1 00:00:00,000 --> 00:00:06,200 Everyone and their dog is still angry over the outrageous pricing of NVIDIA's RTX 4000 series. 2 00:00:06,440 --> 00:00:10,400 But if you're trying to figure out some way to justify the expense, 3 00:00:10,480 --> 00:00:18,080 it's worth considering their DLSS 3.0, the newest version of Team Green's AI-powered tech that gives you much better 4 00:00:18,240 --> 00:00:25,840 performance in games without absolutely thrashing your GPU. And it's only officially available on the 4000 series. 5 00:00:25,960 --> 00:00:31,800 So, is it worth it? If you're not familiar with NVIDIA's previous generations of DLSS, 6 00:00:31,800 --> 00:00:39,640 here's a quick primer. The basic idea is that instead of asking your GPU to render every frame at your preferred resolution from scratch, 7 00:00:39,800 --> 00:00:43,000 it uses a machine learning algorithm that runs on special 8 00:00:43,320 --> 00:00:51,640 tensor cores that live on your graphics card. This algorithm attempts to upscale images with little to no discernible quality loss. 9 00:00:52,000 --> 00:00:56,760 So, if your card is having a difficult time rendering a game at, say, 1080p, 10 00:00:56,960 --> 00:01:05,320 the idea is that you enable DLSS, set it to performance, and instead you just have your GPU cores render the game at 540p 11 00:01:05,400 --> 00:01:12,800 with DLSS filling in the blanks to give you a 1080p image that is almost as good as native resolution for far less 12 00:01:13,160 --> 00:01:19,960 computational effort. But the specifics of how the AI works have gone through several revisions. 13 00:01:20,680 --> 00:01:25,880 DLSS 1.0, released in 2019, relied on a neural network that was trained against 14 00:01:26,320 --> 00:01:29,720 specific games by a supercomputer at an NVIDIA facility. 15 00:01:30,160 --> 00:01:37,880 Reception to the feature was mixed at the time, but DLSS 2.0 was much more warmly received when it debuted in 2020. 16 00:01:38,000 --> 00:01:41,520 Not only did 2.0 run better on your card's tensor course, 17 00:01:41,600 --> 00:01:46,840 but it had a more generalized AI model that was still generated by NVIDIA's supercomputer, 18 00:01:46,840 --> 00:01:52,160 but didn't have to be trained on specific games, making it more broadly applicable. 19 00:01:52,320 --> 00:01:58,280 The big change with DLSS 3.0 is what Team Green is calling the Optical Flow Accelerator. 20 00:01:58,480 --> 00:02:01,760 It kind of sounds like a part of a time-traveling DeLorean, 21 00:02:01,760 --> 00:02:09,360 but what it's actually doing is inserting entirely new frames between the ones that your GPU has already rendered. 22 00:02:09,720 --> 00:02:15,080 Here's how it works. The AI examines two sequential frames from the game you're playing, 23 00:02:15,080 --> 00:02:19,080 specifically looking for how the pixels are moving from frame to frame. 24 00:02:19,600 --> 00:02:24,040 Each pixel is examined for changes in lighting, reflections, and particle behavior 25 00:02:24,040 --> 00:02:29,920 to create what NVIDIA calls an optical flow field, basically a model of pixel movement. 26 00:02:30,160 --> 00:02:33,640 DLSS also keeps track of how geometry in the game changes, 27 00:02:33,880 --> 00:02:38,520 such as how the road moves below the character on their motorcycle in the example that we're showing. 28 00:02:39,000 --> 00:02:43,880 Although it is possible to generate frames for smoother motion just by looking at geometry, 29 00:02:43,880 --> 00:02:48,600 quality can suffer as other effects like lighting won't be tracked correctly, 30 00:02:48,600 --> 00:02:53,080 as you can see in this example where the shadow cast by the motorcycle is distorted. 31 00:02:53,480 --> 00:03:00,760 But DLSS 3.0 combines the optical flow field with geometry tracking to give more accurate inserted frames, 32 00:03:00,760 --> 00:03:05,080 giving you significantly higher frame rates while maintaining image accuracy. 33 00:03:05,320 --> 00:03:09,080 It'll be interesting to see just how high you can push frame rates in popular games, 34 00:03:09,080 --> 00:03:14,760 now that DLSS is inserting entire artificial frames in between frames that were already 35 00:03:14,760 --> 00:03:21,160 being upscaled by DLSS. NVIDIA is claiming that your GPU only has to render one eighth 36 00:03:21,160 --> 00:03:26,120 of the pixels that it would have to for similar performance and quality without DLSS. 37 00:03:26,760 --> 00:03:32,120 Unfortunately, the official line from NVIDIA is that DLSS 3.0 is only supported on the new 38 00:03:32,200 --> 00:03:37,400 RTX 4000 series of cards, though a report has already come out that the restriction is due to 39 00:03:37,400 --> 00:03:42,360 a simple software lock. One player was able to bypass it and double his performance in the 40 00:03:42,360 --> 00:03:49,080 notoriously demanding Cyberpunk 2077 while using a now two generations old RTX 2070, 41 00:03:49,080 --> 00:03:53,160 though there were some stability issues. We're going to keep our fingers crossed that NVIDIA 42 00:03:53,160 --> 00:03:57,320 will officially remove this restriction in the near future, especially since they probably 43 00:03:57,320 --> 00:04:01,880 want to clear out some stock of their existing 3000 series products. Keep in mind though, 44 00:04:01,880 --> 00:04:07,640 even if it does happen, it's unlikely to work as well as it should on 4000 series due to its less 45 00:04:07,640 --> 00:04:14,360 powerful hardware. Another caveat of DLSS is that it's still a very sophisticated form of guessing 46 00:04:14,360 --> 00:04:19,960 as to what an individual frame should look like. So depending on your game and your specific hardware, 47 00:04:19,960 --> 00:04:26,200 your mileage may vary and it could add significant latency to your gameplay even if the image quality 48 00:04:26,200 --> 00:04:32,760 is great. So maybe DLSS 4.0 will come out with such a good AI that it can render artificial frames 49 00:04:32,760 --> 00:04:35,880 and even play the game for me? I can only hope.