1 00:00:00,000 --> 00:00:04,560 The highlight of the NVIDIA keynote for me was when CEO Jensen Huang quipped that 2 00:00:04,560 --> 00:00:10,000 if you bought an RTX 4090, that was probably the best investment you could make. 3 00:00:10,000 --> 00:00:14,320 I mean, just consider all the entertainment happiness that it could give you. 4 00:00:14,320 --> 00:00:21,600 But one small thing, Mr. Huang, when NVIDIA launched the RTX 4090, your stock price was 5 00:00:21,600 --> 00:00:28,480 about $1150. So if someone had bought that instead of a 4090 as their investment, well, 6 00:00:28,480 --> 00:00:34,560 hey, good news, because they should be able to comfortably afford a 5090 or 10. 7 00:00:35,280 --> 00:00:39,920 As for everybody else, well, unfortunately NVIDIA doesn't take happiness, 8 00:00:39,920 --> 00:00:44,880 so you're going to be trading a truly princely sum for the privilege of experiencing their latest 9 00:00:44,880 --> 00:00:51,040 and greatest. But, affordability aside, this was a heck of an event. NVIDIA kicked it off with the 10 00:00:51,040 --> 00:00:56,320 most beautiful real-time render demo that I had ever seen and then followed it up with a slew 11 00:00:56,400 --> 00:01:02,240 of new RTX Blackwell 50 Series GPUs with brand new technologies like DLSS 4.0, 12 00:01:02,240 --> 00:01:09,360 DLSS Transformer, Neural Faces, and some of the GPUs even have prices that, taken at face value, 13 00:01:09,360 --> 00:01:17,680 might actually be kind of exciting. Jensen claimed on stage that the 5070 will offer 14 00:01:17,680 --> 00:01:24,480 4090 levels of performance at $549. Then immediately followed that up with a huge red 15 00:01:24,480 --> 00:01:31,120 flag saying that this feat would be impossible without AI. So I was all like, sick, and then I 16 00:01:31,120 --> 00:01:36,240 was all like, I wouldn't want to see it. Well, no sooner did I say that than some burly guys in 17 00:01:36,240 --> 00:01:40,320 green showed up and said, all right, you cynical tech-tipping son of a bitch, we're going to show 18 00:01:40,320 --> 00:01:45,520 it to you right now, but you've got to do the thing. The thing, I said. Yeah, the thing. 19 00:01:46,480 --> 00:01:49,440 All right, the thing, the segue to our sponsor. 20 00:01:59,120 --> 00:02:05,120 We'll get to some of the AI and data center stuff from the keynote, but first I want to talk G-Force, 21 00:02:05,120 --> 00:02:11,200 and right now we are behind the scenes getting the first ever hands-on with the RTX 5090. 22 00:02:11,280 --> 00:02:15,440 Well, we aren't yet. Now we are. They didn't tell me I could touch it, 23 00:02:15,440 --> 00:02:20,480 but they can't stop me. They're way over there. Before we go deeper, there are a couple of things 24 00:02:20,480 --> 00:02:25,040 that stood out immediately from their sizzle video for this thing. I mean, yeah, there's the 25 00:02:25,040 --> 00:02:30,960 incredibly compact PCB with the state-of-the-art power delivery that enables a sick dual fan, 26 00:02:30,960 --> 00:02:37,520 dual slot cooler with pass through for both of the fans. That's right, both of these fans are 27 00:02:37,520 --> 00:02:45,120 blowing right through the cart, but this, thank you. No more need for fire-prone right-angle 28 00:02:45,120 --> 00:02:50,000 adapters to clean up your cable management. An angled power connector. Okay, okay, Linus, 29 00:02:50,000 --> 00:02:55,680 it has a nice cable, but what about the specs? Oh, I've got specs. Let me put it this way. 30 00:02:55,680 --> 00:03:01,120 If you have a 4090 right now, the 5090 is the GPU that she tells you not to worry about. 31 00:03:01,120 --> 00:03:07,120 In raw numbers, here's what that looks like. This is going to be one heck of an upgrade. 32 00:03:07,120 --> 00:03:14,720 I can't wait to see it in action. So why am I waiting? Let's do it. Here, relegated to the floor, 33 00:03:14,720 --> 00:03:20,400 like a common chicken, is the previous king, the RTX 4090. Why don't we start there? 