’t know about that – we saw 64 ROPs possibly being a limit at lower resolutions too – it just depends on the game.

            • Den
            • 4 years ago

            Could you provide a source on that? Or some explanation. Because everything I’ve seen until now has said that ROPs are used for scaling to higher resolutions and the fact that the Fury (X) scales better than nV cards basically proves its not ROP limited.

            • Ninjitsu
            • 4 years ago

            [quote<] In other areas, the Fury X's theoretical graphics rates haven't budged compared to the 290X, including the pixel fill rate and rasterization. Those are also precisely the areas where the Fury X looks weakest compared to the competition. We are looking at a bit of asymmetrical warfare this time around, with AMD and Nvidia fielding vastly different mixes of GPU resources in similarly priced products. ... The Fiji GPU has the same 64 pixels per clock of ROP throughput as Hawaii before it, so these results shouldn't come as a surprise. These numbers illustrate something noteworthy, though. Nvidia has grown the ROP counts substantially in its Maxwell-based GPUs, taking even the mid-range GM204 aboard the GTX 980 beyond what Hawaii and Fiji offer. Truth be told, both of the Radeons probably offer more than enough raw pixel fill rate. However, these results are a sort of proxy for other types of ROP power, like blending for multisampled anti-aliasing and Z/stencil work for shadowing, that can tax a GPU. [/quote<] [url<]https://techreport.com/review/28513/amd-radeon-r9-fury-x-graphics-card-reviewed/4[/url<]

      • chuckula
      • 4 years ago

      We are just approaching the point where the highest end single-GPU solutions are considered “acceptable” at 4K (for varying values of “acceptable” and Damage likely has higher standards than I do).

      Basically: Titan X isn’t even considered “good” or “great” at 4K, and the less than 30FPS showing at “5K” is pretty much expected.

        • beck2448
        • 4 years ago

        Interesting that 4k adoption is at what, less than 1%, yet all the reviews feature it as critical.

          • Meadows
          • 4 years ago

          Much less than 1%.

    • PrincipalSkinner
    • 4 years ago

    Look at Witcher 3 memory usage [url<]https://www.techpowerup.com/reviews/Performance_Analysis/The_Witcher_3/3.html[/url<] It seems quite frugal.

    • wingless
    • 4 years ago

    I have two 2GB GTX 760’s in SLI. The struggle is real! 2GB is definitely not enough. I want to get a GTX 980 just to alleviate the FPS crash when I hit that frame buffer limit. Titanfall is unplayable with AA on. Shadow of Mordor is crippled with good quality settings. My setup has no future.

      • Meadows
      • 4 years ago

      Serves you right for going with SLI!

        • September
        • 4 years ago

        That’s my observation too, everyone needs to step up from the x60 line to the x70 or x80. I went from 460 1GB to 770 2GB and don’t look back at 1200p 60Hz IPS. But I want to game at 1440p and have a feeling the 770 with 2GB won’t cut it across the board, maybe with GSync but should I upgrade to a 970 3.5/0.5GB or just go for the 980 4GB? I think this is exactly the point Meadows and others are trying to make above (hey, I thought it was polite – and no one is directly bashing you Damage we all just want MOAR charts!) so that we know where the $$$ make sense when we think about an upgrade.

        Edit: no strike-through?

          • Ninjitsu
          • 4 years ago

          use the [ s ] tag.

    • sparkman
    • 4 years ago

    [quote<]Why do the 4GB Radeons suffer when GeForce cards with 4GB don't?[/quote<] Does this honestly surprise anyone familiar with AMD driver quality?

      • Chrispy_
      • 4 years ago

      -1 for jumping to conclusions because they support your hatred of AMD.

      According to Scott’s article, GCN 1.2 gets 14% compression in the Beyond3D suite, whilst Maxwell gets 66% compression. In theoretical terms only, it means:

      a 4GB GCN 1.2 card has 4.5GB of effective VRAM
      a 4GB Maxwell card has almost 7GB of effective VRAM.

      This is [b<]hardware[/b<] compression, a feature designed into the silicon, and that's likely why the Radeons suffer whilst the GeForces don't. If it were drivers the 290X and the Fury would get the same result in that benchmark.

        • MathMan
        • 4 years ago

        You’re assuming that the compression for AMD and Nvidia is size compression in addition to just bandwidth compression.

        Size compression is simply not possible for this kind of application. It’s BW compression only.

          • Chrispy_
          • 4 years ago

          [quote=”Scott Wasson”<]we're more concerned about capacity. Assuming these GPUs store compressed data in a packed format that saves capacity as well as bandwidth, it's possible the Maxwell GPUs could be getting more out of each megabyte by using stronger compression.[/quote<]

            • MathMan
            • 4 years ago

            Yes, Scott is wrong here. See my message below.

            It can not be done based on first principles. Both AMD and Nvidia say that they compress on the fly when frame buffers are written.

            For an immediate mode renderer (AMD and Nvidia GPUs), your reads and writes happen in no particular order. You can first render a triangle at the bottom of the screen, and then one at the top. That means: you need full random read and write access over the full buffer.

            For size compression, the right mental model is a zip file: the only way to make changes is to decompress fully and then compress everything again. That obviously won’t fly.

            Conclusion: it can only be BW compression. (In fact, it does a little bit of size expansion: one way or the other, you need storage to indicate whether or not a particular rectangle in the buffer has been compressed or not.)

            (This is different than texture compression, which has existed since forever: there you only need to reads, which changes everything.)

            • Ninjitsu
            • 4 years ago

            I’m curious: I know 7-zip has solid and non-solid archives. What you’re talking about seems like solid archives; would non-solid be too slow or have too much overhead in this situation?

            • MathMan
            • 4 years ago

            Even with non-solid archives, if you changes the first file in the archive and that file, in compressed form, becomes larger, you still need to move up all the files behind it to make room.
            A better analogy is this: if you’d only compress files that are 10KB and you reserve 10KB for each file, but then you store each file within its reserved 10KB in compressed form.

            In that case, when you do reads or writes, your file traffic is much lower, however, the size of the archive is still the same as what it would have been without compression.

            • eofpi
            • 4 years ago

            You’re correct for framebuffers, but 7680x4320x4 bytes = 127MB. Coincidentally, that’s the same size as 4K with 4x FSAA. Even triple buffered, that’ll fit easily on the slow channel of a GTX 970.

            Much of the memory space is used for textures, which are pretty much read-only. Those would be perfect candidates for compression on load, which really would save some memory space because allocation can be done after compression.

            But here my knowledge of modern game engine GPU memory usage breaks down. Texture space is basically constant for a given GPU, game, and scene. Framebuffer scales with pixel count, but even with triple buffering it doesn’t grow very fast.

            So where does the rest of it go? Does persistent geometry get assembled, textured, possibly lit, and stored in GPU memory for reuse? If so, wouldn’t that be a good candidate for reallocation for efficiency?

