WEBVTT 00:00:00.000 --> 00:00:04.500 Here it is, NVIDIA's new RTX 5080. 00:00:04.500 --> 00:00:11.580 Compared to its predecessor, the 4080 Super, it is around 15% faster in games that use its new features. 00:00:11.580 --> 00:00:16.300 It's faster in AI, as long as you don't need more than 16 gigs of RAM for your model. 00:00:16.300 --> 00:00:22.380 It's more efficient, even though it draws more power overall, and it's got a great cooler design, 00:00:22.380 --> 00:00:25.420 as long as you don't mind heating up your CPU. 00:00:25.420 --> 00:00:29.540 And that's it. That's the most positive spin we can put on this thing. 00:00:29.580 --> 00:00:33.980 Why the f*** do we need all of these caveats when we say 00:00:33.980 --> 00:00:38.380 the new one is better than the old one? I mean, isn't the whole point of the new one 00:00:38.380 --> 00:00:42.100 to be clearly and substantially better? Well, it is better. 00:00:42.100 --> 00:00:45.900 In cranked out cyberpunket 4K, it not only beats the 4080 Super, 00:00:45.900 --> 00:00:49.240 but it even gives the 4090 a run for its money. 00:00:49.240 --> 00:00:54.000 But it's such a mixed bag that in other situations, it barely outcompetes the 7900 XTX, 00:00:54.000 --> 00:01:00.300 the flagship GPU that made AMD think, you know, maybe we just shouldn't do flagships anymore. 00:01:00.300 --> 00:01:03.540 Hey, glass half full? At least they didn't raise the price. 00:01:03.540 --> 00:01:07.940 And we did find some situations where the 5080 offers a pretty compelling use case. 00:01:07.940 --> 00:01:11.940 And if you don't mind the weirdness of generated frames, maybe NVIDIA is onto something 00:01:11.940 --> 00:01:14.940 with the neural rendering stuff. We're gonna tell you more about it. 00:01:14.940 --> 00:01:18.060 After I nearly render, a better way to sell your used hardware. 00:01:18.400 --> 00:01:19.240 Brrrrr. 00:01:22.900 --> 00:01:26.900 ["Side Story of the Wildlands"] 00:01:29.820 --> 00:01:33.100 Before we explore the wild lands of AI assisted rendering, let's get our bearings 00:01:33.100 --> 00:01:38.740 with some good old fashioned 1440P results. All raster, no deal less upscaling, or ray tracing. 00:01:38.740 --> 00:01:44.020 In cyberpunk, the 5080 makes a great first impression with a 20% lead over the 4080 Super. 00:01:44.020 --> 00:01:48.880 Then in Alan Wake, the 5080 brings a decent generation uplift that brings 1% low FPS 00:01:48.880 --> 00:01:54.120 into the triple digits. But also it only barely outpaces the 7900 XTX here. 00:01:54.120 --> 00:01:58.880 In F124, things look even less exciting with the 5080 failing to overtake AMD's 00:01:58.880 --> 00:02:04.000 now two-year-old flagship and only offering a measly seven-ish percent uplift 00:02:04.000 --> 00:02:09.320 over the 4080 Super. Does it get any better? No, no again and triple no. 00:02:09.320 --> 00:02:14.800 Why does the last-gen 4070 Ti Super make everything else here look like an overpriced joke? 00:02:14.800 --> 00:02:19.760 At least Blacksmith Wukong provides a reason for the 5080 to exist, but overall, I'm just kind of getting sad 00:02:19.760 --> 00:02:23.320 reading you all these charts. And you can expect a roughly 10% uplift 00:02:23.320 --> 00:02:28.040 over the 4080 Super in just 7% over the 7900 XTX. 00:02:28.040 --> 00:02:31.600 If you're curious about the 4080 non-super, there's a reason we called it the 4080 cheaper. 00:02:31.600 --> 00:02:36.360 It performs basically identically and you can't buy it anymore. So we opted to use our limited graph space 00:02:36.360 --> 00:02:42.040 for older cards that might actually be worth upgrading from. All right, well, maybe the story changes at 4K. 00:02:42.040 --> 00:02:46.800 Ooh, yeah, okay. So it's a bit happier, but this feels a little bit like going from bridge 00:02:46.800 --> 00:02:49.820 to terabithia to dead poet society. 