|
Rinkles posted:Any indication whether it's artificially being restricted to Ada? Nvidia is claiming it depends on: quote:New Optical Flow Accelerator That and shader execution reordering feel like things that might help RT performance a lot in cards lower in the stack but we won’t find that out for a while. Mobile safari doesn’t want to pull the block diagrams out for whatever reason but DLSS2 vs DLSS3 are on this page: https://www.nvidia.com/en-us/geforce/news/dlss3-ai-powered-neural-graphics-innovations/ Note that they’re implying this is a new “accelerator”, it may “just” be the same thing already in previous RTX GPUs. hobbesmaster fucked around with this message at 19:09 on Sep 20, 2022 |
#
?
Sep 20, 2022 19:06
|
|
|
|
| # ? May 30, 2024 01:07 |
|
|
I wonder how Ada Lovelace and Grace Hopper would feel about their names being turned into GPU marketing. At least Kepler and Pascal have been dead 400 years
|
|
#
?
Sep 20, 2022 19:09
|
|
|
3090 fe is a perfectly fine purchase, and unless you need av1 encoding and potential hypothetical future raytracing performance in games that currently don't exist, just hold on to it unless you're trying to game the used market and want to cover a certain % of a future purchase by selling now
|
|
#
?
Sep 20, 2022 19:11
|
|
|
Dr. Video Games 0031 posted:Hey now, AMD's graphics dept is good now. If AMD improves their ray tracing and prices their cards well, then I think they have a serious opportunity to expand their market share. Yeah, right now id buy a 6800xt over a 3070 or 3080 and chances are Ill buy a 7800xt over a 4070 or 4080
|
|
#
?
Sep 20, 2022 19:12
|
|
|
FuturePastNow posted:I wonder how Ada Lovelace and Grace Hopper would feel about their names being turned into GPU marketing. At least Kepler and Pascal have been dead 400 years
|
|
#
?
Sep 20, 2022 19:13
|
|
|
FuturePastNow posted:I wonder how Ada Lovelace and Grace Hopper would feel about their names being turned into GPU marketing. At least Kepler and Pascal have been dead 400 years i wonder if they had to pay off grace hoppers estate to use her name and image like that, with her dying relatively recently
|
|
#
?
Sep 20, 2022 19:16
|
|
|
I don't give it much time before someone hacks in DLSS3 support for Turing and Ampere GPUs somehow if it turns out that Lovelace really doesn't have any special hardware required for it.
|
|
#
?
Sep 20, 2022 19:19
|
|
|
it wouldn't surprise me even if lovelace does have a better optical flow engine, that doesn't stop them offering a compromised version for the older hardware
|
|
#
?
Sep 20, 2022 19:22
|
|
|
repiv posted:it wouldn't surprise me But then how do you sell the new stuff? I also could see a pretty complicated balance of how to use this hardware for upscaling vs frame interpolation, and the associated compute costs of each. Did they explicitly say that the DLSS frame interpolation is also rendering at a lower render res?
|
|
#
?
Sep 20, 2022 19:25
|
|
|
Wonder how it ends up feeling. Will something be off about playing with extra “fake” frames.
|
|
#
?
Sep 20, 2022 19:25
|
|
|
Apparently there is a separate architecture NDA sometime over the next few weeks that might help answer some of these questions, but I was really surprised to see the big increase in tensor cores in the new arch, as I thought the consensus was that they really aren't used in the gaming side. Wonder if they are vestigial from AD102's double duty as a pro market AI accelerator, or if maybe Nvidia has found a better use for the hardware? Or maybe they need the spacing for thermal density anyway? Seems like a lot of diespace to not be used on a relatively expensive per-wafer process. (Just spitballing here)
|
|
#
?
Sep 20, 2022 19:27
|
|
|
Twerk from Home posted:But then how do you sell the new stuff? 
|
|
#
?
Sep 20, 2022 19:28
|
|
|
Rinkles posted:Wonder how it ends up feeling. Will something be off about playing with extra “fake” frames.
|
|
#
?