            • Freon
            • 4 years ago

            From my understanding, there are also all sorts of other back buffers, environment maps, cube maps, etc. that come to play on top of frame buffer size that may be directly affected by frame buffer size depending on what and how the developer chose to implement.

            Unfortunately not sure there’s a perfect answer to this. Some tools can query VRAM use, but it seems even that may not be the whole picture depending on how the driver is managing memory and paging.

            • Chrispy_
            • 4 years ago

            I was also under the impression that GPU’s reserved VRAM to run shader code too.

            Games with low texture loads like Bioshock Infinite (and its DX9 previous-gen console target) can still gobble up VRAM because there are so many shader effects, postprocessing and other things going on that the GPU has to deal with. The common ones that eat up the most VRAM that I know of are:

            Temporal AA
            Screen-space or horizon-based ambient occlusion
            Depth of field
            Motion blur
            Soft shadows
            Realtime light calculations

            The first four of those increase in VRAM demands with resolution and the last two increase based on the number of lights and resolution of the shadow calculations.

            • auxy
            • 4 years ago

            That’s mostly to do with extra offscreen buffers. (*‘ω‘ *) The last two effects don’t use much RAM, but instead dramatically increase load on shaders.

          • Den
          • 4 years ago

          How can you have BW compression without shrinking the size? The BW a file takes up would be determined by its size, would it not?

            • MathMan
            • 4 years ago

            Here’s the worlds simplest BW compression: you have an on chip memory that has 1 bit per 8×8 pixels.

            When all pixels are black, you set that bit to 1. Otherwise you set it 0.

            When you need to read any of those pixels in that square, you discard the read and return (0,0,0) back right away.

            You still need the memory for cases where not all pixels are black, but when they are all black, you’ve achieved infinite BW compression for that square.

            Expand to more complicated algorithms as needed.

            Check out the Maxwell slides with that mindset: it’s exactly what they’re doing.

        • sparkman
        • 4 years ago

        -1 for jumping to conclusions because they support your hatred of people.

        I own multiple systems based on AMD and Intel and nVidia.

        AMD’s has a well-deserved reputation for having lower quality than either nVidia or Intel. Their hardware and software generally works fine, but…

      • Nevermind
      • 4 years ago

      AMD drivers are fine with two caveats, brand new games and running crossfire on new HW.

      If you’re running 6 month old hardware 1 card at a time, you have zero issues for the most part.
      If you’re early-adopting 380’s and trying to x-fire onto 3 monitors, you might hit a snag.

      • NoOne ButMe
      • 4 years ago

      in their games tested. Would be interesting to see a sufficient sample size of games to get an answer that is at least some reasonable backing.

      Of course, that would be an unfair burden on almost any site to do. Maybe a “if we raise XXXX/XXXXX will you do it Damage? *begs*”

      *************************
      Well, Nvidia appears to have lovely Windows 10 drivers. Good thing it hasn’t been one to the public if anyone puts a little effort into trying it.

      Oh, wait. Months and months.

      Nvidia’s drivers for games are in the worst case on par with AMD (some outlier games exist that break this rule) but for outside of games and professional applications? Nvidia is at best in even footing.

      Windows 10 runs perfect with all my systems (VLIW5, VLIW4 both APUs and Fermi) however I’ve heard enough horror stories about Nvidia on win10 and far fewer on AMD’s side.

      Of course, I am biased towards AMD, but, well. I tend to visit forums with people who own both cards and a lot more issues on Nvidia from what I have seen. But, hey. No WHQL driver for 6 months means AMD is dead.

    • Klyith
    • 4 years ago

    I always come back to the comment by a nvidia guy in the Fury review, who dismissed 4GB as a problem and said that they’ve been using memory inefficiently, because they can always throw more on the card. That strikes me as an honest answer when given the opportunity to score points on the opposition.

    Of course it still might suck for a little while to have a lower-memory card while everyone else is still being inefficient, but in nvidia’s next cards use HBM as well it’ll probably be sooner rather than later.

      • Ninjitsu
      • 4 years ago

      That was an AMD guy.

    • anotherengineer
    • 4 years ago

    I like this resolution scaling better Scott 😉

    [url<]https://techreport.com/review/4104/ati-radeon-9700-pro-graphics-card/5[/url<] edit - 12:00 AM on September 16, 2002 Has it been that long!!??

      • Anovoca
      • 4 years ago

      Parhelia?

        • Firestarter
        • 4 years ago

        [url<]https://techreport.com/review/3735/first-look-matrox-parhelia-512-gpu[/url<]

          • Anovoca
          • 4 years ago

          I remember the geforce 4 release and even the radeon with its rumors of starting houses on fire, but that brand escapes me.

            • Chrispy_
            • 4 years ago

            Matrox, from the days of 3DFX, Cirrus Logic, Tseng Labs and S3.

            • Mr Bill
            • 4 years ago

            I’ve got a Matrox Millennium P750 lying around somewhere and half a dozen G400’s which used to be very handy for quick builds on the old hardware.

        • colinstu12
        • 4 years ago

        Matrox Parhelia. When they were a competitor in the graphics card scene.

      • Ari Atari
      • 4 years ago

      Boy they even spoke of the bit depth of the resolution.

      1600x1200x32

      I remember the time where windows had 8bit, 16bit, and 32bit colors… windows 7 was the last windows to even allow 16bit colors.

        • Firestarter
        • 4 years ago

        I remember somewhat preferring 16bit over 32bit back then, the dithering looked good to me and my S3 Savage 4 Pro needed all the help it could get

          • jihadjoe
          • 4 years ago

          Lol and 3Dfx saying “16bits ought to be enough for everyone” because none of their cards could render in true color for a while.

    • lycium
    • 4 years ago

    Hmmm, I hope my brand new 970 will be ok for 4K overall :/

    Thanks for this Scott, it’s a day too late for my purchase but very informative!

      • Meadows
      • 4 years ago

      You’ve seen the graphs, it should be okay.

    • Prestige Worldwide
    • 4 years ago

    Hi Damage,

    Interesting analysis.

    Just wondering, why weren’t measurements of average and maximum VRAM usage gathered using a program like GPU-Z or MSI afterburner and documented in the article?

    I think it would be a useful metric in this case.

    Cheers

      • Damage
      • 4 years ago

      As I noted:

      [quote<]Trouble is, most of the available tools track video memory allocation at the operating system level, and that's not necessarily a good indicator of what's going on beneath the covers. In reality, the GPU driver decides how video memory is used in Direct3D games.[/quote<]

        • Prestige Worldwide
        • 4 years ago

        Touché

        I control+F’d for “usage” but not “allocation” as I skimmed through the article with this thought in mind.

        Thanks for the enlightening copy / paste.

        • Milo Burke
        • 4 years ago

        I look forward to the possibility of AMD and Nvidia providing you vendor-specific tools.

          • Prestige Worldwide
          • 4 years ago

          I would also be interested in seeing the discrepancy between GPU-Z and vendor-specific tools, since we plebs will likely never get access to said tools.