00:02:49.820 --> 00:02:53.320 On average, the 5080 takes a commanding 18% lead 00:02:53.320 --> 00:02:57.160 over its predecessor, which is good. But that parade gets somewhat rained on 00:02:57.160 --> 00:03:01.400 by its mere 10% lead over the 7900 XTX. 00:03:01.400 --> 00:03:04.560 There are a couple of standouts that merit specific mention here. 00:03:04.560 --> 00:03:07.780 Blacksmith Wukong has the 5080 taking a price-appropriate lead 00:03:07.780 --> 00:03:10.920 over the 7900 XTX with a worthwhile uplift 00:03:10.920 --> 00:03:14.060 over the 4080 Super. And that's also true in Cyberpunk, 00:03:14.060 --> 00:03:17.880 where again, the 5080 nearly catches the 4090 00:03:17.880 --> 00:03:21.520 losing only in the less important average FPS. 00:03:21.520 --> 00:03:26.440 But that one result there, it's less of a trend and more of just an outlier. 00:03:26.440 --> 00:03:31.600 In other titles, it seems like the 5080 is content to just be its disappointing self. 00:03:31.600 --> 00:03:34.600 Better, but not better enough. 00:03:34.600 --> 00:03:38.600 Of course, NVIDIA doesn't think that traditional raster rendering is the way forward. 00:03:38.600 --> 00:03:42.200 No, their way forward is with the AI GPU hardware 00:03:42.200 --> 00:03:46.680 that made them the most valuable company in the world and any conflicting data would almost certainly 00:03:46.680 --> 00:03:52.880 mean a cratering stock price. So with the AI boom in mind, NVIDIA focused most 00:03:52.880 --> 00:03:56.920 of their architectural upgrades on that. If you want to learn more about the architectural 00:03:56.920 --> 00:04:01.800 and technology upgrades, we highly recommend you check out our 5090 review where we spend a bit more time 00:04:01.800 --> 00:04:06.320 explaining what everything means. What really matters for gamers is how this card differs 00:04:06.320 --> 00:04:09.640 from the 4080 and 4080 Super. It's not much. 00:04:09.680 --> 00:04:15.040 We're still on TSMC's 4N process node, but at the same MSRP, we get four more streaming 00:04:15.040 --> 00:04:19.360 multi processors in the 4080 Super, along with a clock speed bump and a memory swap 00:04:19.360 --> 00:04:24.000 from 16 gigabytes of GDR6X to 16 gigabytes of GDR7. 00:04:24.000 --> 00:04:27.080 That all explains why raw gaming performance isn't improved much. 00:04:27.080 --> 00:04:31.760 The bigger improvements come in the form of next generation NVIDIA ray tracing, RT cores, 00:04:31.760 --> 00:04:35.920 and AI accelerating tensor cores. The tensor cores also now have hardware support 00:04:35.920 --> 00:04:40.720 for FP4 calculations for improved performance and memory usage with lightweight AI models 00:04:40.720 --> 00:04:45.440 and are more tightly integrated with CUDA cores to make way for the sci-fi sounding neural rendering. 00:04:45.440 --> 00:04:50.840 For hypothetical future gamers, these neural rendering techniques actually sound really cool and they might turn out great, 00:04:50.840 --> 00:04:55.480 but gamers shouldn't feel pressure to upgrade until a must play title that uses them hits the Steam Store. 00:04:55.480 --> 00:05:00.200 Instead feel pressured to buy our precision multi bit screwdriver bundle from LTCstore.com. 00:05:00.200 --> 00:05:03.720 Anywho, those working with high end video might upgrade for the new media engine 00:05:03.720 --> 00:05:07.800 on the 5000 series that provides support for 422 chroma subsampling. 00:05:07.800 --> 00:05:11.440 In a nutshell, this means much higher quality hardware accelerated video encoding, 00:05:11.440 --> 00:05:14.560 which will be a boon for professionals and anyone who like streaming their games 00:05:14.560 --> 00:05:20.360 to a hand older lightweight decoding box. We're also stoked about DisplayPort 2.1 UHBR20. 