1 00:00:00,000 --> 00:00:04,560 It's here, it is ripping fast, and it's $2,000 US dollars. 2 00:00:04,560 --> 00:00:07,800 But in exchange for your least favorite of kidneys, 3 00:00:07,800 --> 00:00:11,960 NVIDIA promises that their brand new GeForce RTX 5090 4 00:00:11,960 --> 00:00:17,480 will deliver a level of performance that obliterates their only real competition. 5 00:00:17,480 --> 00:00:20,560 NVIDIA, more GPU cores, boom. 6 00:00:20,560 --> 00:00:23,760 More VRAM, and faster VRAM, boom, boom. 7 00:00:23,760 --> 00:00:31,800 Enhanced RT and AI cores, wider memory bus, and PCIe Gen 5, boom, boom, boom. 8 00:00:31,800 --> 00:00:37,000 On top of that, NVIDIA has packed in a deep learning super shed load of new features that I would 9 00:00:37,000 --> 00:00:41,560 love to tell you about. But unless NVIDIA also invented AI teeth extraction, 10 00:00:41,560 --> 00:00:46,520 I won't be able to. So instead, I leave this review in your capable hands. 11 00:00:46,520 --> 00:00:47,400 See you later. 12 00:00:49,920 --> 00:00:52,920 I got it, I got it, I got it. First up, graphics performance. 13 00:00:52,920 --> 00:00:54,000 Let's cut to the chase. 14 00:00:56,520 --> 00:01:04,760 Let's get right to raw gaming results. 15 00:01:04,760 --> 00:01:08,360 No ray tracing, no upscaling, and we're starting with 1440p. 16 00:01:08,360 --> 00:01:13,000 Across our suite of games at 1440p, the 5090 never falls flat on its face, obviously, 17 00:01:13,000 --> 00:01:16,120 but still manages to be underwhelming. 18 00:01:16,120 --> 00:01:21,520 In the Vulcan-based Red Dead Redemption 2, we see less than a 10% improvement over the 4090, 19 00:01:21,520 --> 00:01:24,880 and that lackluster uplift is repeated in F124. 20 00:01:24,920 --> 00:01:30,960 More problematic is the embarrassingly small 30% lead over the 7900XTX, which frequently 21 00:01:30,960 --> 00:01:34,520 goes for a little over two-fifths of the price. 22 00:01:34,520 --> 00:01:40,040 Oof. And Returnal doesn't bring better news. But as we move on to newer, more graphically intensive games, 23 00:01:40,040 --> 00:01:44,640 the 5090 does start to pull away from the pack. In the gorgeous thriller, Alan Wake 2, 24 00:01:44,640 --> 00:01:52,520 it beats the 3090 Ti by more than double and looks great in Blacksmith Wukong, beating the 4090 by 27%. 25 00:01:52,560 --> 00:01:56,800 Cyberpunk is another strong point compared to the previous generations, but as low as they are 26 00:01:56,800 --> 00:02:02,240 on the chart, it's worth noting AMD's strong performance per dollar in this game, at least when ray tracing 27 00:02:02,240 --> 00:02:07,960 isn't enabled. We'll get to that later. For now, this might be obvious, but if you are a 1440p player, 28 00:02:07,960 --> 00:02:11,560 the 5090 is overkill with the current crop of CPUs. 29 00:02:11,560 --> 00:02:15,960 If you're on the latest 9800X3D, you might find that the 5090 exerts a little bit more 30 00:02:15,960 --> 00:02:19,000 of a commanding lead, but I think that anyone with this setup 31 00:02:19,000 --> 00:02:22,720 should be putting their money into a new monitor rather than a new CPU. 32 00:02:22,720 --> 00:02:27,320 Let's move on to 4K testing, where the CPU bottlenecks are less likely to rear their ugly heads. 33 00:02:27,320 --> 00:02:30,400 In Cyberpunk, the 5090 is the first card 34 00:02:30,400 --> 00:02:33,560 to ever crack triple digits at our ultra preset, 35 00:02:33,560 --> 00:02:37,520 scoring a 30 FPS lead over the 4090. 