            • Ninjitsu
            • 4 years ago

            Actually, that might not be the case…
            [url<]https://developer.nvidia.com/nvapi[/url<]

    • K-L-Waster
    • 4 years ago

    I find the Fury X results very interesting. At the extreme resolutions, it tails off to the point of not being able to beat the 290X, even though it has a more powerful GPU and faster memory. That may indicate a more fundamental limitation of GCN in terms of getting data to and from the frame buffer (possibly compression efficiency).

    Of course, this is largely academic, because very few people actually game at that res. At 4K and under it clearly performs very well, and the 4GB capacity isn’t holding it back at all.

    It is good to know that a FuryX or a 980 (non-TI) would be a perfectly workable solution at 4GB and 4K (even more so at 1440).

    • Anovoca
    • 4 years ago

    I can’t believe I am asking this question but would you suspect 8gb on Mantle to have any different result? I don’t care about Mantle specifically but since we cant test low API DX12 just yet that might give us a small forecast on what gains might be there.

      • Damage
      • 4 years ago

      So this question has several answers.

      If you’re talking about a port from DX11 to Mantle, I doubt there would be a big difference in memory usage given the same game assets and display resolution and such. But I really do not know about that.

      With the next-gen APIs, though, the game should change since developers will have lots of direct control over how memory is allocated. I’m not sure what exactly that means for the efficiency of applications, but I suspect it will depend a lot on the developer. Some games will likely be more efficient and others less so.

      Looking further out, the whole point of having more efficient draw calls, thread dispatch, and API overhead generally is to allow games to do more, look better, etc. So eventually, the best DX12/Vulkan games ought to use more GPU RAM than today’s games. That’s exactly what they should to do take advantage of the new APIs.

      Happy to take insights from game devs and such, but that’s a start on it, I think.

        • Anovoca
        • 4 years ago

        Alright, that kinda goes along with my suspicions. The ironic part to that too is that, even if you wanted to “plan ahead” for DX12 and say, spend the extra cash for an 6/8gb card, the price difference alone will be equal to a new card when the next gen is released. (assuming something doesn’t happen in supply chain that causes a dramitic increase in nvidia/AMD pricing model). So even the spend extra now to be prepared for later approach is kinda pointless.

          • Firestarter
          • 4 years ago

          spending extra to be prepared often doesn’t work out when you do the numbers, but you have to remember that you get to enjoy the faster card in the meantime too. If that alone is worth the extra money, then having the card last longer is just the cherry on top

    • Meadows
    • 4 years ago

    You should test 2 GiB cards as well.

    We know from the latest Steam survey that [b<]over 50% of gamers with DX 11 class hardware have less powerful GPUs than a GTX 770 (or equivalent)[/b<], and it's sometimes [i<]significantly[/i<] less powerful, since a great chunk of that figure is composed of low-power and integrated GPUs still. You don't want to show whether Expensive Card A is overkill as opposed to Expensive Card B. What you want to do is give that [b<]50%[/b<] some examples that hit closer to home and give them reasons to upgrade, and by extension, reasons to work your ad revenue.

      • JustAnEngineer
      • 4 years ago

      GeForce GTX960 2GiB vs. 4GiB, for example?

        • Meadows
        • 4 years ago

        Not only that, but “older” cards as well. 600 and 700 series cards from NVidia, and, well, most of AMD’s cards are already old.

        By and large, most people upgrade less and less often nowadays; for this reason, optimally you don’t want to review just those cards that got released in the past 12 months but also older ones that your readers will still have. The readers/audience need concrete reference points and helpful comparisons to better nudge them towards a potential purchase, or at the very least become informed enough.

        For example, the last time I saw a GPU review that was useful to me at TR was in January this year (GTX 960 review); they included a few older Geforces for reference there. That’s the only one in the past 8 months or so. Since then, it’s been exclusively about testing the latest cards against the latest cards, at resolutions literally nobody uses – according to the Steam survey -, and things have kind of drifted off to a strange place for me.

          • bittermann
          • 4 years ago

          I was disappointed to say the least when I saw no 1080p numbers or 2GB cards. More interested in how the lazy or rather un-optimized console porting to pc was affecting 1080P. Oh well…

          • Damage
          • 4 years ago

          We tested a GTX 780 Ti and an R9 290X here:

          [url<]https://techreport.com/review/28612/asus-strix-radeon-r9-fury-graphics-card-reviewed/4[/url<] You seem to be conflating age with price, perhaps. That's a 700-series card and a similarly-aged 290X. Also realize that it's not always practical to do real gameplay testing at 4K with much older cards. They're too slow. Also realize we have been reviewing the new products that have been released in recent months, which are high-end cards. No extra time in my schedule for other stuff. Anyhow, you seem to be making a self-interested request for low-res/low-price testing based on your own interests. Rather than admitting that, though, you're couching it as if you were trying to tell me how to do my job better. I tend to find honest requests from readers more persuasive than lectures of this sort. I think most people would!

            • Meadows
            • 4 years ago

            I realise that many of the GPUs widely used today would be too slow at 4K. Then again, please do realise that according to the Steam survey, approximately 0.07% of gamers use that resolution (yes, less than [i<]a tenth[/i<] of a percent), so pardon me if I bounce that "self-interested" remark right back at you for sticking with your 4K so vehemently in the face of it all. I do apologise but, regardless, I'm not the first person noting these things. I appreciate the work you do and I can see full well that you're in a difficult position, juggling time constraints, important new product releases, and microphone pop filters. It's also not beyond my understanding that 4K brings out the differences between GPUs better, so it has a place in reviews for magnifying strengths and weaknesses. I even understand that your high quality GPU testing methods in general end up being more time consuming than what competing sites are willing to go for. But for pity's sake you even had retro performance comparisons for CPUs recently (featuring quite old ones too!), so it feels like the GPU side of things is lacking love just a tad. One possible way to incorporate testing 1080p and 1440p would be to see whether the GPUs can produce a stable 120 fps. That's a kind of holy grail for many people much the same way how 4K is to you.

            • bittermann
            • 4 years ago

            1080P is still the standard resolution and the most popular one. There is also the nagging question of why 2GB cards are no longer enough for quite a few 1080P games when the graphic’s fidelity is not that much different/better than 2 or 3 years ago. Scott has done a great job with what is there but time constraints would suggest adding 1080P rather than 4320P? Good job regardless!

            • nanoflower
            • 4 years ago

            Here Here!!! While it’s interesting for me to read about 4k and up tests to see just how far the hardware has developed on a more practical basis I’m far more interested in how the various cards perform at 1080P as that’s where I am now and am likely to stay for the next few years. Knowing what cards will perform well at that resolution with the latest games with all the pretty baubles enabled is useful information. Understanding just what resources are needed such as how much memory do you really need at that resolution and up would be great information. Especially if it turns out that 2GB isn’t really enough if you want to enable everything at 1080P.

            • cegras
            • 4 years ago

            I completely agree with you about being able to hit 120 fps. 60 fps is the bare minimum.