00:05:20.360 --> 00:05:24.200 The 58 even drive 4K 240 Hertz with no display stream compression 00:05:24.200 --> 00:05:28.120 and even go higher with DSC, which doesn't impact image quality 00:05:28.120 --> 00:05:32.400 as far as our eyes can perceive. That means no compromises between image quality 00:05:32.400 --> 00:05:35.640 and latency and responsiveness. Let's move on to ray tracing, 00:05:35.640 --> 00:05:39.240 where, yeah, NVIDIA is really just competing against themselves here. 00:05:39.240 --> 00:05:42.480 But as more big titles list ray tracing as a minimum requirement, 00:05:42.480 --> 00:05:45.560 it's becoming something that just can't be brushed off, 00:05:45.560 --> 00:05:50.320 which is a problem for AMD. At 1440p, while AMD can survive lightly ray traced games, 00:05:50.320 --> 00:05:54.680 the path tracing found in Alan Wake and Blackmouth Wukong make AMD look like an eight year old kid 00:05:54.680 --> 00:05:59.640 made it onto the field at an NFL game. Hopefully that will change with AMD's 9000 series of GPUs, 00:05:59.640 --> 00:06:02.800 but we don't know for sure. Today, across our suite of games, 00:06:02.800 --> 00:06:06.960 the 5080 maintains the same performance margin over the 4080 Super that we saw in our raster tests, 00:06:06.960 --> 00:06:10.440 but fails to scale as well as the 4090 does into 4K. 00:06:10.440 --> 00:06:14.880 Maybe the 4090's extra RT cores give it the edge, or maybe it's the VRAM. 00:06:14.880 --> 00:06:18.880 Check out this result in Alan Wake 2. Why did so many cards fail miserably? 00:06:18.880 --> 00:06:22.680 Guess now, did you say lack of VRAM? Well, you're right. 00:06:22.680 --> 00:06:29.680 And if you didn't say lack of VRAM, pretend that you did. 4K path traced Alan Wake asks for over 15 gigabytes of VRAM 00:06:29.680 --> 00:06:33.520 on 50 series cards and over 16 on AMD cards. 00:06:33.560 --> 00:06:38.440 Anything less and it's wholly unplayable. Well, we can't say that it causes a performance hit 00:06:38.440 --> 00:06:42.440 on the 5080. It highlights just how skimpy 16 gigabytes of VRAM feels 00:06:42.440 --> 00:06:46.680 on a $1,000 card and how close we are to a future 00:06:46.680 --> 00:06:49.960 where it's necessary for high end gaming. And the more frustrating thing is that 00:06:49.960 --> 00:06:54.600 if you want more VRAM from Team Green, you literally have to spend twice as much for a 5090 00:06:54.600 --> 00:06:58.120 or go back a generation or two. I digress. 00:06:58.120 --> 00:07:03.440 The rest of the 4K RT results show that the 5080 is still a mean green ray tracing machine. 00:07:03.440 --> 00:07:09.080 It's just not that much meaner than the 4080 Super. Another disappointing result for the 5,000 series. 00:07:09.080 --> 00:07:14.080 For productivity, the story is more mixed. While it doesn't do much to impress in our blender render, 00:07:14.080 --> 00:07:19.080 in Premiere Pro, we can see that the updated media engine on the new 50 series puts in the work, 00:07:19.080 --> 00:07:22.440 beating the 4090 by 5% with a similar lead 00:07:22.440 --> 00:07:25.600 in DaVinci Resolve. We'd love to talk about improvements to encoding, 00:07:25.600 --> 00:07:30.020 but we're working on our benchmarks for that. Let us cook. But who needs to shoot and edit videos 00:07:30.020 --> 00:07:34.460 when you could just generate garbage while wasting precious, earthly resources? 00:07:34.460 --> 00:07:38.820 In AI, the 5080 manages a 20% lead over the 4080 Super 00:07:38.820 --> 00:07:44.620 in our text generation benchmarks, with the only hiccup being a smaller lead in our 5 3.5. 00:07:44.620 --> 00:07:50.660 In our stable diffusion image generation benchmarks, we see 16-ish percent improvement over the 4080 Super, 00:07:50.660 --> 00:07:54.760 which is cool, I guess. Sadly, a lack of VRAM will limit the models 00:07:54.760 --> 00:07:58.620 you can play with. Womp, womp. But this is positioned as a gaming card. 00:07:58.620 --> 00:08:01.700 So let's talk about the AI that makes games run gooder. 00:08:01.700 --> 00:08:05.020 DLSS4 is a massive overhaul from DLSS3. 00:08:05.020 --> 00:08:09.460 It switches from a CNN model to a transformer model. If you want to learn more about the theory, 00:08:09.460 --> 00:08:13.380 you should check out our 5090 review. What it means in practice is that DLSS4 00:08:13.380 --> 00:08:17.180 allows for better image quality at the same settings with some caveats. 00:08:17.180 --> 00:08:21.740 Of course, it's the 50 series. There's going to be caveats. There's a performance hit using the new transformer model. 