Sep 20, 2022 19:28
|
|
|
Twerk from Home posted:Did they explicitly say that the DLSS frame interpolation is also rendering at a lower render res? yeah that's where the 1:8 shading rate number came from 1:4 from resolution scaling, then only rendering every other frame repiv fucked around with this message at 19:32 on Sep 20, 2022 |
|
#
?
Sep 20, 2022 19:28
|
|
|
Rinkles posted:Wonder how it ends up feeling. Will something be off about playing with extra “fake” frames. at low framerates like 30->60 it probably won't feel great nvidia seems to be focusing on 60->120 as the baseline
|
|
#
?
Sep 20, 2022 19:29
|
|
|
Dr. Video Games 0031 posted:
I think amds big issue is competing against itself for fab space. If you are selling all the Zen cores you can churn out there is no reason make a bunch of big rear end gpus that probably make less profit per mm2 of a disk.
|
|
#
?
Sep 20, 2022 19:29
|
|
|
Assholes are gonna scalp the 4090 for like 2500$ so 1599$ retail is a relative bargain right?
|
|
#
?
Sep 20, 2022 19:31
|
|
|
Intel save us, pls.
|
|
#
?
Sep 20, 2022 19:31
|
|
|
infraboy posted:Assholes are gonna scalp the 4090 for like 2500$ so 1599$ retail is a relative bargain right? It really depends on how much supply there will be. scalping requires demand to be higher than supply to work, and I don't think scalpers are capable of having a long-term impact on the market without crypto mining ballooning demand to unprecedented levels. scalpers might scalp the initial demand wave, but after that things will be normal, and I suspect prices will remain fluid and will drop over time.
|
|
#
?
Sep 20, 2022 19:35
|
|
|
Dr. Video Games 0031 posted:
Their engineering is good! Market strategy and manufacturing availability? Not a good recent track record. lamentable dustman posted:I think amds big issue is competing against itself for fab space. If you are selling all the Zen cores you can churn out there is no reason make a bunch of big rear end gpus that probably make less profit per mm2 of a disk. And consoles, which is likely very very low margin but absolute guaranteed pipeline and growing margin over time.
|
|
#
?
Sep 20, 2022 19:36
|
|
|
teagone posted:Intel save us, pls. I returned my A380 because it was terribly unstable and sent my normally stable new system into paroxysms of hard freezes with stuck audio, but I do hope they get things more stable by early 2023. When it worked, I was honestly quite happy with the A380, and its QuickSync support was no joke. They have a shot if they can sort out the drivers.
|
|
#
?
Sep 20, 2022 19:37
|
|
|
Lockback posted:Their engineering is good! Market strategy and manufacturing availability? Not a good recent track record. At least for RDNA3, they will no longer be competing against their consoles since they'll be on a new node. I'm also not sure if that's AMD who places those orders or Microsoft and Sony.
|
|
#
?
Sep 20, 2022 19:38
|
|
|
Since DLSS 3 is creating the newly displayed frames in the GPU and bypassing the CPU, won't that mean that you get the visual benefit of the new frames but the input latency of the non-interpolated frames?
|
|
#
?
Sep 20, 2022 19:39
|
|
|
pyrotek posted:Since DLSS 3 is creating the newly displayed frames in the GPU and bypassing the CPU, won't that mean that you get the visual benefit of the new frames but the input latency of the non-interpolated frames? correct, when interpolating 60fps up to 120fps you'll get the input latency "feel" of 60fps but the visual smoothness of 120fps (with caveats for interpolation artifacts) they seem to be pairing DLSS3 with reflex as standard to try to bring down the input latency through other means
|
|
#
?
Sep 20, 2022 19:41
|
|
|
If AI has gotten good enough at making poo poo up there's no reason to stop at every-other interpolation, just let the GPU subdivide frames as quick as the refresh rate supports.
|
|
#
?
Sep 20, 2022 19:41
|
|
|
at a certain point it seems like the extra frames without the improved input latency won't really have much meaningful benefit idk
|
|
#
?