36 00:02:37,520 --> 00:02:42,440 In Alan Wake 2, the story remains largely the same, offering a noticeable difference in playability 37 00:02:42,440 --> 00:02:47,040 compared to any previous flagship. Blacksmith Wukong at cinematic settings, however, 38 00:02:47,040 --> 00:02:50,520 is the Everest-like summit, where even the 5090 39 00:02:50,560 --> 00:02:56,440 falls short of 60 FPS average. Perhaps an overclocking Sherpa could get us to the peak, 40 00:02:56,440 --> 00:03:02,000 but that's a subject for another day. In Red Dead Redemption 2, the 5090 does not impress, 41 00:03:02,000 --> 00:03:06,160 especially when you consider its price. And in F124, the 5090 continues to operate 42 00:03:06,160 --> 00:03:11,360 at that kind of level of performance that no one else can touch. The problem is that for all the hype, 43 00:03:11,360 --> 00:03:18,320 the performance bump is roughly on par with the price bump, making the 5090 look less like a truly next-generation product 44 00:03:18,360 --> 00:03:22,040 and more like a 4090 Super GT Zikaiburkei. 45 00:03:22,040 --> 00:03:27,360 But what's the deal? I thought that Blackwell was supposed to be the giant leap forward with all that flip metering 46 00:03:27,360 --> 00:03:30,800 and neural rendering and increased rate triangle intersections. 47 00:03:30,800 --> 00:03:34,800 What the fuck? Whoa, whoa, whoa, whoa, all right, hold on. That's a lot of words, 48 00:03:34,800 --> 00:03:37,920 but to understand how they're going to impact performance, 49 00:03:37,920 --> 00:03:41,480 and they will, we need to understand what they mean. 50 00:03:41,480 --> 00:03:45,360 See, Blackwell brings so many new enhancements 51 00:03:45,400 --> 00:03:50,000 that NVIDIA marketing doesn't even call it a GPU architecture. 52 00:03:50,000 --> 00:03:54,200 No, it's called a neural rendering architecture. 53 00:03:54,200 --> 00:03:58,240 So what is that? As far as we can tell, 54 00:03:58,240 --> 00:04:02,200 it's about equal parts, genuine innovation and marketing fluff. 55 00:04:02,200 --> 00:04:07,520 We'll start with the innovation. Up until this point, NVIDIA's AI accelerating tensor cores 56 00:04:07,520 --> 00:04:12,480 could not be accessed by a graphics API like Vulkan or DirectX. 57 00:04:12,480 --> 00:04:18,040 But through collaboration with Microsoft, DirectX now has the Cooperative Vectors API, 58 00:04:18,040 --> 00:04:21,440 which means that gamers can use neural shaders. 59 00:04:21,440 --> 00:04:26,400 Unlike typical shaders, this allows geometry to be imbued with extra properties. 60 00:04:26,400 --> 00:04:30,200 But now that extra property could be a small neural network 61 00:04:30,200 --> 00:04:34,840 that could generate more geometry or help ease ray tracing calculations. 62 00:04:34,840 --> 00:04:40,000 For instance, mega geometry. This one allows for real time generation 63 00:04:40,000 --> 00:04:44,640 of level of detail steps without requiring any normal maps. 64 00:04:44,640 --> 00:04:49,400 Think of like UE5's Nanite, which helps ease jarring LOD change effects 65 00:04:49,400 --> 00:04:52,880 and saves developer time, but with, you know, AI. 66 00:04:52,880 --> 00:04:57,240 To take advantage of these features, NVIDIA loaded the 5090 with the hardware it needs 67 00:04:57,240 --> 00:05:00,880 to accelerate them. It's got fifth gen tensor cores, 68 00:05:00,880 --> 00:05:07,440 which drastically reduces memory usage for simpler AI models that don't require high precision. 69 00:05:07,440 --> 00:05:10,960 As for the non-AI stuff, we get upgraded fourth gen RT cores, 70 00:05:10,960 --> 00:05:16,480 which now double the ray-triangle intersection rate with just 75% of the memory footprint. 71 00:05:16,480 --> 00:05:21,280 And as for the regular old CUDA cores, well, those just don't seem to have changed very much. 72 00:05:21,280 --> 00:05:24,640 So far, the 5090 has managed a best case scenario 73 00:05:24,640 --> 00:05:31,720 of about 30% faster than its predecessor, seemingly entirely thanks to the 33% higher GPU core camp. 74 00:05:31,720 --> 00:05:35,160 This, combined with their reuse of TSMC's 4N process node 75 00:05:35,160 --> 00:05:40,080 from last gen, explains why the new chip is so big and why NVIDIA had to sacrifice some clock speed 76 00:05:40,120 --> 00:05:44,440 to keep their yields, and therefore pricing still attainable to the 1%. 