They're not separate things. DLSS is an advance form of checkerboarding using deep learning.
-
Ever wanted an RSS feed of all your favorite gaming news sites? Go check out our new Gaming Headlines feed! Read more about it here.
NVIDIA publishes RTX 2080 vs GTX 1080 comparison
- Thread starter Alethiometer
- Start date
I think it's incorrect to say that. Checkerboarding is a very specific technique and DLSS works in a completely different way, though both have the goal to improve performance.
Damnit the msi trio 2080ti is 327mm long... Just too long for the fractal define r5. Hopefully i can shift my hard drive cages enough out of the way.
Ah shit, I didn't realize they were that big. I have the same case and figured it's big enough to house anything.
they are all removable, I have no spinning drives in mine nor optic drives so I took everything in the front out
Damnit the msi trio 2080ti is 327mm long... Just too long for the fractal define r5. Hopefully i can shift my hard drive cages enough out of the way.
So dlss reduces the the resolution and checkerboard up like t-aa? Works well in siege for me if that's the case
Lower internal resolution, but no checkerboard pattern or anything like that. Its method for generating the missing information is completely different.
I have a blu ray drive i never use so i could take the top cage out, but i have to use the 5 drive cage sadly. Ill have to move it and measure, it still may not have enough clearance even if i move it all the way to the top or bottom. (these cards too dang thicc)
Wow, that is a ridiculously large GPU.
Damnit the msi trio 2080ti is 327mm long... Just too long for the fractal define r5. Hopefully i can shift my hard drive cages enough out of the way.
I chose the Define R5 for my builds because it has direct airflow over the drives, and the number of drives that it can hold, so removing the cages is not an option for me.
Damnit the msi trio 2080ti is 327mm long... Just too long for the fractal define r5. Hopefully i can shift my hard drive cages enough out of the way.
Shit.. got the fractal meshy case and that is 12mm too long... Guess I have to go with another card. Really wanted the msi trio through...
I have the Define R5 too, but I have no drive cages because just SSDs. So I should be good, right?
If you remove the drive cages that image says it will support a GPU up to 440mm in length, and
EDIT: I'm not sure why they wrote it in small text (and it didn't format properly so it was tiny here), but that poster was referring specifically to an "MSI Trio" 2080 Ti.
The reference GPU is 267mm and will fit without removing the drive cages. I expect most third-party GPUs will too, but you'd have to check that they are under 310mm.
I thought anything over 300mm was a thing of the past.
Last edited:
That teardown is nuts. It's like a rabbit hole the amount of screws he ends up having to take out. I wonder if his joke about the screws accounting for the cost rather than performance meant anything. We'll see tomorrow!
Right now no, since they're all new designs for their coolers we're waiting on reviews to see jwich of the new cards works best.
However I've had good experiences with EVGA, MSI, and ASUS in the past.
I'm going with a stock NVIDIA founder's edition however, since my case is small I don't have much room to spare and I certainly cantc fit the 3 slot cards in lol.
Hadn't even checked on card length for mine yet. I ordered MSI Trio 2080, which is the same size and just checked my case supports up to 330mm card. 3mm to spare. Upgrading from a MSI GTX 980 279mm ->327mm nearly 20% longer.
Damnit the msi trio 2080ti is 327mm long... Just too long for the fractal define r5. Hopefully i can shift my hard drive cages enough out of the way.
I mean, it's going to make them more expensive to manufacture, so I'm sure it's part of it... but the real reason is the R+D and silicon for the RT cores. It's a massive chip remember. Build quality is a factor though from the sound of it.
My case is plenty roomy so I didn't even think to measure it. The two fan EVGA I've ordered (it's the 2.75 slot one) will fit easily in my case. Last case I had was a Shuttle (I was building a PC and knew I'd have to travel with it) and I got sick of having to dismantle the entire thing to take out a stick of RAM.
Intel needs to release a 6Ghz dual core CPU just so gamers can shut up about broken old Crysis.
Completely dependent on wattage from what I've seen. My 650 has 8+8, my 850 has more than that. Older PSUs I've had have only had 6, 8, etc...