            • PancersCloud
            • 4 years ago

            A note on Steam surveys—they can be useful for some usage statistics, but the vast majority of PC gamers do not use Steam, and a large number of Steam users opt out of the totally optional hardware surveys. Just saying—you’re basing a lot on a very limited source of information.

            • Ifalna
            • 4 years ago

            I do think a separate test for 1080p hardware including a few older cards would be useful to folks, since most peeps still game at that resolution and typically monitors outlive several GPU upgrades and effectively limit them in terms or resolution.

            I agree that it would make no sense to test these cards alongside the “elite cards” because they lack the horsepower to do anything at 4K+.

            • Freon
            • 4 years ago

            I find your tone here very disheartening. 🙁

            • Damage
            • 4 years ago

            Eh, read Meadows’ posts again. He’s just rude.

            I’m happy to try to test lower-end cards when I can squeeze it in, and I’m always happy when folks want to see more stuff from us. That’s excellent.

            I just wish people would be nicer to one another and to us. Asking for something by “calling out” the author of an article and telling him why he was wrong to write it–not really the best way to persuade people. Seems Meadows has few other approaches available to him, I guess. Kind of saddens me. Was hoping he’d consider a change.

            • Freon
            • 4 years ago

            I’m rereading the posts over and over and just not seeing anything that I’d call rude… I don’t see any sort of “calling out” here either. At worst “you should…” from the first root post is direct, but I’d not call that rude.

            I think his comments are valid and should as constructive. There’s no abusive language, no name calling, no personal attacks, no “WTF” or “I can’t believe you didn’t…”, and he even backs up his criticism with statistics from Steam install base which are entirely relevant to his post and the subject at hand. Perhaps TR readership is slightly up the food chain with regards to having fancy video cards, but as he says, aiming to expand readership isn’t a bad thing.

            I completely understand you can’t do it all, that’s just life. I don’t see Meadows’ posts as attacking you for not doing everything all the time.

            I just truly don’t understand what you’re upset about in his posts. I’m trying my damnedest but just not seeing it.

            • Damage
            • 4 years ago
            • Damage
            • 4 years ago

            Eh, I give up.

            • Meadows
            • 4 years ago

            I regret to say I tried my best not to be rude. Assertive and perhaps a touch stubborn, yes, but tactful at the same time. …Or at least that was my intention.

            If you would point out anything in particular that offended you, then I’ll try not to repeat that in the future.

            • Wonders
            • 4 years ago

            [quote<]I just wish people would be nicer to one another and to us. Asking for something by "calling out" the author of an article and telling him why he was wrong to write it--not really the best way to persuade people.[/quote<] I agree that commenters here could easily be more respectful of the articles they comment on, of the authors, and each other. It's certainly the case that Meadows was condescending to you (not overtly rude, yet inappropriately disrespectful). At the same time (bear with me), there is also this salient point that GPU testing at lower resolutions could be legit awesome. There's something sweet about the fact that if I can't afford a better monitor, I am rewarded with faster and smoother frame delivery. It's the kind of win-win "freebie" that I can understand many folks getting emotional about. Scrappy they may be, these are people's *rigs* we're talking about here (and everything that goes with that term), so it's natural that emotions quickly become involved. And thus people glom onto articulate comments aligned with their gut instincts, whether tone-deaf or contrarian is no matter at that point.

            • Meadows
            • 4 years ago

            Apologies if some of you found me disrespectful, any jabs I might’ve had were meant to be friendly. Next time I’ll remember to add auxy’s weird emoticons for humorous effect, or something.

          • green
          • 4 years ago

          [quote<]Since then, it's been exclusively about testing the latest cards against the latest cards, at resolutions literally nobody uses - according to the Steam survey[/quote<] i can understand where you're coming from i would love to see a 6 series* card thrown into the mix but from a realistic standpoint, i completely forgive not putting in the older cards i really don't need to see a fury X get bottlenecked by cpu at the world's 2nd most popular resolution of ~1366x768 i mean, if we're including the resolutions that people actually use, then the 2nd most popular resolution should be included ahead of any resolution other than 1080 right? and if we're testing most popular, shouldn't we be benching dota2, cs:go as these are consistently the most popular games played by "gamers"? in the last 8 months, the latest reviews here were for high-end class products from both manufacturers in the high-end, you kind of want to see how far they've pushed those the limits of graphics processing it's good as it gives me an idea of what to expect in the mid-range in 2-3 generations time mid-range class card reviews is where i'd want to see the popular stuff show up (aka. 1080p) and it's the mid-range class cards where i pay most attention note: i mean back to gt6600. i'm that far behind on my upgrades

          • auxy
          • 4 years ago

          “literally” doesn’t mean what you think it means. (*‘ω‘ *)

            • Meadows
            • 4 years ago

            Yes it does, but I wanted to be a bit dramatic.

      • Chrispy_
      • 4 years ago

      Well if a low-end card has 2GB it’ll be fine, since the GPU isn’t powerful enough to push framerates at resolutions requiring more than 2GB.

      If you want to play at 4K on a GTX770 or less, then you’re welcome to, but don’t blame the slideshow on a lack of VRAM beause those cards couldn’t produce playable framerates even if they had 128GB of VRAM.

        • Ninjitsu
        • 4 years ago

        [quote<] If you want to play at 4K on a GTX770 or less [/quote<] Meadows never mentioned 4K. I'm pretty sure they mean testing at 1080p and/or 1440p.

          • Chrispy_
          • 4 years ago

          I mentioned 4K as 2GB is plenty enough RAM for lower resolutions.

          By the time you push resolution and detail settings high enough to use more than 2GB of VRAM you’re already into low framerates for non-VRAM related reasons.

          The GTX[s<]680[/s<]770 is a decent performer at 1440 if you're okay with 30-40fps but it's limited by fillrate far before it runs out of VRAM. AMD's less powerful 2GB cards - the 7870/270X are even more fillrate limited so even 1440p is stretching them too far to worry about VRAM limitations. The interesting question at 2GB is in modern cards that [b<]*do*[/b<] have the fillrate, if only just - and that's the GTX960 and Tonga-based 285 and 380 variants. TR and other sites have already confirmed that 2GB is a poor choice for GPUs with that much pixel-pushing power, and given that 4GB variants are only $10-20 more, why would anyone short-change themselves on RAM? I think it's fair to say that every review I've read of 2GB Tonga/GTX960 has condemned the lack of VRAM as a bottleneck holding back performance scaling, but only because the cards have the potential to perform better with 4GB. With the older cards, VRAM isn't the botteneck because the GPU runs out of grunt first, and by a large margin.

            • Ninjitsu
            • 4 years ago

            Hmmm. I see. What would you say about games that need the extra VRAM for textures, mainly? Would the 770 struggle there as well? I’m mainly curious about such games, because what you’re saying is what I thought as well when the 680 came out. But it seems (from my own testing) textures don’t impact performance much if they have enough VRAM.

