1 00:00:00,000 --> 00:00:03,760 Linus, wake up. The radio 9070 announced would just drop, come on! 2 00:00:03,760 --> 00:00:10,400 That's right! AMD finally announced those GPUs that were first spotted in full retail packaging 3 00:00:10,400 --> 00:00:16,160 weeks ago, and I can't believe I'm saying this, but it's even more exciting than I dared to hope for. 4 00:00:16,160 --> 00:00:23,360 Now, we had some idea what was coming, of course. RDNA 3.5 showed solid uplift over RDNA 3 in handhelds. 5 00:00:23,360 --> 00:00:27,600 We got a look at some version of the ray tracing upgrades in the PlayStation 5 Pro, 6 00:00:27,680 --> 00:00:30,800 and obviously, AI was going to be a big focus. 7 00:00:30,800 --> 00:00:35,200 But now, it's all coming together in a single package. 8 00:00:35,200 --> 00:00:41,680 Literally. They're back to using a monolithic die. I'm talking all new RDNA 4 architecture, an enhanced media engine, 9 00:00:41,680 --> 00:00:44,800 third-gen ray tracing accelerators, second-gen AI accelerators, 10 00:00:44,800 --> 00:00:51,280 powering FSR4 upscaling, and 16 gigs of VRAM at 599 US dollars. 11 00:00:51,280 --> 00:00:57,040 And the best part? If they've been shipping these things for months, maybe stock? 12 00:00:57,680 --> 00:01:02,800 Perhaps? I'm almost afraid to jinx it. But for the first time in as long as I can remember, 13 00:01:02,800 --> 00:01:06,400 a Hymdi might not have screwed up a GPU launch! 14 00:01:07,200 --> 00:01:08,720 Like, I'm going to screw up this, uh, 15 00:01:13,280 --> 00:01:30,800 to our Spoonster? Feels like we've been waiting forever for RDNA 4, but good things come to those who wait, right? 16 00:01:30,800 --> 00:01:34,000 And personally, I'm glad that AMD let their engineers cook. 17 00:01:34,000 --> 00:01:41,920 We're going to dive a lot deeper when our review goes live next week. But for now, AMD is promising some big uplifts in performance per clock, 18 00:01:41,920 --> 00:01:48,080 claiming that each CU or compute unit should be about 40% faster than RDNA 3. 19 00:01:48,080 --> 00:01:52,240 For those of you who are keeping score at home, 40% is a lot, 20 00:01:52,240 --> 00:01:57,440 and it supposedly comes courtesy of improved scalar units and dynamic register allocation. 21 00:01:57,440 --> 00:02:01,200 They've also got an enhanced memory subsystem, which will supposedly 22 00:02:01,200 --> 00:02:09,600 help with the compute portions of ray tracing workloads. Now, it's too early to say if AMD has fully closed the gap with NVIDIA when it comes to ray tracing, 23 00:02:09,600 --> 00:02:14,560 but it seems like they're at least promising a major improvement over their last generation 24 00:02:14,560 --> 00:02:19,360 accelerators. On top of the memory stuff, they've added a second ray intersection engine to 25 00:02:19,360 --> 00:02:23,040 allegedly double throughput for raybox and ray triangle testing. 26 00:02:23,040 --> 00:02:27,760 There's a dedicated hardware block for ray transformation and a new hardware accelerator 27 00:02:27,760 --> 00:02:33,440 feature that they call oriented bounding boxes. That's a lot of jargon, but what it pretty much 28 00:02:33,440 --> 00:02:39,440 boils down to is this thing should trace your rays a lot better, and I'm pretty excited about it. 29 00:02:39,520 --> 00:02:43,520 Even if the real world result doesn't quite match up with NVIDIA's latest. 30 00:02:43,520 --> 00:02:46,400 I'm also excited about AMD's new AI accelerators. 