34 00:03:20,400 --> 00:03:25,120 NVIDIA can't let me show you guys much prior to the performance embargo in a couple of weeks. 35 00:03:25,120 --> 00:03:32,080 We get one game, and only one game, and we can't change the settings. But I can't think of anything 36 00:03:32,080 --> 00:03:38,880 I'd want to change here. Absolutely everything is cranked. And to their credit, they did load up 37 00:03:38,880 --> 00:03:45,680 frame view, so I won't be purely relying on feel to evaluate the difference. Okay, so let's play 38 00:03:45,680 --> 00:03:53,600 on the 4090. Yep, that's a pretty powerful GPU still. And it's even more powerful than you 39 00:03:53,600 --> 00:03:59,440 probably realize. Before we move over to the 5090, this build of Cyberpunk has some new goodness 40 00:03:59,440 --> 00:04:05,840 baked into it, including NVIDIA's new transformer model and DLSS4 Framegen. So the results we're 41 00:04:05,840 --> 00:04:11,200 getting, even on the 4090 here, are not going to be comparable to your system at home. This guy 42 00:04:11,200 --> 00:04:16,960 is performing a little better thanks to some of these changes. It's nice that it's getting the 43 00:04:16,960 --> 00:04:25,680 extra help because this is not close. Even without looking at the numbers, which are double, 44 00:04:27,040 --> 00:04:32,800 this is wickedly smooth. I still remember, I think it was on a ShortCircuit or something, 45 00:04:32,800 --> 00:04:37,600 we were looking at a system from, I want to say main gear. It was my first time playing 46 00:04:37,600 --> 00:04:42,880 Cyberpunk on the 4090 because I hosted the review and everything, but I never actually got to game 47 00:04:42,880 --> 00:04:51,360 on it. And I was like, oh my God. Yeah, this is a lot faster. And it's like, where are the polygons? 48 00:04:51,360 --> 00:05:00,160 You know, where have my polys gone? Sure, the thing is priced at $2,000, a 25% increase, 49 00:05:00,160 --> 00:05:07,680 and is rated at 575 watts. But for the first time in what feels like forever, at least for 50 00:05:07,680 --> 00:05:14,000 getting more for the money. This feels like a truly halo tier product, the 512-bit memory bus, 51 00:05:14,000 --> 00:05:20,240 5,000 more CUDA cores, and noticeably better performance. I can't believe that this is 52 00:05:20,240 --> 00:05:28,080 Cyberpunk running at 4K with ray tracing. I can't believe it because it isn't. FrameGen and AI-assisted 53 00:05:28,080 --> 00:05:33,280 ray tracing are doing a lot of heavy lifting here. Jensen gave us a bit of insight into this, 54 00:05:33,280 --> 00:05:39,120 telling us that out of the about 33 million pixels in his demo, only 2 million of them were 55 00:05:39,120 --> 00:05:45,440 calculated with the rest extrapolated using AI. But okay, I mean, it doesn't matter as long as it 56 00:05:45,440 --> 00:05:51,840 looks good, right? So how good does it look? Well, I'm supposed to be just showing this off. 57 00:05:51,840 --> 00:05:55,360 What I've actually been doing is looking around for a couple of interesting examples 58 00:05:55,360 --> 00:06:02,480 of how DLSS4 really doesn't handle everything perfectly. See the way this text kind of jitters 59 00:06:03,280 --> 00:06:08,480 and blurs if I move around? It's not as clear as the rest of the frame. It's okay if I strafe. 60 00:06:08,480 --> 00:06:14,080 And I've noticed some instances of particularly bright or high contrast objects having little 61 00:06:14,080 --> 00:06:19,600 ghosts on them. With that said, I know what I'm looking for. Other reviewers are going to know 62 00:06:19,600 --> 00:06:27,040 what they're looking for. If I were to put one of my siblings in front of this, it looks pretty good. 63 00:06:28,000 --> 00:06:32,320 And NVIDIA was quick to point out that this is both running at lower latency 64 00:06:32,320 --> 00:06:38,560 and with fewer artifacts than DLSS 3.5. Enough about the 5090 that, let's face it, 65 00:06:38,560 --> 00:06:42,320 most people aren't going to buy. Let's talk about the rest of the family. 