NVIDIA Announces GeForce RTX 3070, 3080 and 3090 (With Even More Shader processors)

Discussion in 'Frontpage news' started by Hilbert Hagedoorn, Sep 1, 2020.

  1. Humanoid_1

    Humanoid_1 Master Guru

    Messages:
    960
    Likes Received:
    66
    GPU:
    MSI RTX 2080 X Trio
    NVidia are not being quite honest again. Their shader counts are Half what they are claiming. Their Ampere shaders do 2 calculations per clock and so are claiming they have 2x the number of cores. This is why performance is not near their claimed TFLOP numbers.........

    watch from about 17.11 (or 18.07 for the exact part with no background info)



    Really looking forward to HH's review and am more interested to see what AMD bring to the table. Ampere's performance is certainly still a really nice step in the right performance direction from what I have seen, if not quite what was hinted at by NVidia.
     
    Last edited: Sep 6, 2020
    Loobyluggs likes this.
  2. Supertribble

    Supertribble Master Guru

    Messages:
    938
    Likes Received:
    157
    GPU:
    Noctua 3070/3080 FE
    Mr. Adored looked at other aspects of the architecture to see what was going on but ultimately he sounded, perhaps begrudgingly so, impressed with Ampere.
     
  3. Fediuld

    Fediuld Master Guru

    Messages:
    723
    Likes Received:
    418
    GPU:
    AMD 5700XT AE
    I have the feeling Ampere will fall over on RT without DLSS and normal DX11/Vulkan games.
    Tensor cores are twice as fast and possibly the RT cores also, however normal shader performance might not have moved much if at all from the 2000 series. And we are talking about a >50% power consumption increase here.

    Also Nvidia changed the measurement types AGAIN. From RT Ops to RTFlops etc. Possibly because quickly people would be able to calculate from the RDNA2 diagram on Xbox the performance against upcoming AMD GPUs.

    [​IMG]

    And we see here that there is no indication where 2000 series was standing in comparable performance metrics.
    https://wccftech.com/nvidia-geforce...ster-than-rtx-2080-in-opencl-cuda-benchmarks/

    I feel is all smoke and mirrors for the pre-order crowd.
     
  4. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    8,073
    Likes Received:
    907
    GPU:
    Inno3D RTX 3090
    It has already gone better than native, there are videos about this. It really depends on the title and the amount of training for now, but it's as possible as Google lens on a shitty mobile processor.
    https://www.eurogamer.net/articles/digitalfoundry-2020-image-reconstruction-death-stranding-face-off

    This is information that literally only NVIDIA has (and perhaps a few developers). You have no idea what you're talking about, as the only data we have is the latest version matching native, and sometimes exceeding it.

    They haven't promised any "final" DLSS. DLSS 3.0 was something Moore's Law is Dead talked about to pay rent.

    Again, I don't know how DLSS works internally, and neither do you. The above looks like a comment from someone who doesn't necessarily understand how deep learning networks work. I don't know either, but I know enough that it's not like this, it really depends on the model, and we don't know how they do what exactly, how generalised DLSS is etc.

    The comment below is probably more helpful than either of ours.

     
    yasamoka and nizzen like this.

  5. DesGaizu

    DesGaizu Ancient Guru

    Messages:
    3,712
    Likes Received:
    74
    GPU:
    AORUS 3060TI
    So DLSS can only been used on small linear games that are "static" (what ever the frack that means) and takes ages to "learn"

    well I guess Kojima is a wizard then because death stranding is a huge open world that's world constantly changes with dynamic buildings and the game released with DLSS 2.0
     
  6. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,809
    Likes Received:
    3,370
    GPU:
    6900XT+AW@240Hz
    1st) Game has ass quality DLSS
    2nd) learn to recognize apple from leaf of grass. Then learn to recognize what details leaf of grass should have and what apple should have.
    3rd) Throw a lot of grass and apples on scene => It Just Works
    4th) Throw in some monkeys running around and enjoy deep dream
     
  7. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    8,073
    Likes Received:
    907
    GPU:
    Inno3D RTX 3090
    You're just reinforcing my point that were all talking out of our a*s about something we have only the word of Nvidia and the results.

    We only know it's a Deep Learning Network.

    Is it an RNN, is it an ANN/CNN nobody knows. On top of that, even if we did, we could only guess how they train them or tweak for biases, or anything else. For my money it's probably a Convolutional Neural Network, but I am no expert. Even experts probably couldn't figure out what it can or cannot do unless they have the actual training data from Nvidia. I bet that not even Nvidia engineers are 1000% certain on the limits.