            • Chrispy_
            • 4 years ago

            I think it’s fine, Textures do impact VRAM but all the high-VRAM using games tend to be ones that have pretty standard texture quality but bucket-loads of fancy graphical effects and shaders.

            Ergo, I think it’s a combination of graphics shaders (quality settings) and higher resolutions that chew up VRAM the most, and for those two things the older GPUs can’t process the shaders or push the pixels fast enough at the resolutions we’re worried about for 2GB.

            Maybe a new game will come out that needs 8K textures and be light on graphical eye-candy. In that scenario I can absolutely imagine 2GB cards struggling, but I don’t know of any current or planned games that fit this bill, more and more it’s down to shaders and GPU-computed stuff like SSAO, realtime shadow processing, volumetric fog and lighting etc – and these are things I had to turn down on my 7970 years ago with Metro 2033. Cards have been fillrate and compute limited long before VRAM was an issue, and it’s only really become and issue at 2GB since 4K gaming came along 😉

            • Ninjitsu
            • 4 years ago

            Hmmm. I play Arma 3 a lot (which is why I keep talking about it so much), and while my GTX 560 is indeed a bottleneck (so at 1080p, it may not be able to cross 40 fps at “high” settings), my CPU is an even bigger bottleneck. So most of the time my GPU isn’t utilized above 60% or 70%. I’m in the 900MB zone of VRAM utilization, so I know “ultra” textures and terrain will push it over 1GB, but I really won’t see much impact in frame rates.

            Similarly, Total War Rome II easily exceeds 1GB, and in such a case automatically sets everything to low, even though the GPU isn’t really being used much because of the CPU bottleneck.

            I use custom settings in both games, but point is that both games (especially Rome II) could cross the 2GB barrier without much trouble at 1080p.

            I don’t remember how it used to be with Far Cry 3 (in terms of VRAM), and Far Cry 4 seemed okay on a friend’s 670, so I suppose it’s okay. I remember an Nvidia guide stating that 2GB is enough, though the game caches data in as much VRAM as it gets.

            I don’t think I’ve faced VRAM issues with other games, either my GPU or CPU craps out before that’s a problem.

            EDIT: Though I think I should add that I don’t have any non-indie 2015 game on my PC, except Cities: Skylines. Maybe even 2014 for that matter.

            EDIT2: I do- Wargame: Red Dragon

            • September
            • 4 years ago

            I agree with you so much I want to see this in a chart!

      • Krogoth
      • 4 years ago

      I think the fill-rate and shading power of the older generation of GPUs will be more of a bottleneck than the memory capacity.

      A comparison between 960 2GiB and 4GIB will help make this very clear since it has roughly the power of a 660Ti/670.

        • bittermann
        • 4 years ago

        There are plenty of 2GB 680’s out there that could be used for testing as well. To think 2GB cards and 1080P is irrelevant is simply denying the current main user base. Pretty important if you ask me. They have plenty of power to push past other factors than vram. Denial is a beautiful thing. 😐

      • derFunkenstein
      • 4 years ago

      2GB cards already can’t run the highest texture settings well at 1080p in recent games. For example, one of my PCs has a 2GB R9 270X. Picking the highest texture settings in GTA5 isn’t possible without ignoring suggested limits, and then once you override that, turning a corner slows down pretty considerably.

      If you want Scott to test it to satisfy your own curiosity, you can do that yourself with DSR and your GTX 770. If you want him to test it to see if it’s worth skimping $20 to get a 2GB GTX 960 over a 4GB GTX 960, then don’t get a 2GB card. It’s already limited in recent games.

        • Meadows
        • 4 years ago

        What I personally have is an overclocked GTX 660 Ti. According to my tests, it’s something like two hairs short of the performance of a factory-clocked GTX 680 and that’s a result I’m actually satisfied with. It’s actually not far off the level of a 770, but not quite there either.

        Thing is – if you guys want to make this personal -, I really have no reference points to decide just how much better a prospective purchase would be for me.

        Then again, full disclosure: I will not buy a new GPU this year regardless of any test results and even H1 2016 is dubious. I just thought to speak out for the silent masses because it’s statistically impossible that I should be alone.

        • Freon
        • 4 years ago

        I almost suspect that at 1080p the highest resolution textures in the most obtuse games (like SoM) never actually have their highest detail mipmap rendered to screen. Besides shoving the virtual game camera into a wall so a texture is blown up on screen, most surfaces are far enough away, and viewed through a relatively (to the texture resolution) rough screendoor of 1080p. Those texture pixels are often just super tiny in terms of view arc, much smaller than the view arc of the screen resolution pixels.

        I’m guessing in many cases there’s virtually no difference in image, and those higher mipmaps sit fairly idle and un- or rarely-used in memory. Maybe Nvidia pages those out of VRAM more efficiently than AMD…

        I’d love to see a test, but running direct diffs of screen caps would also be interesting as a leveling of what really matrers.

          • derFunkenstein
          • 4 years ago

          If you’re already at a point at 1080p that you can’t max out a game, then why does it matter how it scales beyond that?

            • Meadows
            • 4 years ago

            He never said anything like that.

            • derFunkenstein
            • 4 years ago

            Correct. I did. He conceded it when he said it didn’t matter. So why is scaling with low VRAM interesting again? At higher resolutions, you’re compromising image quality to begin with.

            • Meadows
            • 4 years ago

            It’s interesting because 1080p is relevant (and, to a lesser extent, 1440p is too) and the majority of new releases *can* still be cranked to “Very High” quality levels on previous-gen GPUs, if not to Ultra outright (depending on how much you can ignore hiccups), even if they have 2 GiB of VRAM.

            It also makes sense to test more recent GPUs at 1080p because high framerate matters to a subset of gamers. If I had to choose, I’d personally pick 1080p at an average of 120/144 fps rather than 2160p slogging at an average of 40 fps, but sadly we don’t have a reliable impression on high-framerate gaming and I’d rather not base my opinions on the tests of websites other than TR.

            People in general seem reluctant to upgrade their monitors even when they happen to upgrade their PC hardware, meaning that testing 1080p carries an added benefit to readers, showing them prospective upgrade gains in an apples-to-apples comparison.

            • bittermann
            • 4 years ago

            Ridiculous how so many people are against testing 1080P/2Gb cards when its the most common resolution. If what they say is true then test it to find out once and for all. Don’t just throw out magical numbers in a post. I do not believe the 2GB vram threshold is whats holding these cards back in EVERY game. Prove it to me…

            • derFunkenstein
            • 4 years ago

            Good grief. Nobody said every game. I said GTA5. The post is about one particularly harsh title. I don’t see why you’d buy a new 2GB card today, though, because that’s where things are going. A 2GB GTX 960 whena 4GB version is a handful of dollars more. Same with a 4GB R9 380. When the difference is like 2% of the total system cost or 5% of the cost of the card, why are we even considering buying a 2GB card?