I want this game to be great, and I want to enjoy these RTX effects, but this frame rate, whew boy.
Damn dont think there's a way to keep that 5 drive and have space for a long card, can you even attach the cage at the very top? Or put 3 drives on the bottom cage, rest 2 somehow in the optic drive cage...
I haven't read the full white paper, but my understanding was that the precise mehcanism by which the "new" pixels were generated was a product of machine learning and would vary from game to game. Speculation on my part, but It doesn't seem inconceivable that some kind of pseudo-checkerboarding may be occurring in certain titles, or perhaps all of them.
What I'm mainly unclear on is whether Nvidia has tightly controlled the problem space that the ML is searching in, so it would more or less be finding optimal variations on a common theme, or whether it's really broad and could be using wildly varying techniques that have little in common with each other.
Oh, it is. Without a doubt. Delay the reviews as long as possible so people can't cancel their pre-orders before the cards actually launch, or grab the impatient people that just say "fuck it" and order one before the reviews are out anyways. I don't even plan to get one until next year and I'm still annoyed :P
More of a walkthrough of RTX using the Atomic heart engine. I've seen multiple trailers for this game... and this video looks like something that's actually not in the game and is just setup as a small tech demo/showcase for Nvidia. Still cool I guess.
There is lot of work to be done.
I cannot wait to see proper information on how much time was spent on additional shading and so on.
Last edited:
Do we have any examples of DLSS working? A screenshot or two for comparisons. We should be able to tell fairly easily if any sort of checkerboarding is taking place.
Speculating here as well, but I'm guessing it just doesn't work in the same way at all. I think the AI takes the fully rasterized frame and upscales it with magic AI that will probably one day will be upscaling us for lunch, but I digress. :p Unless it has access to other parts of the rendering pipeline (does it?) it couldn't do the type of interpolation checkerboading does.
Sure. But it would be a waste to use checkerboarding on Turing when you have VRS and TSS to use instead of just plain checkeboarding.
Pretty much. AI just draws on top of a lower resolution image to make it look like a higher resolution one. Results will obviously differ.
DLSS 2x is a lot more interesting as it uses native resolution image and then DL AA it to make it look like a 64x SSAA image - considering that tensors are able to run in parallel with general FP processing this may result in somewhat of a "free supersampling".
It's interesting, because you can see the (current) performance hit for turning on RTX in Metro... but the difference in IQ is stark. I wonder how this compares in performance to VXAO, because I'd argue it's much better looking than VXAO (which itself had a large performance cost). The scene with the shutters opening and closing was pretty jaw dropping, but clearly RTX off is running *much* smoother than on. Hopefully gsync can clear that up. It's hard to estimate by eye just how far below 60 this is running with RTX on, but I don't believe it's dipping below 30.
Edit: I was right. DF threw their analysis tools at this footage and RTX on runs from around 45 to 60 fps.
Last edited:
Got an EVGA RTX 2080 XC and XC Ultra pre-ordered on Amazon. Not sure which one I want to keep. Benchmarks can't come soon enough.
Thought this was interesting. Afterburner app creater gave his thoughts on the Nvidia one click overclocking and it sounds like it's coming to an update in afterburner.
https://www.reddit.com/r/nvidia/comments/9gvqr7/msi_afterburner_app_creator_shares_his_thoughts/
https://www.reddit.com/r/nvidia/comments/9gvqr7/msi_afterburner_app_creator_shares_his_thoughts/
I believe that update just pushed, at least based on a new thread in the same sub?
Yeah. The two aren't mutually exclusive - the gigantic vapor chamber is legitimately unique and potentially very good, but that doesn't mean too much of the cost isn't going to stupid screws. That many screws have to be a significant portion of the HSF's BoM.
Steve does mention though that due to the complete lack of airflow to the PCB it's possible that the tolerances are extremely low, everything probably has to be sub-millimeter to make sure every single IC that needs cooling is contacting the heat spreader somewhere.
Yeah, it's mentioned that (similar to TAA) it uses temporal data, just instead of just anti-aliasing it's actually supposed to keep the detail from the previous frames also and use AI + motion vectors to know what detail to keep.