31 00:02:47,120 --> 00:02:51,520 Okay, I'm not that excited about it, but someone is. Maybe even you. 32 00:02:52,400 --> 00:02:58,880 With only 16 gigs of VRAM apiece, the two cards that AMD revealed today won't be able to handle 33 00:02:58,880 --> 00:03:03,920 larger models, but the claimed performance improvements for the models that they can handle 34 00:03:03,920 --> 00:03:09,280 are pretty big. Thanks to more math pipelines and expanding the capabilities of their AI 35 00:03:09,280 --> 00:03:15,280 accelerators, AMD is claiming a two times performance uplift in FP16 and four times in 36 00:03:15,280 --> 00:03:20,080 Int8 compared to RDNA3. As for how that translates to real world results, 37 00:03:20,080 --> 00:03:28,000 they showed some slides later comparing the 9070XT to the 7900 GRE, and it does look pretty impressive. 38 00:03:28,000 --> 00:03:32,320 And who knows, maybe you'll actually use the AI features that AMD is coming out with, 39 00:03:32,320 --> 00:03:37,200 like the optional image inspector, which is an algorithm that monitors game frames and looks 40 00:03:37,200 --> 00:03:45,920 for image corruption during gameplay, and the AMD chat? Oh good, it's a local large language 41 00:03:45,920 --> 00:03:51,760 model that can supposedly help you configure your GPU and tune it. Neat, like how you can 42 00:03:51,760 --> 00:03:57,440 keep your desk setup neat with magnetic cable management from LTTstore.com. Speaking of things 43 00:03:57,440 --> 00:04:03,520 looking nice, the image quality improvements to the media engine look great. If their slides are 44 00:04:03,520 --> 00:04:07,840 to be believed, you could actually stream to Twitch with one of these cards without it looking 45 00:04:07,840 --> 00:04:12,720 like hot garbage. Let's just hope that, along with the better encoding, that they have finally 46 00:04:12,720 --> 00:04:20,160 fixed that weird 1080p hardware encoding bug. Regardless of that, they've also promised 8K 47 00:04:20,160 --> 00:04:27,040 ADFPS encode and decode, support for HEVC AV1 and H264, accelerated post processing effects, 48 00:04:27,040 --> 00:04:31,040 and no limit on the number of sessions or encoding streams. 49 00:04:31,120 --> 00:04:37,040 Team Green, meanwhile, locks this behind their professional card paywall, so uh, good guy AMD. 50 00:04:37,040 --> 00:04:45,360 But also, darn it AMD! Let's take a look at the hardware. The 9070 and the 9070XT are two 51 00:04:45,360 --> 00:04:51,600 surprisingly different cards. They both appear to use the same die, they're both equipped with 16 52 00:04:51,600 --> 00:04:59,760 gigs of memory on a 256-bit bus, but the XT is clocked way higher at darn mere 3 GHz and pulls 53 00:04:59,760 --> 00:05:07,200 a whopping extra 84 watts. On top of that, it has 8 more compute units and RT accelerators, 54 00:05:07,200 --> 00:05:16,080 and 16 more AI accelerators for an extra 392 tops. All of which would be totally fine if the 9070 55 00:05:16,080 --> 00:05:23,360 wasn't priced a mere $50 lower. AMD, you're doing the thing again. The 7900XT got terrible 56 00:05:23,360 --> 00:05:29,280 reviews at launch because it was only $100 less than the XTX and it was way slower. 57 00:05:29,760 --> 00:05:34,960 Months later, you discounted it when it wasn't selling and it picked up, but the reviews were 58 00:05:34,960 --> 00:05:42,160 already out of the gate. There's still time to fix this, though. 499. Do it. Do it! Anywho, 59 00:05:42,160 --> 00:05:48,720 both cards, regardless, are quite future-ready feature-wise, with PCI Express Gen 5 interfaces, 60 00:05:48,720 --> 00:05:57,360 HDMI 2.1b and DisplayPort 2.1a, although it is worth noting that not all DP 2.1a ports are made 61 00:05:57,360 --> 00:06:04,240 equal, and AMDs support only 54 gigabit per second compared to the 80 gigabit per second on NVIDIA's 62 00:06:04,240 --> 00:06:09,360 50 series. In daily use, this is unlikely to matter much, since these cards aren't really 63 00:06:09,360 --> 00:06:14,720 expected to be as powerful as NVIDIA's top GPUs, and with display stream compression, you should 64 00:06:14,720 --> 00:06:21,200 still be able to run a 4K 240Hz monitor, but it is worth noting. But what should we expect 65 00:06:21,520 --> 00:06:28,400 performance-wise? Well, thanks to FSR4, maybe a lot? As per usual, you're gonna have to wait 66 00:06:28,400 --> 00:06:32,560 for the reviews to drop from ourselves and others next week before you pull the trigger. 