00:08:21.740 --> 00:08:25.820 While the effect is mild for things like super resolution, Digital Foundry saw that older cards 00:08:25.820 --> 00:08:31.020 can take a real hit from enabling certain features with the new model. But it scales better, which means that the 50 series 00:08:31.020 --> 00:08:36.340 and all of its tensor core chutzpah generates up to three frames for every rendered frame, 00:08:36.340 --> 00:08:39.900 essentially quadrupling your FPS. Multiframe Gen is impressive. 00:08:39.900 --> 00:08:44.060 And despite generating so many frames, the tech does not impact latency substantially, 00:08:44.060 --> 00:08:48.740 at least over top of your base frame rate. Sadly, more frames is not more performance. 00:08:48.740 --> 00:08:53.780 If you were to run a game natively at 30 FPS, each frame would have 33 milliseconds of latency. 00:08:53.780 --> 00:08:57.020 Running at 60 FPS would have your latency to 16.66, 00:08:57.060 --> 00:09:00.060 doubling your FPS again, 8.33 milliseconds. 00:09:00.060 --> 00:09:04.260 You might see where I'm going with this. If you quadruple your 30 FPS game with frame gen, 00:09:04.260 --> 00:09:08.620 you will still have that 33.33 milliseconds of latency, which is 25 milliseconds worse than 00:09:08.620 --> 00:09:11.940 rendering 120 FPS natively, a large disparity. 00:09:11.940 --> 00:09:16.140 Meanwhile, if you quadruple your 120 FPS game to 480 FPS, 00:09:16.140 --> 00:09:19.420 well, that only results in about a six millisecond difference 00:09:19.420 --> 00:09:24.260 in latency, which is far less noticeable. But here's the kicker. MFG has a performance overhead 00:09:24.260 --> 00:09:27.860 that will reduce your base frame rate. And the more frames you gen, 00:09:27.860 --> 00:09:33.460 the more overhead you need to contend with. That's why we don't see a literal quadrupling of frames. 00:09:33.460 --> 00:09:38.300 This means that using it to bump up an already low frame rate will make it feel even worse to play 00:09:38.300 --> 00:09:41.540 as it adds more latency. There's a ton of nuance to this tech, 00:09:41.540 --> 00:09:45.020 and we would recommend checking out Hard Run Box's excellent video on it. 00:09:45.020 --> 00:09:49.340 The good news is that multi-frame gen scales about as well in the 5080 as it did in the 5090. 00:09:49.340 --> 00:09:55.060 Hopefully this trend carries forward to the lower 50 series because it did not carry over to the lower end 00:09:55.060 --> 00:09:59.940 of the 40 series, sheesh. So TLDR, the higher your base frame rate, 00:09:59.940 --> 00:10:03.820 the better your frame gen will feel, but the less you need it. 00:10:03.820 --> 00:10:08.340 NVIDIA likes to pretend that frame gen is comparable to native rendering in their marketing, but it's just not. 00:10:08.340 --> 00:10:11.940 Should you forego MFG and just stick with the new DLSS4 transformer model? 00:10:11.940 --> 00:10:16.420 On the 5080, you'll see about a 5% lower frame rate for the improved image quality. 00:10:16.420 --> 00:10:20.380 And we think it's worth the trade-off. Now, all of this talk of frame rates and latency 00:10:20.380 --> 00:10:25.100 might have you wondering, how does the 5080 fair in eSports? Well, it does a dang good job. 00:10:25.100 --> 00:10:28.260 We didn't test the rest of the lineup today because frankly, any of the high-end cards 00:10:28.260 --> 00:10:32.820 will give you a great experience in eSports titles, but this is a reminder that despite our dour 00:10:32.820 --> 00:10:36.980 and disappointed tone, it's still a very high-end, very performant card. 00:10:36.980 --> 00:10:41.100 And that also means it has big power draw. The 5080 gets the same sleek 00:10:41.100 --> 00:10:46.700 and innovative double flow-through cooler as the 5090, allowing it to stay at a reasonable 70 degrees in combustor 00:10:46.700 --> 00:10:49.740 despite pulling as much as 403 watts. 