77 00:05:44,440 --> 00:05:49,720 GDDR7, on the other hand, is kind of a big deal. It boasts double the data rate of GDDR6 78 00:05:49,720 --> 00:05:54,400 while using half as much power per bit. This is an enlarged part thanks to the shift 79 00:05:54,400 --> 00:05:59,100 to PAM3 signaling. PAM, short for pulse amplitude modulation, 80 00:05:59,100 --> 00:06:02,520 is akin to how we store data in multi-level cell flash storage. 81 00:06:02,520 --> 00:06:05,800 GDDR6 uses PAM4, meaning that each clock 82 00:06:05,800 --> 00:06:09,520 can be encoded for four different states, rather than just two. 83 00:06:09,520 --> 00:06:15,120 But it came with a big trade-off, the error rate, since the signals are so similar in amplitude 84 00:06:15,120 --> 00:06:18,720 that sometimes they can be hard to tell apart, especially when there's interference. 85 00:06:18,720 --> 00:06:23,240 PAM3 improves the situation by just trying to handle three states instead of four, 86 00:06:23,240 --> 00:06:27,320 giving a little bit more room between each of them. This improves signal integrity, 87 00:06:27,320 --> 00:06:30,480 allowing GDDR7 to run at higher frequency 88 00:06:30,480 --> 00:06:37,040 while consuming less power to make up for the trade-offs. And let's not forget that we finally got 32 gigs of VRAM. 89 00:06:37,040 --> 00:06:42,040 This will be a huge jump for AIDarks, and maybe gamers someday. 90 00:06:42,040 --> 00:06:46,960 But there are some other cool things, like NVIDIA's new ninth-gen N-Vanc hardware video encoders, 91 00:06:46,960 --> 00:06:50,160 which support higher-quality 422 10-bit video. 92 00:06:50,160 --> 00:06:53,320 This, for the right people, is a huge deal, 93 00:06:53,320 --> 00:06:57,960 and might make Blackwell a must-have upgrade. And for the folks out there who own monitors, 94 00:06:57,960 --> 00:07:03,520 hi, Ploof, we finally get a card that can actually take advantage of DP 2.1 UHBR20, 95 00:07:03,520 --> 00:07:07,280 a new DisplayPort standard that can drive 4K 240Hz 96 00:07:07,280 --> 00:07:11,280 without display-stream compression. And all of this while talking to your computer 97 00:07:11,280 --> 00:07:16,940 at PCIe Gen 5. It's 2025, and ray-tracing is no longer an afterthought, 98 00:07:16,940 --> 00:07:20,680 or even a choice in some cases, with the new name of Jones being the first game 99 00:07:20,680 --> 00:07:24,580 to outright require support. So let's talk about it. 100 00:07:24,580 --> 00:07:30,000 For RT testing, we use the highest settings, starting at 1440, and I want to get this out of the way. 101 00:07:30,000 --> 00:07:34,200 AMD does not ray-trace well. Now, in fact, 2 makes for a very playable experience 102 00:07:34,200 --> 00:07:37,600 on the 5090, with 1% lows well above 60 FPS. 103 00:07:37,600 --> 00:07:41,580 Numbers, it can't quite hit yet on the absolutely brutal Blackmith Wukong, 104 00:07:41,580 --> 00:07:45,440 though it is playable, unlike the poor 7900XTX. 105 00:07:45,440 --> 00:07:50,800 Ouch! In Cyberpunk, the 5090 has just a 20% lead over the 4090, 106 00:07:50,800 --> 00:07:54,820 but compared to the 4080 Super, it maintains its price-to-performance ratio, 107 00:07:54,820 --> 00:07:58,480 which I generally consider to be pretty darn acceptable for a Halo-class card. 108 00:07:58,520 --> 00:08:03,320 In the lightly ray-traced F124, AMD comes back to life a little, 109 00:08:03,320 --> 00:08:07,520 performing well against the 4080 Super, and the same can be said for Returnal, 110 00:08:07,520 --> 00:08:12,520 but there's no question that the 5090 is king for RT at 1440P, 111 00:08:13,600 --> 00:08:19,160 with a crown that only gets more dazzling at 4K. Blackmith Wukong falls below what we consider playable 112 00:08:19,160 --> 00:08:23,000 for an intense action game. Don't worry, we'll talk about AI upscaling later, 113 00:08:23,000 --> 00:08:28,320 because first, dang, look at this thing! Maintaining performance in the 50s at these settings, 114 00:08:28,480 --> 00:08:32,160 this economy? Dang, NVIDIA, that's pretty impressive. 115 00:08:32,160 --> 00:08:35,300 And if you care more about absolute cinema than framerate, 116 00:08:35,300 --> 00:08:40,480 well, it holds above 30 FPS in Alan Wake 2, which should go great with your popcorn. 