Nvidia DLSS 3 on RTX 4090 - Exclusive First Look - 4K 120FPS and Beyond

preview_player
Показать описание
Alongside the arrival of RTX 4000 series cards, Nvidia revealed the new DLSS 3. Frame Generation combines with the existing DLSS 2 and Reflex latency reduction technology to take 4K gaming performance to new heights. But what is DLSS 3? How does it work and what should you expect from the experience? Rich and Alex team up to explain DLSS 3, from first principles to performance to image quality and more.

00:00:00 Introduction
00:03:55 What is AI Frame Generation?
00:06:23 Performance Tests - Up To 5x Perf vs Native 4K
00:14:41 The Practical Reasons Why We Need Frame Generation
00:17:16 Limitations - Latency And The Effect of Nvidia Reflex
00:22:57 DLSS Frame Generation vs Topaz Video Enhance AI & Adobe After Effects
00:29:51 Wrap Up
Рекомендации по теме
Комментарии
Автор

I suspect the main reason the frame generation is far superior to offline methods is because the GPU has access to motion vectors and other buffers, while offline methods have only the image to work with, it is a huge advantage to have when interpolating frames.

Bankoru
Автор

Since DLSS 3 frames are generated per native frame, I would love a comparison of just the AI frames to just the native frames as a way of comparing the generated image quality.

erikharrison
Автор

For those wondering, I think the corridor benchmark at 13:30 is running at ~17-18fps at native 4K. This would make DLSS2 performance mode ~55 fps, and DLSS3 (performance mode + frame generation), ~93fps. I got this by just going through the video frame by frame, counting each time the elevator moved in Native 4K, over the course of 1 second (60fps container). Which was 17 times. I think this is an accurate way to measure sub-60fps footage in a 60fps container. Just looking at the footage and judging by eye, my estimate was very low 20fps so I'm pretty confident. It's this heavy because it's pathtracing, not simple raytracing.

monkfishy
Автор

I would be interested to see if you could frame skip a recording to only string together ai generated frames (60fps from a 120fps recording) and compare that to the traditionally rendered frames at 60fps. im not sure how much that comparison would help in identifying image quality but it would be an interesting comparison.

cqbSarge
Автор

Fun idea, split a capture every other frame, 60FPS split screen with Even frames (native ones) on the left and Odd frames (results of interpolation) on the right, would be very interesting to see

DavidFregoli
Автор

29:17 I feel another important thing to consider is visual persistence. We don't technically see in a perfect refresh rate. Our brains take shortcuts to compile large amounts of information into something understandable and while we can be sensitive to some issues like microstutters or irregular frametimes, I feel like most people would be unable to see slight visual inconsistencies. I'm sure to a very large degree. A spot like that fence pole would just be filled in so long as there were no "jumping" or "flickering" motions.

muneeb-khan
Автор

Amazing products, incredible step ahead in gaming, horrendous prices and marketing moves from Nvidia.
What a shame...

RandyCR
Автор

@13:12 oof.. you can tell the 4090 at native 4K in that portal sequence is STRUGGLING based off of the speed of those particles. Looks like anywhere between 30-40fps. Full path* tracing stuff is still so crazy demanding.

Nick
Автор

One benefit I can see right away is that even if games are limited to 60/30fps by the engine, DLSS3 "should" be able to insert it's own without affecting the actual speed of the game, or it's physics simulation.

Things like this are fairly rare in modern PC games, but a bunch of console ports from the PS3/360 era could definitely benefit from DLSS3.

cuplaxi
Автор

26:46 I assume DLSS3 does it better than offline software because it has motion vector information, while offline sw is just looking at frames.

franzusgutlus
Автор

DLSS3 artifacts beside lags - 5:15 look at people moving in center of screen - 7:13 buildings in background left side of center of screen behind lights and palm tree... - 9:06 big tower in the middle ... many many more caused by frame generation

jakejoyride
Автор

13:19 in the portal demo, while 500%, the particle effects on the two portals is having trouble keeping those particles rending at full animation, it seems to be vanishing more frequently

ZomgZomg
Автор

The fact that Gsync has to be on prevents you from using BFI on OLED TVs, that's a deal breaker for me, I can already double the motion Clarity by using Motion Pro High on my LG C1. My excitement about DLSS 3 frame generation was to easily get 120 HDR or to get 240 motion Clarity by using BFI on top of 120 frame generation. I'm sure they are locking the feature due to input lag but give me the option to use it with Vsync on without Gsync and of course make it available for the 20 and 30 series that already have optical flow accelerator albeit slower

plasmatvforgaming
Автор

Really excited to see if the errors are perceivable when you try it with a Base-FPS of 40-50 to get over 60!

TilSkywalker
Автор

If DLSS3 interpolates between the current and previous back buffers, it has to introduce a latency of a half engine-frame plus time of the interpolation.
Since for each Present, it takes time to generate the mid-frame past in time (and queue the true backbuffer for a driver-only swap at the next half mid-frame)

So you are playing with the double framerate, but with an added latency slightly higher than a half engine frametime.

patrickgrison
Автор

Fantastic technology. It's a shame that Nvidia is being scummy with their Sku labeling and pricing. Love this type of content so I can at least see what I won't be spending my money on.

matthewbarrios
Автор

The funny thing about DLSS is that the human brain basically has a form of it. Let's call it DLSS 9000.
If our brains didn't use predictive algorithms to help us see the world, everything would be a jumbled mess. Makes sense that AI would use a similar technique.

Numberofthings
Автор

One thing of interest to note is that the motion blur is 2 or 3x longer than it should be when frame generation is turned on. This could cause a strobing effect. The game is generating motion blur with a length that makes sense for 100fps but then it is being displayed at 200fps which will cause extra bluriness. In a perfect world the game would know that frame generation is on and cut the shutter angle of the motion blur in relation to how many frames are being added by dlss 3.0

kimrobingraham
Автор

Excited to get my hands on this




After 4 to 5 generations when I can afford it

ToadyEN
Автор

When DLSS 1 came out, I always wondered if they could use AI to interpolate frames without the artifacts. I'm glad it's actually a thing. I would've liked to see the latency without reflex but the fact that they'll be shipped together and that reflex makes the latency negligible makes DLSS 3 an absolute win. Especially if you have 4k 120hz display

volatilelyle