Sep 20, 2022 19:44
|
|
|
MarcusSA posted:Yeah it’s kinda AMDs game to lose at this point. 
|
|
#
?
Sep 20, 2022 19:46
|
|
|
i suppose interpolation is the wrong word for this, since it implies looking at two frames and then generating the frame between them (the laggy way that TVs do frame insertion) this is more like frame extrapolation, generating a frame based on the previous frame before the next real frame has even arrived
|
|
#
?
Sep 20, 2022 19:50
|
|
|
lih posted:at a certain point it seems like the extra frames without the improved input latency won't really have much meaningful benefit idk As always, its probably going to depend on the game. I can absolutely think of lots of CPU limited games that I would love to double the framerate on, even if those frames didn't have any extra inputs or sim work going on. Flight Sim was their first example for good reason.
|
|
#
?
Sep 20, 2022 19:55
|
|
|
FuturePastNow posted:I wonder how Ada Lovelace and Grace Hopper would feel about their names being turned into GPU marketing. At least Kepler and Pascal have been dead 400 years The first time I heard the Lovelace term I had to double check because wasn't all that aware of Ada Lovelace and the only famous Lovelace I could think of was Linda.
|
|
#
?
Sep 20, 2022 19:58
|
|
|
I didn't see this chart before (didn't watch the presentation, only read the thread/news sites) This is a really bad and confusing selection of games to use since Nvidia cards tend to be all over the place in Ubisoft titles and RE Village, but the RTX 4080 12GB I guess is generally worse than the 3090 Ti without DLSS 3. Using their own numbers for RE: Village, it would seem as though the 4080 12GB is actually quite close to the 3080 12GB in rasterization performance. You're paying $200 more for better ray tracing support and DLSS 3 support. That blows lmao. edit: to be extra clear, you should take nothing on this chart at face value. It is comparing DLSS 2.0 to 3.0 in games that support both. I used RE:V because it's not one of the games with advertised DLSS 3 support (neither are AC:V or Div2—all the others are tested with DLSS3 for the 40-series and 2 for the 3090 Ti) Dr. Video Games 0031 fucked around with this message at 03:38 on Sep 21, 2022 |
|
#
?
Sep 20, 2022 19:59
|
|
|
repiv posted:i suppose interpolation is the wrong word for this, since it implies looking at two frames and then generating the frame between them (the laggy way that TVs do frame insertion) That makes sense, at least it won't add any lag to the equation. It just worries me when I see things like the Cyberpunk comparison on the DLSS 3 page that shows 22 FPS with DLSS off vs. 85 FPS with DLSS 3 on. I imagine the input latency will probably be more like it is at 1/2 of the DLSS 3 framerate to take into account the old-style DLSS generated frames, so it will "feel" more like 42 FPS. We'll find out soon enough, I suppose.
|
|
#
?
Sep 20, 2022 20:01
|
|
|
repiv posted:i suppose interpolation is the wrong word for this, since it implies looking at two frames and then generating the frame between them (the laggy way that TVs do frame insertion) No, it's interpolation. It takes two frames and generates an in-between one. https://www.nvidia.com/en-us/geforce/news/dlss3-ai-powered-neural-graphics-innovations/ "Ada’s Optical Flow Accelerator analyzes two sequential in-game frames and calculates an optical flow field. The optical flow field captures the direction and speed at which pixels are moving from frame 1 to frame 2. "
|
|
#
?
Sep 20, 2022 20:02
|
|
|
I think using DLSS Frame Generation will probably increase input lag by a lot.
|
|
#
?