117 00:08:40,480 --> 00:08:44,080 F124 and Returnal are similar stories as the 1440P results, 118 00:08:44,080 --> 00:08:48,880 just with more pixels and fewer FPS. All of this taken together means we're looking at 119 00:08:48,880 --> 00:08:53,600 a greater than 30% lead over last gen at a 25% higher price, 120 00:08:53,600 --> 00:08:57,020 meaning the new RT cores are providing some benefit, 121 00:08:57,020 --> 00:09:01,420 but it's pretty small compared to the impact of NVIDIA just plunking in more of them. 122 00:09:01,420 --> 00:09:04,540 This is obviously a downer compared to the good old days 123 00:09:04,540 --> 00:09:09,140 when we used to get yearly GPU refreshes with dramatic improvements to performance per dollar. 124 00:09:09,140 --> 00:09:13,840 But it's clear that unless cutting-edge semiconductor manufacturing miraculously gets cheaper, 125 00:09:13,840 --> 00:09:18,940 those days are never coming back. So if we compare this more to, say, 126 00:09:18,940 --> 00:09:23,240 adding a second card in SLI, a feature NVIDIA no longer supports, 127 00:09:23,240 --> 00:09:29,540 then the glass half-full-take is, hey, at least it costs less than $240.90s. 128 00:09:29,540 --> 00:09:32,900 But NVIDIA still has some tricks up their sleeve. 129 00:09:32,900 --> 00:09:37,060 Holy heck, that's one tight leather jacket. How the heck did you fit that stuff in those sleeves? 130 00:09:37,060 --> 00:09:41,820 With DLSS4 and multi-frame gen, by making use of flip metering and swapping them 131 00:09:41,820 --> 00:09:46,020 from convolution neural networks to transformer-based models. 132 00:09:46,020 --> 00:09:49,900 There's a lot of words again, lots of words again. Let's break it down. 133 00:09:49,900 --> 00:09:53,980 DLSS4 is NVIDIA's latest suite of AI enhancements, 134 00:09:53,980 --> 00:09:58,260 and it's the biggest change in years. Previous versions of DLSS included 135 00:09:58,260 --> 00:10:01,460 a convolutional neural network or CNN. 136 00:10:01,460 --> 00:10:04,940 A CNN can be thought of as a series of filters 137 00:10:04,940 --> 00:10:08,740 that look for specific details. When used for image processing, 138 00:10:08,740 --> 00:10:15,220 one layer could be looking for vertical edges, one for horizontal edges, and one for contrast, et cetera. 139 00:10:15,220 --> 00:10:19,980 The neural network then observes the results from the filters and can use that information 140 00:10:19,980 --> 00:10:24,340 to identify things. Like if an image contains a dog or a stop sign 141 00:10:24,340 --> 00:10:27,340 and seems convoluted, well, it literally is. 142 00:10:27,340 --> 00:10:30,860 On 4,000 series GPUs, this information was combined 143 00:10:30,860 --> 00:10:34,780 with an optical flow accelerator that interpreted the motion in the scene 144 00:10:34,780 --> 00:10:38,460 to upscale or generate frames. So why the switch? 145 00:10:38,460 --> 00:10:41,940 Scaling. Each filter can only scan and compute 146 00:10:41,940 --> 00:10:46,340 a small number of pixels at a time. When you have millions of pixels, 147 00:10:46,340 --> 00:10:50,140 dozens of times per second, increasing performance can be tough. 148 00:10:50,140 --> 00:10:54,940 So DLSS uses a new transformer model, which, as NVIDIA explains, 149 00:10:54,940 --> 00:11:00,140 allows them to evaluate the relative importance of each pixel across an entire frame 150 00:11:00,140 --> 00:11:04,580 and over multiple frames to achieve a deeper understanding of the scenes 151 00:11:04,580 --> 00:11:08,580 that offers greater stability, reduced ghosting, higher detail in motion, 152 00:11:08,580 --> 00:11:11,780 and smoother edges. They also scale better, 153 00:11:11,780 --> 00:11:16,660 which is part of why they have become so heavily used in things like large language models. 