1
00:00:00,000 --> 00:00:04,800
Deep Learning Super Sampling. This probably sounds like some kind of hokey way to balance

2
00:00:04,800 --> 00:00:09,840
your energy field or something, but it's actually a new AI-powered method from NVIDIA to make your

3
00:00:09,840 --> 00:00:15,440
games look better. Now, you might have already been familiar with traditional super sampling. That is,

4
00:00:15,440 --> 00:00:20,480
when your graphics cards render frames at a higher resolution than what your monitor can support,

5
00:00:20,480 --> 00:00:25,120
but then downscales the image to fit your display. Although it's very computationally taxing,

6
00:00:25,120 --> 00:00:29,520
it can provide a significant quality boost that other anti-aliasing methods may not be able to

7
00:00:30,080 --> 00:00:34,240
achieve. DLSS, however, is quite different than standard super sampling, and it's actually

8
00:00:34,240 --> 00:00:39,520
less demanding on your GPU. Instead of simply forcing your GPU to render higher resolution

9
00:00:39,520 --> 00:00:44,640
frames from scratch, it uses a neural network to predict what the frames should look like.

10
00:00:44,640 --> 00:00:49,040
The neural network is trained by an NVIDIA supercomputer that feeds it correct frames

11
00:00:49,040 --> 00:00:53,840
from certain games to help it learn how to generate extra pixels accurately. These

12
00:00:53,920 --> 00:00:59,680
correct frames are actually 16k images, so the AI will have a very granular level of detail to

13
00:00:59,680 --> 00:01:04,800
learn from. The AI model is then sent to your GPU via driver updates so it can be run locally. The

14
00:01:04,800 --> 00:01:09,440
idea here is that it's easier for your GPU to run this neural network when you're playing a

15
00:01:09,440 --> 00:01:13,920
graphically intense game running at a low frame rate than it is to keep drawing new frames from

16
00:01:13,920 --> 00:01:20,320
scratch. Regardless of how difficult the game is to run, DLSS uses a fixed amount of time for frame,

17
00:01:20,320 --> 00:01:25,120
so it often doesn't take as long for your GPU to spit out a DLSS-assisted frame

18
00:01:25,120 --> 00:01:30,480
as it does to render one the old-fashioned way, partly thanks to newer RTX GPUs having specialized

19
00:01:30,480 --> 00:01:34,800
tensor cores that are supposed to be optimized for running AI. They kind of are.

20
00:01:34,800 --> 00:01:40,080
And a new version, DLSS 2.0, was recently rolled out and features a few key improvements. For

21
00:01:40,080 --> 00:01:45,360
starters, DLSS 2.0 aims to provide near native resolution quality while the GPU renders well

22
00:01:45,360 --> 00:01:49,840
under half the pixels it would otherwise need to handle. There are also efficiency improvements to

23
00:01:49,840 --> 00:01:54,000
the neural network that should help it process images faster, ultimately boosting frame rates.

24
00:01:54,000 --> 00:01:58,320
Another big strength is that now the neural network is much more generalized. Instead of

25
00:01:58,320 --> 00:02:03,360
needing to separately train it for every game, NVIDIA now uses more general visual content that's

26
00:02:03,360 --> 00:02:07,600
supposed to be more representative of a variety of different games, which means improvements can

27
00:02:07,600 --> 00:02:13,280
be delivered to users more quickly and more games will end up supporting DLSS. Users can now decide

28
00:02:13,280 --> 00:02:17,600
how much they want to utilize DLSS technology rather than just leaving it completely up to

29
00:02:17,600 --> 00:02:22,960
the GPU and driver. Gamers can now choose between three modes for image quality, prioritizing either

30
00:02:22,960 --> 00:02:27,760
greater image quality or higher frame rates. The benefits of all of this can mostly been seen when

31
00:02:27,760 --> 00:02:32,480
looking at the finer details in a scene, not just at the edges of objects where you traditionally

32
00:02:32,480 --> 00:02:37,520
see jaggies, if the anti-aliasing isn't up to par. Things like text, chain link fences, and

33
00:02:37,520 --> 00:02:42,800
details on faraway buildings should be a lot clearer with DLSS without game breaking slowdowns.

34
00:02:42,800 --> 00:02:47,680
And in some cases, DLSS even appears to be able to make visual elements look more detailed

35
00:02:47,680 --> 00:02:52,400
than they even would be with native resolution, despite rendering fewer pixels from scratch,

36
00:02:52,400 --> 00:02:56,800
if the AI figures out how to enhance a texture or shadow to a greater extent than what the

37
00:02:56,800 --> 00:03:01,120
game's code originally calls for. Of course, with other parts of an image, it might still be

38
00:03:01,120 --> 00:03:05,760
inferior to standard rendering. But the hope is as processing power increases and the algorithm

39
00:03:05,760 --> 00:03:10,960
continues to be fine-tuned, it'll become easier and easier to run more and more games at buttery

40
00:03:10,960 --> 00:03:16,080
smooth frame rates without compromising quality. Thanks for watching. Like, dislike,

41
00:03:16,080 --> 00:03:19,760
check out our other videos, comment with video suggestions, and don't forget to subscribe and

42
00:03:19,760 --> 00:03:20,560
follow below.