I think of something like a moving sword: If the sword is moving from left to right, with the motion vectors the AI will know it can use data of that sword's previously rendered output and transpose that on top of the new rendered output - essentially getting multiple samples in a single image. They key really will be in how they know what data is valid and what will need to be thrown away, so that specular highlights and such work properly even if the angles are changing slightly. Will be very interesting to see if there are any noticeable artifacts from the process.
Whichever one of these overclocking software suites has support for it by the time mine arrives is the one I go with.
THREE MORE DAYS FAM.
For you! I have yet to receive a shipping notification despite ordering 2 minutes into the presentation and using PayPal :(.
Or it just uses a previous frame to create AA for the current one in parallel to the current one being rendered.
I knew they were gonna put out a raytracing benchmark, but
#1. Why are they making a benchmark that can run at a good framerate? This shit should run 10fps on a 2080Ti to be legit Futuremark. Is UL going to make them go soft?
#2. Much less impressive than Nvidia's several tech demos. Pretty simple visually.
Sure, it's a benchmark, but I remember ~15 years ago the new 3DMark was a major event. I know it can't be that way again, but even Time Spy looked real good, I think. I hope this is the 'low spec' benchmark.
I'm banking on my Best Buy pick up in store date still holding. But we'll see.
I'm sure there will be an Extreme version, just like Fire Strike.
I wonder how they can extremify this? Just add a bunch of mirrors all around?
So, like TAA? In which case DLSS would be yet another AA that doesn't work with SLI.
Incase you haven't noticed, SLI is kind of dead. NVlink may yet show some promise, but AFR just isn't a technique that offers much any more. Maybe that'll change, but I wouldn't expect it to.
Since DLSS is only on cards with nvlink, it would not be a problem in SLI. In fact taa should work just fine now too, the lack of bandwidth is why taa tanked framerates on sli.
I think some of the confusion is the result of the training of the DNN for DLSS being separate from the current DLSS functions.
During training, the DLSS relied on 'back propagation' by dynamically weighting the iterative differences between 64x supersampled reference images and the respective raw inputs. After sufficient training however, the AI now 'knows' how to produce output that approximates the reference quality, and now functions by analyzing multiple inputs within a given set of frames, as described in the excerpt of the Turing whitepaper below:
So DLSS = Performance Mode, and DLSS 2X = High Quality Mode
And we can see here how effective the result is:
The DNN AI also has to intelligently combine inputs based on its 'understanding' of the motion of the scene, instead of strictly following motion vector values like TAA, which gives us results like this:
Something else also worth mentioning is the AI Super Rez technology, which effectively increases resolution from a lower resolution sample by 'intelligently' inferring new pixel data to preserve detail:
Though this seems to only work for video and pictures, not interactive media.
At any rate, I'm working on a thread that goes over all of the new features of Turing, as described in the whitepaper. Hopefully that will help consolidate all of the information we have on it so far.
Not the case. The 2070 cards have DLSS but no nvlink. Furthermore, Nvidia haven't really announced or gone into any details for what we can expect for nvlink beyond regular old SLI in high end gaming GPUs. Ideally you'll be able to have two GPUs appear as a single GPU to windows, but even with nvlink you're still going to hit bandwidth issues (since the bandwidth between GPU and onboard ram is much higher). Not to say it isn't a solvable problem, but I genuinely think we'll be moving away from AFR as a multi gpu solution. That's just my amateur opinion of course, but AFR's place in DX12 (which unless I'm mistaken every DXR title is going to be within) is of shrinking utility.
I wasn't expecting Nvidia to ship these out so soon. I thought my 2080 would arrive next Tuesday, but it's scheduled for delivery this Thursday.
The last time I changed a graphics card was when I replaced a TNT card for a Radeon something or other in 2001 (16 MB of RAM up to 32, baby). I hope I don't fuck this up.
The last time I changed a graphics card was when I replaced a TNT card for a Radeon something or other in 2001 (16 MB of RAM up to 32, baby). I hope I don't fuck this up.
