WEBVTT 00:00:00.000 --> 00:00:04.160 Making your games look better, without having to drop a bunch of money on a new graphics 00:00:04.160 --> 00:00:09.280 card, is something that GPU manufacturers are keenly aware that gamers want, especially 00:00:09.280 --> 00:00:13.760 now that a new GeForce or Radeon card is going to cost you a spare kidney on the secondary 00:00:13.760 --> 00:00:19.760 market. Now, NVIDIA and AMD have both tried to use their own secret sauces to make this happen, 00:00:19.760 --> 00:00:25.760 both NVIDIA's DLSS and AMD's FSR have made plenty of headlines recently, but Team Green 00:00:25.760 --> 00:00:30.440 is adding another feature to its arsenal that it hopes will give it a leg up on AMD and 00:00:30.440 --> 00:00:38.840 maybe give you a leg up in your games. They call it DLDSR, or Deep Learning Dynamic Super Resolution, and if that sounds familiar, 00:00:38.840 --> 00:00:43.240 it's because NVIDIA's regular dynamic super resolution, you know, without the deep learning 00:00:43.240 --> 00:00:47.840 part, has been with us since the days of the GTX970 and 980. 00:00:47.840 --> 00:00:53.600 So let's talk about what exactly that is, and how deep learning improves upon it. 00:00:53.600 --> 00:01:01.400 DLDSR is a super sampling technique. Basically, it uses your graphics card to render your game at a higher resolution than what 00:01:01.400 --> 00:01:06.960 your monitor natively supports, then it scales that down to match your screen. 00:01:06.960 --> 00:01:11.160 This produces better looking images than just rendering at your display's native resolution. 00:01:11.160 --> 00:01:18.760 But why? Well, DSR more specifically works by taking samples of groups of neighboring pixels from 00:01:18.760 --> 00:01:23.920 the higher resolution rendered image and blending them into a single pixel that you see on your 00:01:23.920 --> 00:01:28.200 lower resolution display to make the resulting image more accurate. 00:01:28.200 --> 00:01:34.200 This is essentially the same way that Super Sample Anti-Aliasing, or SSAA, works. 00:01:34.200 --> 00:01:39.520 SSAA has been around for longer than DSR, but games had to implement it individually. 00:01:39.520 --> 00:01:44.600 DSR has the advantage of not requiring work from game developers and also puts the image 00:01:44.720 --> 00:01:49.240 through NVIDIA's own filter that takes into account additional surrounding pixels to further 00:01:49.240 --> 00:01:55.200 improve image quality. DSR has a habit of making images a little softer than traditional SSAA, meaning that 00:01:55.200 --> 00:02:01.200 some gamers like it and others don't. Regardless, it's a popular way to improve image quality, especially as it allows more 00:02:01.200 --> 00:02:04.960 granularity with exactly how hard you want your GPU to work. 00:02:04.960 --> 00:02:09.400 Instead of rendering the image at two times or four times your monitor's resolution, 00:02:09.400 --> 00:02:15.000 you could specify something like 2.25 times, depending on how powerful your GPU is. 00:02:15.000 --> 00:02:20.720 But both DSR and SSAA are very computationally expensive, since you're basically asking 00:02:20.720 --> 00:02:25.680 your GPU to do way more work and then throw away most of it. 00:02:25.680 --> 00:02:32.520 That is where DLDSR comes in. It leverages the tensor cores in RTX series GPUs that are designed with machine learning 00:02:32.520 --> 00:02:38.960 and AI in mind. The idea behind DLDSR is that the neural network and the tensor cores can figure out 00:02:39.040 --> 00:02:45.240 what the image should look like by sampling a smaller number of pixels than regular DSR 00:02:45.240 --> 00:02:53.000 or SSAA. One example NVIDIA gave was from the game Prey, which was rendered at 1620p rather than 4K, 00:02:53.000 --> 00:03:00.640 then downscaled to 1080p. This was accomplished with nearly no frame rate drop compared to native 1080p rendering, 00:03:00.640 --> 00:03:05.520 and the image actually looked a little better than the traditional DSR image that downscaled 00:03:05.520 --> 00:03:12.000 all the way from 4K. Similar to regular DSR, DLDSR doesn't need per game support, so it can actually make 00:03:12.000 --> 00:03:18.120 your older games look refreshed. But if you're using a title that supports DLSS, you can combine the two features for 00:03:18.120 --> 00:03:22.280 an extra quality boost without sacrificing too much performance. 00:03:22.280 --> 00:03:29.360 Don't expect it to work miracles, though. Early reviews have said that the quality improvement over regular DSR is pretty marginal, and of 00:03:29.360 --> 00:03:34.000 course, turning up the scaling factor is still going to slow your frame rate down. 00:03:34.000 --> 00:03:38.680 Notice that the impact won't be as huge as it would be with traditional super sampling. 00:03:38.680 --> 00:03:43.000 You'll need an RTX card to take advantage of DLSR, but don't feel completely left out 00:03:43.000 --> 00:03:47.800 if you have a card from Team Red, as AMD is set to launch its own AI-based Radeon Super 00:03:47.800 --> 00:03:53.680 Resolution technology in early 2022, which should work on any RDNA2 GPU. 00:03:53.680 --> 00:03:58.240 AMD cards, at least for now, don't have dedicated machine learning cores like NVIDIA's 00:03:58.240 --> 00:04:02.240 do, so it'll be interesting to see how the two super sampling methods ultimately stack 00:04:02.240 --> 00:04:15.760 up against each other. I just hope they'll both be super.