Sep 20, 2022 20:05
|
|
|
I missed that Cyberpunk is getting another ray tracing upgrade
Cyberpunk 2077’s neon-illuminated environments are key to its aesthetic, and with the new Ray Tracing: Overdrive Mode their level of detail is taken to the next level: -NVIDIA RTX Direct Illumination (RTXDI) gives each neon sign, street lamp, car headlight, LED billboard and TV accurate ray-traced lighting and shadows, bathing objects, walls, passing cars and pedestrians in accurate colored lighting -Ray-traced indirect lighting and reflections now bounce multiple times, compared to the previous solution’s single bounce. The result is even more accurate, realistic and immersive global illumination, reflections, and self-reflections -Ray-traced reflections are now rendered at full resolution, further improving their quality -Improved, more physically-based lighting removes the need for any other occlusion techniques (I imagine this’ll be out of range for my 3060ti)
|
|
#
?
Sep 20, 2022 20:06
|
|
|
Dr. Video Games 0031 posted:"Ada’s Optical Flow Accelerator analyzes two sequential in-game frames and calculates an optical flow field. The optical flow field captures the direction and speed at which pixels are moving from frame 1 to frame 2. " that could mean they're using the optical flow vectors between frames A and B to extrapolate from frame B that way the optical flow would be a frame behind, but they wouldn't have to wait for frame C to arrive if they are actually buffering a frame ahead then it's going to suck repiv fucked around with this message at 20:11 on Sep 20, 2022 |
|
#
?
Sep 20, 2022 20:06
|
|
|
Rinkles posted:I missed that Cyberpunk is getting another ray tracing upgrade I saw that. This is like the whole shebang, right? Every light source is now traced. You have multi-bounce GI. Full-res reflections. They already had RTAO (though it was broken sometimes) and RT shadows from the sun and artificial lights. Is there anything missing here?
|
|
#
?
Sep 20, 2022 20:11
|
|
|
Dr. Video Games 0031 posted:That blows lmao. Lookin at the RE Village chart, its really hard to guess if thats like 3090 non-Ti performance or like 3080 performance. Assuming its 3080 Ti level, that 4080 12gig would be the same as current 3080 Ti street prices for the same performance, with 65w lower TDP and better RT/DLSS3. Thats not ... awful, but also not great. Anyone who was around for the old insane generational performance days will get A Sad. And clearly anyone with value on the mind should stay away and buy firesale Ampere/RDNA2 or wait. This sounds more and more like Turing, honestly.
|
|
#
?
Sep 20, 2022 20:12
|
|
|
Rinkles posted:I missed that Cyberpunk is getting another ray tracing upgrade it may actually be more efficient, RTXDI is supposed to allow sampling a arbitrary number of lights using a fixed ray budget per pixel, rather than tracing shadow rays to every light on every frame. in the research they've pushed it ridiculously far, like sampling hundreds of lights using a single ray per pixel. it has its limits but it's pretty impressive nonetheless. repiv fucked around with this message at 20:28 on Sep 20, 2022 |
|
#
?
Sep 20, 2022 20:15
|
|
|
|
| # ? May 30, 2024 01:07 |
|
|
Dr. Video Games 0031 posted:It really depends on how much supply there will be. scalping requires demand to be higher than supply to work, and I don't think scalpers are capable of having a long-term impact on the market without crypto mining ballooning demand to unprecedented levels. scalpers might scalp the initial demand wave, but after that things will be normal, and I suspect prices will remain fluid and will drop over time. The X factor in 4090 supply is the persistent rumor from multiple sources that Nvidia ordered WAY too much allocation on the 4nm+ node, and supposedly TSMC is forcing them to stick to that schedule albeit delayed by a quarter or so. I actually don't expect nearly the impacted initial stock that we had with Ampere, especially since there are also numerous indications that Nvidia postponed this launch and is therefore likely to be sitting on lots of stock. In short I think odds are extremely good that launch stock will be high, AND we can likely expect some discounts in pricing into Q1, especially post holiday season. Of course they might pull some Apple poo poo and artificially limit initial units to generate the image that the cards are selling like hotcakes, but they can only do that for so long. Q1 2023 is when everything gets revealed but in short it just, to me, doesn't look like we'll have anywhere near the issues we had with Ampere. This is all based on speculation of course, but numerous sources are conflating the same general situation.
|
|
#
?
Sep 20, 2022 20:16
|
|
