I thought that was the whole idea of DLSS. To be a better and more efficient way to upscale than checkerboard and other methods.
It comes across weird though. Either use DLSS for every title, or drop all the titles that can't play in native 8k (or in this case drop the DLSS titles).
I think the paper has a different focus. It's focused on classical upscaling. DLSS is a little different. When DLSS came out my first thought was this has to use some form of GAN and/or convolutional autoencoder. Meanwhile Nvidia published it is indeed based on an AE. I don't think the exact details are published anywhere, but I could be wrong. From what I know, developers have to provide certain data to Nvidia who then run the games (or data?) on their clusters to create the AE that is then embedded into the drivers for the game. So technically it's probably more a reconstruction of a high-res original (on a Nvidia server) from a low-res sample on a consumer PC and not a "upscaling". The nice thing is the use of their tensor cores which are perfect for convolutions to make it very efficient. Last time I checked (long ago, IIRC Titan V/V100), we were talking about 1 mul-acc operation per GPU clock vs. 1 4x4 mat-mul-acc per GPU clock. Not sure about current gen, please correct me if I'm wrong.
Power/Watt is exactly the reason I'm not impressed by the performance of the new RTX series. May seem to buy NVidia's marketing hook line and sinker without a skeptical eye. It's silly, especially when often the same voices crop up ready to call out any other company's marketing B.S.
An increase of over 100 watts isn't significant? Maxwell to Pascal decreased TDP for performance, in fact, that's what used to happen all the time every new gen. Somehow with the last few gens of CPUs and GPUs we've been trained to accept "well it has more performance, so it has to have higher power draw".
Mea culpa, it's only a 40% markup over Pascal.
If you don't care about power consumption that's your prerogative.
I feel like I should clarify, I don't hate NVidia's gpus or even this gen. They are pushing the envelope in many ways, like raytracing (which I've warmed up on from a year ago) and their new "broadcast" tech. They haven't been like Intel and only improving performance a tiny bit year over year, consistently offering 30%-40% performance uplifts each gen. There's no question that NVidia makes good GPUs. (unless it's Thermi)
I wouldn't say unimpressive, the price is certainly impressive. However, performance is misleading. While there is a improvement for games, the new cards still can't run 4k max setting for every game >60fps. They still have drops to lower fps. So their claim to get 8k gaming is a little misleading. Sure it will work and it requires DLSS. But it won't work for every game with useable framerates and people might still choose to play with lower resolution / less detail, just like the 20x0-series cards. Was really looking forward for 4k60 across the board for the one or two games I play every year, since I play them on a TV.
Well. I kinda agree what you are trying to say. Don't get me wrong, but talking about buying into marketing hype in "Apple" community? Well.... we all buy into Apple's don't we?
Personally I try to keep my expectations in check. In the case of Apple however, we do see quite a bit of negative or skeptical feedback (see several threads on MacRumors even), which is good in any case for any company. Apple at the very least has about as many detractors as they do loyal fans, it sort of balances it out I'd say. We need those "devil's advocates."
That's also what I guessed. The performance was misleading, which we expect for any product release. Though the cards aren't duds, we (consumers in general) should be ready to look at it with a very skeptical eye.
Nvidia hasn’t had good competition till recently to force them the lower their prices. Some would argue that they still don’t have any competition. I mean I am kind of baffled that folks think getting 2080ti performance for half the price is still too expensive.
Yeah for onscreen rendering mGPU isn't really worth it. For compute and offline rendering though, it is great.Franken-CrossFire: Radeon RX 5600 XT Joins Radeon RX 5700
CrossFire is gone, but we still have multi-GPU.www.tomshardware.com
I am not sure there is just 1 autoencoder. Seems like they have at least 2, one for denoising and the other for upsampling.
I don't think that's what it is about. It's pretty clear that we can't use Nvidia cards in the macOS environment, that ship is sailed. The big question is, where do we go? While AMD is currently in Apple hardware, the support is very limited to the Apple software eco-system, focused on Metal and Apple's Libraries, without much for OpenCL, ROCm and the likes.
Doubtful, as that's not what an AE is for unless you combine it with a CNN or hybrid approaches with, say a SVM.
Not sure, but did Apple mention GPU support for open source frameworks? I can't imagine it's difficult to get Tensorflow running on a Apple Silicon CPU. After all, I've been running Tensorflow and PyTorch on ARM for ages.
Yes and No. I don't really want to miss the Apple eco system. Everything is well integrated, macOS, iOS and the watch. This is personal thing for me. Of course I could switch, I simply don't want to. My research workflow is optimised as well, DevonThink, Scrivener, Latex and Bookends. There are alternatives, none of which I personally think are as well thought out as these.