66 00:06:42,320 --> 00:06:47,520 The RTX 5070 comparison to the 4090 should be taken with a grain of salt. I mean, 67 00:06:47,520 --> 00:06:53,280 obviously that's going to be with DLSS version 4. Obviously that's going to be using their new 68 00:06:53,360 --> 00:06:59,760 MFG or multi-frame generation at 4x. That feature is limited to 50 series cards and 69 00:06:59,760 --> 00:07:05,520 is powered by AI. And obviously the situation might change depending on the resolution you're 70 00:07:05,520 --> 00:07:12,080 running. I mean 4K needs a fair bit of VRAM. So can a few hundred more CUDA cores net such a 71 00:07:12,080 --> 00:07:18,480 wild performance boost, especially with so much less VRAM? It's a bold claim. But with this mix of 72 00:07:18,480 --> 00:07:24,240 next-gen CUDA, Tensor and RG cores, along with DDR7, it could very well achieve it, at least in 73 00:07:24,240 --> 00:07:29,600 some titles. And if it looks anywhere near as good as this does, while costing less than half 74 00:07:29,600 --> 00:07:36,720 of what the 4090 did at launch, well, I certainly wish that it had 16 gigs of VRAM, but even without 75 00:07:36,720 --> 00:07:45,040 it, they are going to sell a lot of these. Not to me, you might say. I want 16 gigs of VRAM. Well, 76 00:07:45,920 --> 00:07:52,320 then you are going to need to take another step up to the 5070 Ti or another another step to the 77 00:07:52,320 --> 00:08:00,320 RTX 5080. And about these. If prices had stayed the same, I'd probably be pretty pissed off right 78 00:08:00,320 --> 00:08:06,800 now. But the two 70-class cards they showed today are $50 less than last-gen. And I don't know if 79 00:08:06,800 --> 00:08:15,600 that's because of competition, but hey, progress. And it actually gets better. The RTX 5080, with all 80 00:08:15,600 --> 00:08:20,880 the other bells and whistles that we've mentioned so far, and sporting an additional 1000 CUDA cores 81 00:08:20,880 --> 00:08:27,600 compared to its predecessor, is set at just $999. I say just. I know a thousand dollars is still 82 00:08:27,600 --> 00:08:35,360 really expensive, but that is down $200 from the RTX 4080 or about 16%. It's got next-gen 83 00:08:35,360 --> 00:08:40,400 everything, a thousand more CUDA cores, more than double the AI tops, and more memory bandwidth, 84 00:08:40,400 --> 00:08:44,480 meaning it should deliver legitimately flagship-class performance at that price, 85 00:08:45,120 --> 00:08:50,080 even if there does exist an even higher end card for the Mega Ballers out there. 86 00:08:50,640 --> 00:08:55,920 Right alongside their new desktop GPUs, NVIDIA announced 50 series for laptops as well. They 87 00:08:55,920 --> 00:09:03,120 start as low as $1299 with the RTX 5070, and they are claiming 4090 performance using half of the 88 00:09:03,120 --> 00:09:11,600 power. And then the lineup goes all the way up to an RTX 5090 at $2899. By the way, if you want to see 89 00:09:11,600 --> 00:09:17,680 our quick first initial reactions to MSI's 50 series cards when we were under embargo, and to 90 00:09:17,680 --> 00:09:21,520 watch us print Colton, why don't you check out our behind-the-scenes video over at LMG.gg 91 00:09:21,520 --> 00:09:27,440 slash flowplain. But wait, there's more. Jensen brought a wafer prop onto the stage to illustrate 92 00:09:27,440 --> 00:09:33,840 the incredible technology that's built into their NB72 AI racks. He described it as equal to 20 cars. 93 00:09:34,640 --> 00:09:40,080 I'm not 100% sure if he meant in terms of complexity with the 600,000 parts or price. 