00:10:49.740 --> 00:10:54.260 If you're going to small form factor, you're far less likely to bump up against the thermal moments of this card 00:10:54.260 --> 00:10:58.620 than with the 5090. When gaming, the average temps drop further to 65 degrees 00:10:58.620 --> 00:11:04.660 thanks to the lower power budget. In F124, we see the 5080 pulling an average of 337 watts, 00:11:04.660 --> 00:11:09.700 nearly 15% more than the 4080 Super. This nets around 14% more frames, 00:11:09.700 --> 00:11:13.540 which indicates no efficiency gains, just like we saw in the 5090. 00:11:13.540 --> 00:11:16.540 Bummer. Feeling deflated? Me too. 00:11:16.540 --> 00:11:20.020 Taking everything into account, we have a next-generation card 00:11:20.020 --> 00:11:26.420 that outperforms its last-gen counterpart by 10 to 20% using roughly 15% more power. 00:11:26.420 --> 00:11:30.900 I mean, at least the price is the same. That's good. Even then, it feels like it could have been 00:11:30.900 --> 00:11:36.140 a mid-cycle refresh or a 4080 Ti. Given it doesn't catch the 4090 in raw performance, 00:11:36.140 --> 00:11:40.420 something we really hoped it would do. At best, if you're considering the 4080 Super, 00:11:40.420 --> 00:11:44.620 the 5080 might be enough to push you over the edge to buy, at least at MSRP, 00:11:44.620 --> 00:11:49.340 but that's the hardest working S in the world. Because last-gen cards are nearly impossible 00:11:49.340 --> 00:11:53.140 to find at NVIDIA's suggested price, and if the 5090 is anything to go by, 00:11:53.140 --> 00:11:55.580 this generation will be the same or even worse. 00:11:57.980 --> 00:12:02.140 It feels like NVIDIA just doesn't care much about gamers. That's because they don't. 00:12:02.140 --> 00:12:05.180 They care about AI. That's what made them one of the most valuable companies 00:12:05.180 --> 00:12:10.140 in the world, and they're trying to use their large market share to move the market in their direction. 00:12:10.140 --> 00:12:13.220 They're saying, no, no, no, no, no. You don't want more raster. 00:12:13.220 --> 00:12:16.540 What the industry wants, nay, what it needs is AI. 00:12:16.540 --> 00:12:21.540 And AI just so happens to need the proprietary features enabled by the chips that we produce. 00:12:21.540 --> 00:12:26.780 I get it. If they were to admit that AI hardware wasn't the solution to every conceivable problem, 00:12:26.780 --> 00:12:31.100 then it would crater their stock price, a thing that actually just happened on Monday. 00:12:31.100 --> 00:12:35.620 The moment a Chinese AI company with limited access to NVIDIA's best AI hardware 00:12:35.620 --> 00:12:41.220 supposedly overtook chat GPT, Team Green's stock dropped 17 points. 00:12:41.220 --> 00:12:45.580 If the hype for AI, and more specifically to them, AI hardware deflates, 00:12:45.580 --> 00:12:49.620 Jensen might have to trade his crocodile leather jacket for a regular leather jacket. 00:12:49.620 --> 00:12:53.260 But with all of that said, without a clear competitor in the GPU space, 00:12:53.260 --> 00:12:57.620 NVIDIA will still sell a buttload of these to gamers and will continue to drive the industry 00:12:57.620 --> 00:13:02.020 in the direction of their choosing. So until AMD or Intel brings the heat, 00:13:02.020 --> 00:13:05.860 this is what we got. Really kind of feels like the Intel, 00:13:05.860 --> 00:13:10.940 hey, that's a nice quad core you got there. How about this shiny new quad core days? 00:13:10.940 --> 00:13:16.460 All I can do is hope at this point then that NVIDIA can nearly render some more innovative ideas 00:13:16.460 --> 00:13:20.020 before they suffer the same eventual fate that Intel did. 00:13:20.020 --> 00:13:24.460 Just like I'm gonna nearly render this segue to our sponsor. 00:13:24.460 --> 00:13:30.800 What, you expected an AI to come up with a new segue? I mean, it was trained on my segue to our sponsor. 00:13:30.800 --> 00:13:34.540 Thanks for watching guys. If you liked this video, check out our 5090 review 00:13:34.540 --> 00:13:38.380 or check out lttlabs.com for even more info on these new cards. 00:13:38.380 --> 00:13:40.260 The site is always getting updated.