            I mean, seriously, ten bucks.
            [url<]http://www.newegg.com/Product/Product.aspx?Item=N82E16814202150&cm_re=Sapphire_R9_380_2GB-_-14-202-150-_-Product[/url<] [url<]http://www.newegg.com/Product/Product.aspx?Item=N82E16814202149&cm_re=R9_380_4GB-_-14-202-149-_-Product[/url<] Or $20 for Nvidia, which isn't quite as great but hardly a reason to skimp. [url<]http://www.newegg.com/Product/Product.aspx?Item=N82E16814487093&cm_re=2GB_GTX_960_evga-_-14-487-093-_-Product[/url<] [url<]http://www.newegg.com/Product/Product.aspx?Item=N82E16814487133&cm_re=4GB_GTX_960-_-14-487-133-_-Product[/url<] Why is this a concern? If someone is trying to make themselves feel better about their older card, and if they're happy with their older card, that's great. 14/16nm GPUs are bound to come next year (though I have no first-hand knowledge). They would be wise to wait for those if they're satisfied with current performance. But those people shouldn't color someone else's purchase with their own rose-colored 2GB glasses.

            • bittermann
            • 4 years ago

            WTF…who said anything about buying a new 2GB card? You don’t think there are plenty out there already? Have you not seen the latest steam survey? Go cry somewhere else if you don’t like my common sense replies. I have to justify nothing for you. The numbers speak for themselves.

            • derFunkenstein
            • 4 years ago

            Sure, there are plenty out there, but the people that have them can try DSR at whatever resolution they want and find out if the games they want to play perform well.

            • bittermann
            • 4 years ago

            Why would I use DSR for 1080P?

            • Meadows
            • 4 years ago

            In place of standard 4x AA, for example, but it’s not a stellar idea because on midrange cards you might run out of pixel pushing power even faster than VRAM.

            • Meadows
            • 4 years ago

            You’re entirely off track. Nobody’s explicitly recommending buying a new 2 GiB card today.

            I recommended [i<]testing them[/i<] so that we have more points of reference for prospective upgrades, i.e. to answer the questions, "[i<]how much better[/i<] will this new card be for me, in [i<]which games[/i<], and [i<]why[/i<]?"

            • Ninjitsu
            • 4 years ago

            Doesn’t the same argument apply to GPU testing in general, though (why does 4k matter when 1440p, 1080p can’t always be run without rendering longer than 16.7ms)?

            • derFunkenstein
            • 4 years ago

            No, not really. My example was a game where a user is locked out of an option completely, unless they choose to bypass it. I’m clearly in the minority, I guess.

      • itachi
      • 4 years ago

      You’re talking about peasant stuff right there.

      Joke aside, alot of kids on Steam play CS:GO or Dota 2 for instance which doesn’t require a good graphic card, let alone a good PC, but you’re right though, just for reference.

      Usually VRAM is a concern more for the like of us high end user, but I will mention that AMD killed it by increased VRAM on their 7000’s series card, to 3gb and later on with r9 200’s series card with 4gb, less so with the Furys…lol.

    • klagermkii
    • 4 years ago

    One thing i do wonder about is if any of the cards start “cheating” when under memory pressure. Would it be possible for cards to recompress existing textures at smaller size/lower quality if they find themselves overcommitted on RAM? Could you hide this kind of action behind the DirectX/OpenGL API, or are there guarantees of losslessness.

    It feels like one of the easiest ways to avoid the performance cliff, and really, who’s looking at the fine details at 3840×2160+.

      • auxy
      • 4 years ago

      Higher resolutions are WHEN you are looking at the fine details, as you have enough image data to do so. At lower resolutions is when you can hide poor textures, because until you get pretty close to them they’re barely relevant anyway.

      As far as your actual question, well, I know games used to do texture recompression on the fly by design, but I don’t know how much of that actually happens now. I suspect not much, as compression is computationally intensive and games are using gigabytes of texture data in a single scene. (*’▽’)

        • klagermkii
        • 4 years ago

        I don’t know if it’s true that higher resolutions means you’re looking at finer details. You’re running at a much higher res on a similar size monitor so your pixels are way smaller and you’re less likely to notice the kind of detail that gets lost in lossy compression. It’s like if you’re doing a 320×200 game each pixel is significantly more important and the fine details in the edges matter more. Again, I don’t know, it just seems like it would be the case that in a sea of pixels the importance of each individual pixel is diminished.

        On the second point I think the rise in framebuffer compression on newer cards has shown that the computation cost of even repeated compression/decompression on a single frame can be cheaper than moving around more data.

          • auxy
          • 4 years ago

          No, I’m sorry, that’s not how this works. Nevermind that I can discern individual pixels on a still image on a 24″ 4K monitor; in motion, it’s MUCH more obvious — even a single pixel. This is why anti-aliasing is so important; it’s not (and has never been) about “jaggies”, it’s about texture shimmer and jitter caused by insufficient resolution in the original sample.

          This is a really basic concept — more pixels on screen means more chances to convey color data from a texture. That’s all there is to it. Arguing otherwise is like arguing that maybe lower velocity gives less momenum.

            • klagermkii
            • 4 years ago

            You can’t tell me that aliasing is more noticeable on a 24″ 3840×2160 screen vs 24″ 1920×1080. And I’m not saying you need less data for the textures you use 3840×2160 vs 1920×1080, just that you don’t need to quadruple your data to achieve the same apparent quality because the visible details are significantly smaller. That’s where the opportunity for compression comes in.

            • auxy
            • 4 years ago

            No, and in fact I said that this is the reason we need more resolution: to reduce aliasing. Why would you say that I said literally the opposite of what I said? ( ゚Д゚)

            The thing is, visible details AREN’T smaller. You don’t change your viewpoint by changing resolution; it’s not a still image. You are actually sampling the same object four times as many times (and that gets multiplied if you’re multisampling or supersampling!) when at 4K UHD vs. 1080p FHD, so you can see four times as much detail. The visible details aren’t smaller, they’re clearer, which means compression is only going to make it look that much worse.

            • Den
            • 4 years ago

            If you have 4 pixels for every one, then if you improve 3 of them and leave the 4th one as being the same as one of the other 3, it would still be clearer than 1080p while not being as demanding as true 4K, would it not?

            • auxy
            • 4 years ago

            How will you ‘improve’ the other 3 pixels?

            • Den
            • 4 years ago

            In 1080p, you have 1 pixel and its one color. In 4K you have 4 pixels who can be colored independently. It improves the image. Even if you don’t treat every single pixel individually, you could still get an improved image with 4K.

            • Meadows
            • 4 years ago

            If you’re talking about texture interpolation to “fill out” the extra screen space, then 4K actually increases the ranges of interpolation needed and it’ll look no better unless the output depth is greater than 8 bits per channel (24-bit colour).

            Since all games output 24-bit colour today, the banding/textures will look exactly as they do at 1080p and your only two advantages will be sharper model edges and greater useful viewing distance on textures with alpha (before they get mushy).

            • Den
            • 4 years ago

            Nearest-neighbour “interpolation”. Yes you need more than 1080p, but I’m saying that in demanding situtation it could possibly “cheat” by increasing interpolation more than it normally would.