94 00:09:41,600 --> 00:09:46,720 Either way, each one draws 120,000 watts, contains 130 trillion transistors, 95 00:09:46,720 --> 00:09:55,920 and has 1.4 XFLOPS FP4. But do what with it? Well, I'm not sure yet, but it's exciting, 96 00:09:56,000 --> 00:10:01,440 and he wasn't done. Another big announcement is Cosmos, a world foundation model that's designed 97 00:10:01,440 --> 00:10:06,800 and created to understand the physical world. There are three levels of models, and they're all 98 00:10:06,800 --> 00:10:12,560 openly licensed and available on GitHub, Nano, Super, and Ultra. They're saying this is kind of a chat 99 00:10:12,560 --> 00:10:18,880 GPT moment for robots to help them interact with the real world, which is super cool and also kind 100 00:10:18,880 --> 00:10:23,760 of scary, especially when they brought the army of robots onto the stage. That was cool. Anyway, 101 00:10:23,760 --> 00:10:28,320 the point is you should be able to give a robot a prompt and then just watch it do the thing in 102 00:10:28,320 --> 00:10:34,560 the virtual world so that it can train itself virtually. They did it. The Matrix. But it's 103 00:10:34,560 --> 00:10:40,480 the humans bamboozling the robots. Oh, dodged a bullet there. Anyway, a big part of the reason 104 00:10:40,480 --> 00:10:43,920 for this is that when you're training something like a car, you can just do it in the real world. 105 00:10:43,920 --> 00:10:48,400 We're driving around with cameras all the time. We've got lots of data. But when it comes to 106 00:10:48,400 --> 00:10:52,960 humanoid robots, it's a lot more challenging because they aren't just running around doing 107 00:10:52,960 --> 00:10:57,040 stuff because if they were, it'd be extremely dangerous. So the idea with Cosmos is to take 108 00:10:57,040 --> 00:11:02,880 the wear and tear out of the equation, which will hopefully mean that costs can go way down in robotics 109 00:11:02,880 --> 00:11:08,080 development on a physical level. Just exactly what this is going to mean in the long term for everyone, 110 00:11:08,080 --> 00:11:11,760 including potentially a lot of warehouse workers that aren't going to need to move things around 111 00:11:11,760 --> 00:11:19,360 anymore. Well, I guess we'll see. But wait, there's even more. NVIDIA has announced, you guessed it, 112 00:11:19,360 --> 00:11:23,280 a ton of partnerships around robotics and machine vision promising to take the training that's 113 00:11:23,280 --> 00:11:28,400 happening in Cosmos and apply that to the real world. That was coupled with their new Thor AV 114 00:11:28,400 --> 00:11:33,920 computer, a blackwell based robotics processor that also has their grace ARM CPU cores on it that 115 00:11:33,920 --> 00:11:39,360 can handle a ton of visual data at once like cameras, radar and LiDAR. They're boasting 20 times the 116 00:11:39,360 --> 00:11:44,240 processing power of their previous generation, which is absolutely wild. And last but not least, 117 00:11:44,240 --> 00:11:49,200 they included a couple allusions to the multi trillion dollar industries that they're going 118 00:11:49,200 --> 00:11:54,240 to be addressing with all of this technology. So well, will NVIDIA stock price or even higher 119 00:11:54,240 --> 00:12:00,800 thanks to our new robot overlords? Again, I'm not sure because as NVIDIA carefully noted, 120 00:12:00,800 --> 00:12:06,320 all of these are forward looking statements and can't be taken as pure fact. Finally, last but not 121 00:12:06,320 --> 00:12:12,080 least, they surprised everyone, including Nintendo by announcing the Switch 2. I'm kidding. I'm 122 00:12:12,080 --> 00:12:17,680 kidding. I know that didn't happen. Anyway, that's it. The nearly two hour keynote ended and I had 123 00:12:17,680 --> 00:12:23,120 just enough time after trying the RTX 5090 to give you this segue to our sponsor. If you guys 124 00:12:23,120 --> 00:12:27,920 enjoyed this video, thank you for watching. Check out the rest of our CES coverage. Maybe the AMD 125 00:12:27,920 --> 00:12:31,040 announcement? Gotta go.