154 00:11:16,660 --> 00:11:22,500 The transformer is the T in chat GPT, so while a CNN can see this picture and say, 155 00:11:22,500 --> 00:11:25,820 there is a cat, there is a product from LTTstore.com. 156 00:11:25,820 --> 00:11:31,340 A transformer might say, there is a cat enjoying the premium CRT-themed peck cave 157 00:11:31,340 --> 00:11:36,380 from LTTstore.com. However, while a transformer can process 158 00:11:36,380 --> 00:11:40,340 complete images faster, it does require more data for training, 159 00:11:40,340 --> 00:11:46,100 and honestly, with the side-by-side comparisons, it is tough to tell the difference in image quality 160 00:11:46,100 --> 00:11:49,500 between the two models, like really tough. 161 00:11:49,500 --> 00:11:53,540 There are clear benefits in specific areas that NVIDIA points out, 162 00:11:53,540 --> 00:11:56,580 things like fences, power lines, and barbed wires, 163 00:11:56,580 --> 00:12:01,620 but there's still obvious artifacts when dealing with semi-transparent objects, 164 00:12:01,620 --> 00:12:07,300 or just very busy scenes. A lot of the artifacts are different than DLSS3, 165 00:12:07,300 --> 00:12:11,660 but are still present. On the bright side, at least on high-end cards, 166 00:12:11,660 --> 00:12:15,900 the transformer models don't show a substantial performance hit. 167 00:12:15,900 --> 00:12:19,300 Enough theory, I wanna talk about MF-ing G. 168 00:12:19,300 --> 00:12:23,580 Multi-frame Gen is perhaps the most game-changing tech landing with these new cards. 169 00:12:23,580 --> 00:12:27,940 Like the previous version of Framegen, it uses AI to generate in-between frames 170 00:12:27,940 --> 00:12:32,700 based on optical flow data and rendered frames, but Multi-frame Gen now allows users 171 00:12:32,700 --> 00:12:35,980 to generate up to three in-betweens, rather than just one, 172 00:12:35,980 --> 00:12:39,300 boosting frame rates to up to four times native. 173 00:12:39,300 --> 00:12:43,980 Does it work? Well, according to our charts, yes, very yes. 174 00:12:43,980 --> 00:12:49,420 The numbers double, triple, and quadruple, and make the 5090 look absolutely ridiculous, 175 00:12:49,420 --> 00:12:52,620 at least in the charts. But as big as that bar is, 176 00:12:52,620 --> 00:12:56,420 the real frames haven't changed. So what's the deal? 177 00:12:56,420 --> 00:13:01,540 Well, MFG's pretty wild. DLSS3 Framegen required specific optical flow 178 00:13:01,540 --> 00:13:05,380 accelerating hardware on GPUs, and combined that with the game data, 179 00:13:05,380 --> 00:13:09,660 like depth and motion vectors to generate in-between frames. And it was an okay solution, 180 00:13:09,660 --> 00:13:13,900 but you had to have two bits of hardware processing each frame, 181 00:13:13,900 --> 00:13:17,520 and that's just inefficient, and could even cause the GPU to throttle, 182 00:13:17,520 --> 00:13:20,620 resulting in a lower base frame rate to multiply off of. 183 00:13:20,660 --> 00:13:24,380 That's why you didn't just see a straight doubling of frame rate when you turned Framegen on. 184 00:13:24,380 --> 00:13:28,500 The 5090 and multi-FrameGen is Shu Aida's optical flow accelerator, 185 00:13:28,500 --> 00:13:32,620 and instead utilized tightly integrated Tensor and CUDA cores in Blackwell 186 00:13:32,620 --> 00:13:35,660 to run a lightweight AI optical flow model. 187 00:13:35,660 --> 00:13:39,980 Not hardware accelerated, it's just an AI model. This means that single Framegen 188 00:13:39,980 --> 00:13:43,980 should now run 40% faster while using 30% less VRAM. 189 00:13:43,980 --> 00:13:48,700 We know it works, but how does it look? Well, it depends on who you ask. 190 00:13:48,700 --> 00:13:51,820 If you want to see your FPS number much higher then it works great. 191 00:13:51,820 --> 00:13:56,900 Nothing else sort of hacking your FPS counter will let you get nearly 600 FPS in Cyberpunk. 192 00:13:56,900 --> 00:14:00,860 But visually, it's not perfect, and Framegen weirdness still persists. 