            • Meadows
            • 4 years ago

            I haven’t seen nearest-neighbour in games in 15 years.

            • Den
            • 4 years ago

            What if a card was doing it secretly when under demand? Perhaps only near the edge of the screen. Could people reasonably just not be noticing that happening at 4K+?

      • cmrcmk
      • 4 years ago

      I’m not sure about lossy compression merely for high resolution, but at least at high DPI (which usually coincides with high res), you could afford some lossy-ness.

      As for deeper compression, you can’t go too far or the time required to decompress becomes more severe than just pulling from system RAM. All this would need to be done in hardware for peak performance, too, so it couldn’t be patched on through drivers and still be effective.

        • klagermkii
        • 4 years ago

        I don’t think your decompression time is going to scale that much, for most compression algorithms the increased time cost of heavier compression is paid at compress time. Anyway I think they’ve got pretty decent texture decompression hardware after their experience stretching back to the S3TC days.

        The other thing they have to work with is that most textures are going to be loaded upfront, so the driver needing to perform an intermediate compression step should mostly hit level load times and the occasional stream rather than individual frame times. There’s a lot more flexibility available to do such things when you’re not bound to work within the 16ms frame budget.

        Anyway it’s just a theory of what could be possible for hardware vendors to work around the demands of texture heavy games. Neither AMD or NVIDIA have ever been shy about aggressively making changes to the way things work behind the scenes to get FPS on their cards (c.f. shader code replacement via drivers). I could imagine texture compression overrides could be one more tool in their arsenal.

      • MathMan
      • 4 years ago

      The thing about texture compression is that it’s fixed ratio: in JPEG, you select a quality ratio and the amount of bit that rolls out is variable. With texture compression, you select a compression ratio that results in a fix amount of bits, and the quality will differ based on the content.

      So if you want to reduce the size of the compressed textures, you’d have to choose a significantly larger compression ratio. This is likely to be noticed.

      • Pancake
      • 4 years ago

      That’s what mip-mapping is for.

    • southrncomfortjm
    • 4 years ago

    Finally, some evidence to throw at those people that say, anecdotally, that 4GB is no longer enough. Thanks Damage!

    Any chance you could run a similar test with 2gb of memory? Really only has to be done at 1080p and 1440p.

      • anotherengineer
      • 4 years ago

      Ya and that’s a worse scenario type of game. The witcher 3 is fairly new and look what it uses (at bottom of page)

      [url<]https://www.techpowerup.com/reviews/Performance_Analysis/The_Witcher_3/3.html[/url<] For 1080p/budget users 2GB should suffice just fine.

        • southrncomfortjm
        • 4 years ago

        I figure, but more data is better.

        I run a 2gb GTX 760 and don’t have any issues at 1080p, then again the hardest games I throw at it are Crysis 2 and Witcher 2… that will change when Battlefront 3 rolls out.

    • wizardz
    • 4 years ago

    Thanks Scott. Articles like this are the reason i’m a subscriber.

    keep up the good work!

    • MathMan
    • 4 years ago

    Scott, I don’t think there’s any feasible way for compression to save capacity. It only improves bandwidth.

    Reason: the compressed data still needs to support read/modify/write operations and random accesses.

    That can not be done with size compression. (Or at least not while maintaining O(1) access times.)

      • eofpi
      • 4 years ago

      So they’re compressed on the GPU, sent to memory, and then decompressed in memory? And then recompressed in memory, sent back to the GPU, and decompressed there? That’s some awfully smart DRAM.

      Edit: Oh, you’re talking about allocation space, not the amount of data actually being moved around.

      • Ryszard
      • 4 years ago

      GPUs don’t access the raw surfaces in DRAM. They read the data in via the texture hardware, operate on it in the core in registers, then it goes back out to memory. So there’s no notion of O(1) access to the data in DRAM.

      So that data can and is (very) compressible in size. You have a compressor on the far side of the GPU core between it and DRAM somewhere, and a decompressor on the way in in roughly the same place.

        • MathMan
        • 4 years ago

        It doesn’t matter whether or not there is a bunch of hardware in between a shader and the DRAM to determine the access time to memory. My point is simply this: whether your surface is 128×128 or 4096×4096, the random access time to a pixel of that surface will be the same. That’s the O(1) notation. (Same is true if it were stored in the cache.)

        And that has a major implication: you can NOT use a stream compressor like zip or lz that works on the full surface, because for a full buffer compression you need to do a full decompression before you can access your data: the access time becomes O(n), where n in the number of pixels in the surface.

        None of that means that the data isn’t compressible: it obviously is highly compressible, but you have to do something to keep your constant access time.

        In the case of textures, you solve that by enforcing a fixed compression ratio (say: 4:1) on 4×4 pixel blocks, but this has the consequence that it has to be lossy.

        But texture compression isn’t at issue here: that’s pre-baked, read-only. The memory compression that is being talked about for Tonga and Maxwell is about compression of your render target, which requires on-the-fly compression for writes and decompression for reads.

        Here you don’t have the luxury of a fixed ratio compression on the whole image, because that’d be lossy. So what you have to do is divide the image in thousands of 8×8 pixels, and compress those individually (or not, if it’s not possible.) And keep track of that square is compressed or not. You still need to allocate all the memory that’d be needed for the full uncompressed image, but, square by square, you don’t need to access all its data if it is compressed.

        In other words: you save on BW, but you can’t magically squeeze a bigger buffer in the same amount of DRAM.

        (Note: this may be a bit different for deferred rendered like your’s at Imagination, in that you can deal with much larger square than 8×8, but that still doesn’t materially change the fundamentals.)

          • Ryszard
          • 4 years ago

          You still don’t access memory like that in a GPU, in-place. Access to RT or texture surfaces, read or write, serviced through the texture hardware or not, IMR or tile-based or not, doesn’t happen directly in DRAM.

          It happens in the core. So if you put a compressor/decompressor at the GPU boundary in front of DRAM, servicing all GPU memory requests, it not only saves bandwidth, but space in DRAM too.

          The GPU never operates on compressed data with a design like that, the compression and decompression is completely transparent and driver controlled.

          You’re right about how it’d work (block-based, lossless).

            • MathMan
            • 4 years ago

            I really would like to know how you can do lossless block based compression while still saving space in DRAM.

            They are fundamentally in direct opposition, irrespective of whether it’s a GPU or not.

            • Ryszard
            • 4 years ago

            So you can’t even imagine using something like RLE to represent a DRAM storage format for a pixel block? Why are they fundamentally in direct opposition?

            Apologies ahead of time if I’m missing something here!

            • MathMan
            • 4 years ago

            You can use any compression algo for your pixel block. Let’s say RLE.

            But you cannot store those RLE pixel blocks tightly packed next to each other in memory. Because when you change any pixel in that pixel block during a ROP operation, the compression may be less, and then there won’t be any room to store it.

            So each pixel block has to start at a fixed location, and that location is determined by the worst case: no compression possible at all.