    Yet here we are, full of certainty, in a tech forum :p
     
  8. fantaskarsef

    fantaskarsef Ancient Guru

    Messages:
    13,485
    Likes Received:
    6,322
    GPU:
    2080Ti @h2o
    On your first point I totally agree. I too had lots of hope for DLSS personally, but the adaption was in way too few games. That ray tracing wasn't quite there yet, I was aware of it, but as I mentioned I didn't buy it for RTX.
    Second point, the price was indeed steep. I wrestled with myself for months and eventually took a good offer I got here on the forums. And I didn't buy the card because I needed to do that, or thought of it anything else than it was a good traditional card for, admittedly, an extra price tag which wasn't exactly delivered. But that was my own choice, so I don't complain about it. I decided to buy the card out of traditional benchmarks and for that, the card served me well.
    I too believe that Ampere destroys Turing in RT and probably many other things too, especially on higher resolution and RT combined. But at the same time I can't be upset that a later generation does batter than my last gen card. Because that's the way it's supposed to be, remember, we want improvements. Complaining would feel to me as if I would only complain out of buyer's remorse, which I don't have.
    After all, Nvidia is a business, and business wise they absolutely played their cards to sell as many Turing cards as they could. Such things are part of the game. The big surprise is that they actually offer a remarkable improvement, probably because AMD does so too. If there wasn't at least the hint of competition, we might not even have 7nm/8nm lithography for GPUs by now.
    If Nvidia treated their customers poorly... maybe yes. Their marketing department seems to have done an incredible job, I too was hyped far more than the adoption of new tech held up with the hopes I had. But overhyping happens at the buyer's side, and there's always people doing that with each release. If we should be compensated? Maybe, like with a really good upgrade for a relatively lower price than what we had. But Nvidia only made half a step into that direction... being Nvidia and being ahead, such things happen, sadly. But there won't be any chance for serious consequences... Ampere will just sell, like Pascal probably.
     
  9. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    8,073
    Likes Received:
    907
    GPU:
    Inno3D RTX 3090
    I think that Nvidia has done a stellar job getting killer prices from Samsung, who struggles to get customers, and Ampere will prove they have the way to mass produce fairly big dies for a demanding customer, and put them a bit on the map.
    Their 5nm looks OK actually, but I wonder where Nvidia will go after they (inevitably) burn this bridge too. Maybe they buy ARM and GloFo and start making full computers :p
     
  10. NewTRUMP Order

    NewTRUMP Order Master Guru

    Messages:
    618
    Likes Received:
    244
    GPU:
    rtx 3080
    IF you watch the leather jacket guys video he has to use two hands to lift it up. Should rename it Godzilla 2020. [​IMG]
     
    Fediuld likes this.

  11. NewTRUMP Order

    NewTRUMP Order Master Guru

    Messages:
    618
    Likes Received:
    244
    GPU:
    rtx 3080

    2080 isn't exactly hard to beat, and it's not a flagship turing card. My point was if they are so great why not show off it's strengths in the release video.
     
    Fediuld likes this.
  12. Supertribble

    Supertribble Master Guru

    Messages:
    938
    Likes Received:
    157
    GPU:
    Noctua 3070/3080 FE
    The 2080 is regularly beat by exactly two cards, the 2080ti and 2080S, so no, it isn't easy to beat, as for Ampere being disappointing, the digital foundry video showing its performance shows the 3080 being anything but disappointing but blah blah, doom Eternal is well optimised/easy to run, but it's all relative, isn't it. The 3080 clearly and easily beats the outgoing 2080ti under the same workload at half the price.
     
  13. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,809
    Likes Received:
    3,370
    GPU:
    6900XT+AW@240Hz
    What's your problem again? Is it myself? Because as far as my hyperbolic description goes, you basically placed bet on exactly same base. Yet you feel urge to disagree. Nobody knows limits, but maybe you should at least try to use freaking google to read few hundred of pages which explain cost of progress and what is needed. Because that we know for current methods.
    And no, even nVidia does not know why it does what it does. They set it up, let it run and hope for the best. Technology been around for very long time, but it is no different than that crystal which "can identify" numbers from 0 to 9.
    It does it somehow, but you can't tell exactly what part of it does exactly what. And it is like simplest network. If they could tell, DLSS would be inside driver and applied everywhere already.

    And do everyone a favor. Go look on how SSAA and better techniques look like. What is supposedly ground truth on 16K resolution. And what ends up on screens of people who use DLSS.
    Thing is they are not training it for purpose of SSAA, they are training it for purpose of faking SS out of low resolution. Getting details where there were none before. And that AA is just side effect of adding this kind of detail where none was before.
    And you know why it basically succeeds only on doing AA? Because from all those scenes, environments, objects, there is common thing. You get a very good AA on Ground Truth images where there was no AA before at low resolution.