193 00:14:00,860 --> 00:14:05,260 Look at the combing on these crosswalks in Cyberpunk, or this bottle phasing in and out in the benchmark, 194 00:14:05,260 --> 00:14:09,140 or the doubling of the fan blades in these large HVAC units. In Alan Wake 2, 195 00:14:09,140 --> 00:14:12,900 there's obvious artifacting around the player model and around the edge of your flashlight, 196 00:14:12,900 --> 00:14:17,500 which is sadly exactly where you will be looking 100% of the time. 197 00:14:17,500 --> 00:14:22,100 And it's worth noting that these artifacts are not present when we're just using DLSS for upscaling. 198 00:14:22,100 --> 00:14:26,780 Curiously, while Cyberpunk and Alan Wake were both updated with explicit support for multi Framegen, 199 00:14:26,780 --> 00:14:30,780 the feature can be forced in any DLSS3 single Framegen 200 00:14:30,780 --> 00:14:36,340 supported games through the NVIDIA driver. And the game that fared the best was Dragon Age of Veilguard. 201 00:14:36,340 --> 00:14:39,740 The world's full of magic, so who's to say exactly what a vortex of shadow 202 00:14:39,740 --> 00:14:44,220 is supposed to look like? And since it already ran at a solid 70 FPS 203 00:14:44,220 --> 00:14:48,820 with all the settings cranked, no Framegen, it kept input latency manageable, 204 00:14:48,820 --> 00:14:52,860 which is really important, because if your base FPS is only 30 frames, 205 00:14:52,860 --> 00:14:57,220 well, Framegen will make it look smooth, but you'll observe many visual anomalies, 206 00:14:57,220 --> 00:15:01,060 and latency is still dictated by your true frame rate, 207 00:15:01,060 --> 00:15:05,460 meaning that the game feels far less responsive than it looks like it should be. 208 00:15:05,460 --> 00:15:09,460 The good news, says NVIDIA, is that MFG at least isn't adding any latency 209 00:15:09,460 --> 00:15:13,500 compared to the base frame rate, but we felt like it would be a good idea to verify that. 210 00:15:13,500 --> 00:15:17,700 And verify we did, using our trusty LDAT, our click to photon test results 211 00:15:17,700 --> 00:15:22,260 showed that Framegen does not increase latency over native with reflex on. 212 00:15:22,260 --> 00:15:25,300 In fact, it seems to actually lower the latency slightly. 213 00:15:25,300 --> 00:15:28,420 It doesn't really make sense to us, so we're gonna chalk that up to sampling error, 214 00:15:28,420 --> 00:15:32,580 but if there is any effect on latency, it's so minor that it's not noticeable 215 00:15:32,580 --> 00:15:36,180 compared to our total system latency, and that is mighty impressive. 216 00:15:36,180 --> 00:15:40,260 But the issue remains, you can't beat your base frame rate's latency, 217 00:15:40,260 --> 00:15:44,860 and when it's disconnected from the motion you see on screen, it almost makes it worse. 218 00:15:44,860 --> 00:15:48,340 Perhaps the situation will improve with reflex too, but that's not here yet, 219 00:15:48,340 --> 00:15:51,500 so we aren't gonna dwell on it, even if the tech is really cool. 220 00:15:51,500 --> 00:15:55,620 Our take on multi-frame gen right now is that it has the same key flaws before. 221 00:15:55,620 --> 00:15:58,740 It's a win more feature that works the best 222 00:15:58,740 --> 00:16:03,420 when it makes the least sense to use, which means it is definitely not the silver bullet 223 00:16:03,420 --> 00:16:06,940 that NVIDIA's graphs make it out to be. With all this power, 224 00:16:06,940 --> 00:16:11,820 it would make a lot more sense to use the 5090 to make some money, right? So let's talk productivity. 225 00:16:11,820 --> 00:16:16,340 Hey, you might not know me, but I do this kind of thing and this kind of thing. 226 00:16:16,340 --> 00:16:19,340 And NVIDIA's new architecture has benefits for me too. 227 00:16:19,340 --> 00:16:22,660 The encoder and decoders provide support for 422 chroma sampling, 228 00:16:22,660 --> 00:16:27,740 which will make working with high-end video files much faster, especially for multi-camera video edits. 229 00:16:27,740 --> 00:16:31,220 The encoders also provide better quality at smaller file sizes. 