            The thing that is in opposition is lossless compression and guaranteed smaller size.

            • MathMan
            • 4 years ago

            To state it yet another way: you ask the driver to allocate a render target of 1920×1080 pixels. You may render random gibberish that’s not compressible at all, or a single rectangle that covers the whole screen with 1 color.
            You have an immediate mode rendered, so you don’t know the order in which pixels will be written to DRAM.

            How much physical memory are you going to reserve?

            Answer: 1920x1080x4 bytes.

            • Ryszard
            • 4 years ago

            Sorry, but that’s just not how it has to work. I wish I could tell you why explicitly, but to do so would explain how our compressor works.

            You absolutely do know the order in which the pixels will be written at the back-end because the rasteriser has an explicit order.

            • MathMan
            • 4 years ago

            That’s why I wrote ‘immediate mode renderer’ a couple of times.

            This article is about Nvidia and AMD. They are immediate mode renderers.

            • Ryszard
            • 4 years ago

            The compressed block size might not be guaranteed, but you don’t need to commit contiguous worst-case block allocation either for the whole surface.

            • MathMan
            • 4 years ago

            In an immediate mode renderer? Where the pixel block is only 8×8 pixels (if we are to believe the Maxwell slides) ?

      • Ninjitsu
      • 4 years ago

      [quote<] By far the most consequential innovation in Tonga is a new form of compression for frame buffer color data. GPUs have long used various forms of compression in order to store color information more efficiently, but evidently, the method Tonga uses for frame buffer data is something novel. AMD says the compression is lossless, so it should have no impact on image quality, and "delta-based." Tonga's graphics core knows how to read and write data in this compressed format, and the compression happens transparently, without any special support from applications. [/quote<] That's what Scott wrote in the R9 285 review. Seems like only the frame buffer is compressed, which would be tiny anyway, compared to other things like textures. And the talk there is only about bandwidth. So it certainly seems like that's not the cause of the difference observed here.

        • MathMan
        • 4 years ago

        Frame buffer and other render targets. It doesn’t have to be that much of a size, it’s sufficient that it tilts the total amount of memory need over the boundary of what’s available on tbe GPU card.

          • Ninjitsu
          • 4 years ago

          I see. Thanks a lot for all the posts btw! 🙂

    • Ninjitsu
    • 4 years ago

    At last! Thanks a lot. I guess 4GB is plenty for 1080p, even with the 970’s partitioned config.

    I’m curious about 2GB and 4GB at 1080p, for something like a 960. Both Total War Rome II and Arma 3 can go over 1GB at 1080p quite easily – I just don’t know how far.

    • chuckula
    • 4 years ago

    There’s an interesting corollary too: Memory bandwidth isn’t a super-critical factor either.

      • Ninjitsu
      • 4 years ago

      Actually, I don’t know – quite possible that the Fury X is bottlenecked elsewhere, not allowing bandwidth differences to come to light.

      • nanoflower
      • 4 years ago

      A question does arise of whether memory bandwidth can help alleviate memory issues when you run out of VRAM. I expect we all thought the Fury X would help answer that question but it seems other design contraints may be limiting the Fury (X). Next years redesign from AMD and the Pascal entry from Nvidia should help answer that question as both should get much more performance with HBM2 over the current HBM1 designs.

        • Nevermind
        • 4 years ago

        Why would bandwidth help if you “run out” of vram? I don’t get… that..

          • Meadows
          • 4 years ago

          It wouldn’t. High bandwidth helps whenever you have to “touch more of the RAM” in a set amount of time. Foggy effects and transparent overlays take more bandwidth, higher resolutions take more bandwidth for flipping a larger frame buffer, and so on.

          Actually, if you run out of VRAM, then an overclocked PCI-E bus and fast system memory would be the things that might help some.

            • Nevermind
            • 4 years ago

            Fast system memory and memory bus affects EVERYTHING.

    • Bryan H
    • 4 years ago

    Good read.

    HardOCP keeps bringing up Dying Light when talking memory constraints.

    I’ve never played the game, but they think 4GB is causing issues even at 1440p, if I recall.

    I’d be curious to know what is really going on there.

    • Anovoca
    • 4 years ago

    A good read, thanks Scott.

    I think this is something to put in the back pocket and revisit once DX12 games are in full swing.

    • klagermkii
    • 4 years ago
    • klagermkii
    • 4 years ago

    Nice busting the myth of the benefit of just adding more GPU RAM. It’s one of the most trumpeted specs in the marketing of a GPU and on its box, but until you slip below a hard limit it’s not going to make much difference.

    Now we just need to see the influence of the number of PCIe lanes/speed on the performance of GPUs.

      • Anovoca
      • 4 years ago

      Easy enough to test on your own. Most mobos (enthusiast class not included) only run full lanes to one pci slot. just move your GPU 1 or 2 slots down and test for a drop in performance.

        • klagermkii
        • 4 years ago

        Not going to work on my motherboard. I have PCIe 2.0 x16 in one slot, and then PCIe 2.0 x1 in two other slots. Despite being unsure of the performance differences of 4x/8x/PCIe 3.0, I’m pretty sure PCIe 2.0 x1 is going to be terrible.

    • chuckula
    • 4 years ago

    Here’s my very unscientific and entirely too linear line of reasoning:
    1. I got a GTX-770 with 2 GB of RAM on a 1920×1200 display. VRAM doens’t appear to be a limiting factor in anything I’m doing at this resolution.

    2. 4K is (almost) 4X the resolution of my current display. So, 2 GB of RAM * 4 = 8 GB as an amount of VRAM that ought to be fine.

    Yes, I’m well aware that memory scaling is not a simple function of pixel count, so don’t take this as the be-all end-all metric, but it seems that Damage’s testing shows that it’s also not a horrible rule-o-thumb either.

    P.S. –> Anybody with more graphics knowledge than me about how VRAM requirements scale at either less-than-linear or even greater-than-linear rates with respect to resolution can feel free to add more.

      • Meadows
      • 4 years ago

      It scales less-than-linearly because while the frame buffer and screen post-processing will indeed take significantly more memory, the game’s underlying assets (textures, mostly) remain unaffected by what resolution you view them at.

      • Nevermind
      • 4 years ago

      Resolution is just one aspect of it..

      • MEATLOAF2
      • 4 years ago

      I like to think of it as an “overhead”.

      You will need a static (not counting things like loading/streaming new areas and assets) amount of VRAM at 0x0 resolution, just for the assets etc., that would be the “overhead”, you will then need a growing amount of VRAM depending on your resolution and AA(or other resolution dependent things that I may not be aware of).

      I wouldn’t know if it’s linear or not once you remove the “overhead” though, maybe it is if you don’t use AA.

    • f0d
    • 4 years ago

    and just as i thought
    anything over 4gb of memory usage has a useless (under 30, but for me personally i hate under 60) framerate

    that extra 4gb of memory of 390’s is useless

    well done scott

Pin It on Pinterest

Share This