    Same reason why DeepDream has so many freaky eyes. It has been trained on animals and all of them have eyes. Most common thing became dominant function of network to achieve results that are as close as expected. nVidia engineers are by now screaming "WE do not want more eyes, give us more hair or fins where we want them." Except their Eyes equals to AA and hair/fins are details which should appear on particular objects on very particular places. (That's why it does good AA job, and bad job at adding details.)
    And with neural network, you do not know what made that one pixel you really like in that pace to be placed there, and why that other pixel is missing or why there is some undesirable extra detail in some place. Change that removes one bad thing may add more bad things or remove good things in process.
    If you tracked thousands of fractional contributors, you would still be no smarter than before. You would be like: This thing contributed 0.1% while that other thing contributed 0.11% of this particular pixel's final RGB value.

    But as I wrote before as an hint. That blind devotion is same as P=NP statement without looking at NP problems.
     
    Last edited: Sep 6, 2020
    Fediuld likes this.
  14. Supertribble

    Supertribble Master Guru

    Messages:
    938
    Likes Received:
    157
    GPU:
    Noctua 3070/3080 FE
    That's a lot of text and I've had one too many beers to process it all. My own experience with DLSS 2.0 has been generally positive. In control the benefits greatly outweighed the few negatives but that game has appalling TAA anyway that made the image look like blurry sh1t even at native 4K when viewed on a large screen UHD TV. Death stranding's TAA implementation isn't nearly as bad and in my opinion DLSS 2.0 generally doesn't look as good as a 4K image with TAA.
     
    PrMinisterGR likes this.
  15. Noisiv

    Noisiv Ancient Guru

    Messages:
    8,190
    Likes Received:
    1,458
    GPU:
    2070 Super
    @Fox2232

    Lets assume we don't have any deep knowledge about DLSS. Because we don't and because it's unimportant for what I am trying to say here.

    Take a look at the most common, most generic aliasing problem. Consider a frame buffer that is being outputted to 1920x1080 display. What we get on our monitor are different kinds of jaggies and imprecision patterns which we call aliasing.
    A tough problem to solve, computationally, resource wise yadayada - depending on the original content it can be hard to solve to our eyes satisfaction. We all know this...

    YET...and this is my point - intuitively we KNOW that all the information is there. Our brain knows that the picture on the monitor is wrong, and furthermore our brain pretty much knows what the pic should look like.
    Answer me honestly: Does this problem (aliasing) not look like a perfect candidate for a neural network algo?
     
    yasamoka, pharma and PrMinisterGR like this.

  16. Fediuld

    Fediuld Master Guru

    Messages:
    723
    Likes Received:
    418
    GPU:
    AMD 5700XT AE
    You forget few things
    a) 8nm Samsung is effectively 10nm. And Nvidia used Samsung because cannot use those Godzilla die sizes at TSMC 7nm.

    b) If Nvidia goes down to 5nm like AMD does next year, needs to move to MCM and new design from ground up. Monolithic dies are DEAD.

    c) Proof? Simple look at the spec sheets for 5nm from the various companies. They are all EUV with maximum size of the lenses that are 1/3 the 3080/3090 die size (630mm2) and that with atrocious yields.

    Personally I doubt AMD can do something bigger than 64CU at N7P also (between 360-370mm2). Biggest die currently made is the Xbox APU (361mm2) and this one needs 2CU chopped to improve yields and lower prices. 5700XT is just 250mm2 tiny compared to the 2070S at 545mm2 even by scaling the node. Radeon VII is 331mm2.

    Again doesn't need something bigger than 64CU to compete. A 64CU RDNA1 GPU could have been be 20%-30% faster than the 2080Ti (over 60% faster than the 5700XT) so RDNA2 with it's gazillion improvements who knows.
    If AMD pulls a 80CU chip it will be $1500 range as it would have bad yields on something close to 450-500mm2. (I do not thing there is any N7 chip made bigger than 361mm2 currently).

    Also in relation to yields we see the 3080 & 3090 are on the same die size. Nvidia cannot afford to make 2 distinct products due to sheer cost to make those big dies.
     
    itpro likes this.
  17. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,809
    Likes Received:
    3,370
    GPU:
    6900XT+AW@240Hz
    1st) This is not what DLSS does.
    2nd) That what you described is exactly only function I ever wanted from DLSS and it is best use for neural network meant to improve image quality of current games. Because entire fight is about finding edge and doing precise AA.
    3rd) DLSS does AA as part of upscaling. And that brings in main issue of having low resolution image as source, where most of image is covered by surfaces and little by edges/intersections of objects. DLSS can pretty much deal with edges, but screwes image quality for surfaces as it is unable to add fine detail which you would get on native resolution.
    4th) because it is trained to do 2 very different things, it is not excelling even on AA. Yet, when initial FF DLSS screenshots came, I wrote that while some places are bad, others have finest AA I seen till that day. (They should do one thing and do it properly.)

    nVidia does this upscaling for performance purpose as compensation for DX-R's low fps. Yet it ends up in games that have no DX-R or image quality good enough to perform poorly on average HW to even need DLSS performance boost.