230 00:16:31,220 --> 00:16:36,460 Sadly, we can't verify that for you today as we are currently re-evaluating our encoding benchmarks, 231 00:16:36,460 --> 00:16:40,220 but the new media engine is almost certainly playing a role in Puget Bench, 232 00:16:40,220 --> 00:16:43,540 where we see a nice 9% bump in Premiere Pro performance 233 00:16:43,540 --> 00:16:47,500 and an even nicer, nearly 20% improvement DaVinci Resolve 234 00:16:47,500 --> 00:16:51,140 when compared to the 4090. In Blender, NVIDIA has us considering 235 00:16:51,140 --> 00:16:55,140 finding a new benchmark as the 5090 has brought Barbershop render times 236 00:16:55,140 --> 00:16:59,100 to less than half a minute, more than double the speed of the 3090 Ti. 237 00:16:59,100 --> 00:17:03,820 Nice. Overwrought editing transition here. Double the 3090 Ti, you say? 238 00:17:03,820 --> 00:17:08,900 AI nerds, rejoice! If you're like me and for your sake, I hope you're not, 239 00:17:08,900 --> 00:17:12,380 you've been dying for NVIDIA to release a new 32-gig consumer card. 240 00:17:12,380 --> 00:17:17,700 So with all the bragging NVIDIA's been doing about AI tops, I'm expecting some big numbers. 241 00:17:17,700 --> 00:17:22,340 And in the Procyon text benchmarks, what the fuck? Number is not big. 242 00:17:22,340 --> 00:17:27,260 Sure, the 5090 is still the best card on the charts, but I was expecting more than this. 243 00:17:27,260 --> 00:17:30,620 We'll see roughly 20 to 30% improvement over the 4090, 244 00:17:30,620 --> 00:17:35,180 depending on the benchmark, and 60 to 70% over the 3090 Ti. 245 00:17:35,180 --> 00:17:40,060 I can see why they keep talking about AI tops and not specific performance. 246 00:17:40,060 --> 00:17:44,580 In ML Perf, the story remains largely the same in the time to first token 247 00:17:44,580 --> 00:17:50,020 and the token generation rate benchmarks. For image generation, our preferred Procyon benchmark 248 00:17:50,020 --> 00:17:54,620 doesn't support the 5090 yet. So we tested using the benchmark provided by NVIDIA. 249 00:17:54,620 --> 00:18:00,260 Prepare your salt grains for the taking. In the Procyon Flux FP8 image generation, 250 00:18:00,260 --> 00:18:03,780 the 5090 leads by a margin in line with the rest of our benchmarks. 251 00:18:03,780 --> 00:18:08,500 But when we switch to FP4 precision, the 5090 shows how powerful the native hardware support 252 00:18:08,500 --> 00:18:13,420 can be, taking less than one quarter of the time to generate images compared to the 4090. 253 00:18:13,420 --> 00:18:17,900 We'd love to see how this fairs against the 3090 Ti, but this NVIDIA provided benchmark 254 00:18:17,900 --> 00:18:21,440 doesn't support older cards. I was expecting AI to be the place 255 00:18:21,440 --> 00:18:24,620 where this card really shines, but I guess potential buyers will have to settle 256 00:18:24,620 --> 00:18:29,140 for just having the best consumer friendly grade card for AI. 257 00:18:29,140 --> 00:18:34,460 We'll call it the nifty feinty. Somehow there is still more review to go. 258 00:18:34,460 --> 00:18:39,660 Good thing this is so digestible and uncomplicated so far, right? Blackwell's main efficiency improvements over ADA 259 00:18:39,660 --> 00:18:42,660 seem to come from what they're calling Max-Q functionality. 260 00:18:42,660 --> 00:18:46,220 It boils down to a few small but significant changes. 261 00:18:46,220 --> 00:18:50,360 Improvements to power gating, thanks to an additional power rail and improved logic 262 00:18:50,360 --> 00:18:54,080 allows more of the GPU to switch to a low power state more rapidly. 263 00:18:54,080 --> 00:18:57,660 In CPU bound or frame cap scenarios, this could help save some power, 264 00:18:57,660 --> 00:19:01,540 especially on mobile chips. But in GPU bound full load scenarios, 265 00:19:01,580 --> 00:19:06,520 this monster will absolutely draw its fully rated 575 watts, 266 00:19:06,520 --> 00:19:10,940 including transient spikes of as high as 637 watts. 