    Would they just do DLAA (SSAA) or whatever they would call pure AA, nobody could ever complain. Because it has potential to beat any other AA in terms of IQ. And do it with damn small performance impact.
    And they could have had pure tensor-based-AA in place long time ago in driver as optional post process for each and every game.

    Nobody would have to suffer TAA. Or even worse, either DLSS or TAA as options.
    (But not to be negative F1 has pretty impressive TAA.)
     
    Dragam1337 and itpro like this.
  18. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    8,073
    Likes Received:
    907
    GPU:
    Inno3D RTX 3090
    If this is a convolutional network, which is what the Nvidia presentation implies when showing it (see link below), then it has tens of different parameters to tweak, and they all involve a ton of effort to get right. If it was a magic eight ball, everyone else would have had it by now.

    Here's a link in case anyone wants to start learning how all this works.
    (hint: It's different than the crystal game).

    It cannot be globally applied because each game apples TAA differently and DLSS requires a motion vector to work. Since DLSS 2.0 it's a single network for all applications and it is included in the driver.

    DLSS = previous final frame + motion vector + new low resolution frame + deep learning

    You can see Jensen explaining it here (the Youtube link is timestamped). He also says how it is in the driver.

    So it won't work unless all engines implemented TAA the same way, and they don't. The above is another example of not knowing what you're talking about, and being annoyed for that being pointed out.

    There is a link here to check what it does and it does not. It surely doesn't only do AA. As you yourself wrote, AA is the after effect.

    Examples (look at the face detail):
    [​IMG]
    [​IMG]

    On this last one you can see the artefacts that DLSS produces as it doesn't know what to do (yet) with the edged of things like teeth. Yet it still produces detail that isn't there in the source 720p image.

    [​IMG]

    Right. DeepDream had the freaky effects due to back propagation, and it was also made in 2015. It was created to explain how the "black box" behaved, but the images got popular fast and the meme was out.
    Here is a very interesting article that shows that it's not the issue at all, and what DeepDream was actually made to do. I read this one something like a year ago, it's still great if you have the time with a cup of coffee.
    It was a good starting point, but it says nothing about how things are now.

    This is completely correct. But that's why the "secret sauce" is in the training, which we have no idea about. We only have NVIDIA's word about it, and what it can or cannot do (and the results like the ones from above).

    This is not correct. Quantity and quality of training data is one of the most important factors for any mechanism that learns this way. As for P=NP, it's literally irrelevant in this case. Neural networks work with confidence intervals/scores, and not "absolute" results.

    DLSS will never look like the 16k image, but we don't care as long as it looks like it close enough.
    You trust the network after a time to mostly give you what you consider an acceptable solution. No network will give you a confidence interval of 1.
     
    yasamoka, itpro, pharma and 1 other person like this.
  19. devastator

    devastator Master Guru

    Messages:
    204
    Likes Received:
    47
    GPU:
    palit rtx 3070ti
    the 3000 range doesnt have doubleing of the cores i think rtx 2080 has 2944cores and 184 tmu on the 3070 it has 5888 and still 184 tmu there for it has only 2944 cores
     
  20. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    8,073
    Likes Received:
    907
    GPU:
    Inno3D RTX 3090
    Yeap, that Samsung is a "half node". I don't know why Nvidia used Samsung, the rumour mill is that Samsung asked for payment by the complete chip, and not the wafer. It is TSMC that is making GA100, which is a 815mm2 behemoth, and the "big" Ampere chip, and not Samsung.

    They would have zero problems producing the relatively tiny consumer GPUs.

    Yes, but, like why? Monolithic dies still make sense for packages that are power-sensitive, that's why AMDs APUs are monolithic. The fabric uses too much energy too much of the time. If they have a proper offer from TSMC/Samsung. why would they ever bother? If anything, the smaller process would make larger GPUs even more viable. But we will see.

    Those chopped CUs might be there for Microsoft's margins, not a "real" need. I don't think that AMD would have any issue pushing above 600mm2 if they wanted to, but this is just a personal belief. I also believe that the Xbox Series X has what is the equivalent of a midrange GPU for 2021.

    The 5700XT has 40CUs. A 64CU "5900XT" cannot be 60% faster, unless you think it will scale super-linearly or something. I honestly think they need 80 "RDNA 1" CUs as a minimum to even compete on traditional tasks with the 3080, unless they merge them or name them differently for RDNA 2.
     
    Denial, itpro and Noisiv like this.

Share This Page