267 00:19:10,940 --> 00:19:14,900 And you can see even in real world gaming, it will pull space heater levels of power 268 00:19:14,900 --> 00:19:18,620 with an average of 554 watts in F124. 269 00:19:18,620 --> 00:19:22,460 And to manage all that power, they had to make a unique cooler design. 270 00:19:22,460 --> 00:19:26,100 While everyone else was making four slot behemoths, NVIDIA built something very different. 271 00:19:26,100 --> 00:19:30,660 And man, does it look good. Not only is it classy, it's also innovative. 272 00:19:30,700 --> 00:19:34,700 Over ambitious even. The main board of the video card is just the middle section 273 00:19:34,700 --> 00:19:39,500 and the outputs in PCIe connector are on daughter boards connected via what they call a flexible PCB. 274 00:19:39,500 --> 00:19:44,220 It's like a stiff ribbon cable, I guess. This allows for a double flow through design 275 00:19:44,220 --> 00:19:50,380 with fans blowing through dense heat sinks. Anecdotally, the fans run quiet, too quiet even. 276 00:19:50,380 --> 00:19:54,660 So like the 4090, most of the time, the loudest thing about the card will be its coil wine. 277 00:19:54,660 --> 00:20:00,420 It's not the worst we've heard, but it's noticeable. And if you expected it to run cooler than the 4090 founders, 278 00:20:00,420 --> 00:20:05,580 it doesn't. But given how much smaller it is, not to mention the enormous thermal load it's dealing with, 279 00:20:05,580 --> 00:20:10,060 I'd say it's doing a great job. But the new cooler style brings new build considerations. 280 00:20:10,060 --> 00:20:13,540 We know that flow through coolers can have a noticeable impact on CPU temps, 281 00:20:13,540 --> 00:20:17,300 especially for those using tower heat sinks. So as a double flow through, double bad, 282 00:20:17,300 --> 00:20:21,340 we took the 5090 and the 4090 FE and put them in a Corsair 4000D. 283 00:20:21,340 --> 00:20:24,500 And even with both running at 450 Watts to control the experiment, 284 00:20:24,500 --> 00:20:29,100 our poor Noctua NHD15 saw CPU temps that are roughly three degrees higher 285 00:20:29,140 --> 00:20:34,300 in both synthetic and gaming workloads. A CPU cooler upgrade, perhaps an intake-mounted radiator, 286 00:20:34,300 --> 00:20:38,220 could be in order for some folks. Wow, that was a lot to talk about. 287 00:20:38,220 --> 00:20:41,500 So what's our conclusion? Well, if you're a professional game developer 288 00:20:41,500 --> 00:20:45,780 or any other professional, or basically anyone who can use the new performance 289 00:20:45,780 --> 00:20:48,820 and especially the new features to make money, 290 00:20:48,820 --> 00:20:52,900 it's a no-brainer. As for the gamers, well, if the 4090 was stupid, 291 00:20:52,900 --> 00:20:57,540 stupid price, but stupid performance, the 5090 is stupider. 292 00:20:57,540 --> 00:21:03,660 It provides 30-ish percent more performance by using 33% more hardware and 30-ish percent more power, 293 00:21:03,660 --> 00:21:08,980 and it gets a roughly 25% price increase. It's a 4090 plus plus. 294 00:21:08,980 --> 00:21:13,060 On the one hand, you could look at this and say, wow, this doesn't look good 295 00:21:13,060 --> 00:21:16,380 for the rest of the 50 series lineup, but it's also worth considering 296 00:21:16,380 --> 00:21:19,380 that the 4090 was kind of an outlier for 40 series, 297 00:21:19,380 --> 00:21:24,340 offering a huge boost over its predecessor with the rest offering smaller upgrades. 298 00:21:24,340 --> 00:21:28,820 So if you're on 30 series and still not in the mega, mega baller income bracket, 299 00:21:28,820 --> 00:21:34,180 I guess all we can do is wait and see. See if Lydus has the strength to do the segue. 300 00:21:34,180 --> 00:21:39,180 Who are you? If you like this video, 301 00:21:39,180 --> 00:21:43,540 check out the one you get on ACCS chronging, 302 00:21:43,540 --> 00:21:45,740 watch this good nugget for now. 303 00:21:46,980 --> 00:21:49,660 I wish I could eat a nugget right now.