67 00:06:32,560 --> 00:06:37,920 If their image comparisons are real, and I don't have reason yet to believe they're not real, 68 00:06:37,920 --> 00:06:43,920 the image quality has improved so much that maybe you won't mind bumping down from quality mode to 69 00:06:43,920 --> 00:06:49,520 performance. AMD, like NVIDIA, is leaning on AI and neural rendering for this, and they're calling 70 00:06:49,520 --> 00:06:55,760 it FidelityFX Super Resolution 4 ML-powered upscaling. This was teased almost a year ago, 71 00:06:55,760 --> 00:07:00,880 and hey, it looks like it's finally here. The only issue is the lack of early support. 72 00:07:00,880 --> 00:07:07,200 They've got a decent list of 30-plus games at launch, with over 75 more on the way for 2025, 73 00:07:07,200 --> 00:07:12,000 and I would expect that to grow, but don't buy promises. Buy what you can use today, 74 00:07:12,000 --> 00:07:16,320 because there's no telling if your favorite game will or won't end up being supported. 75 00:07:16,320 --> 00:07:24,080 A feature that does support a much more sizable 1,000-plus games list is HyperRX, sort of. 76 00:07:24,080 --> 00:07:27,840 It's a one-click solution that's been around for a while now and enables features like 77 00:07:27,840 --> 00:07:34,480 anti-lag, boost, and FSR with Fluid Motion Frames 2.1 joining that roster and more coming soon in 78 00:07:34,480 --> 00:07:40,800 2025. However, since not all of those games benefit from every single one of those features, 79 00:07:40,800 --> 00:07:44,560 your performance improvements will vary on a game-to-game basis, 80 00:07:44,560 --> 00:07:49,520 and not everyone is a fan of frame generation, regardless of whether those frames are 81 00:07:49,520 --> 00:07:55,440 green fake frames or red fake frames. Speaking of Team Green, let's talk about availability, 82 00:07:55,440 --> 00:08:00,800 because after all, who gives a single flying f*** about anything that AMD said today 83 00:08:00,800 --> 00:08:05,840 if the cards can't be bought? I mean, I was super excited about the 50-series 84 00:08:05,840 --> 00:08:10,480 until they were proven to be made out of unobtainium, so you'll have to forgive me for being a little 85 00:08:10,480 --> 00:08:17,520 jaded, but I mean, these cards have been in the hands of retailers for months. We saw samples 86 00:08:17,520 --> 00:08:24,320 at CES in early January with our own eyes. I've got one in my hand now! So when AMD says 87 00:08:24,960 --> 00:08:32,400 wide availability, March 6th, 2025, there's a small, small ray of hope inside me that hasn't 88 00:08:32,400 --> 00:08:38,240 quite withered and died yet and wants so desperately to believe them. So for the love of all that is 89 00:08:38,240 --> 00:08:44,240 good, please, AMD, please don't f*** this up. If this is another paper launch, it'll just be my 90 00:08:44,240 --> 00:08:50,160 13th reason why the GPU market right now is terrible and gamers should despair. And I don't want 91 00:08:50,160 --> 00:08:56,080 another reason, like Intel's B580 came out and buoyed my spirits. It's a great budget card. 92 00:08:56,080 --> 00:09:01,760 It's just that it's hard to find at or near MSRP, and then NVIDIA's 50-series has been trickling 93 00:09:01,760 --> 00:09:06,720 out into the market as if each one of them is artisanally whittled out of silicon by like 94 00:09:07,280 --> 00:09:14,480 one guy. Gamers want reasonably priced cards, they want them now, and AMD, guys, you've got 95 00:09:14,480 --> 00:09:20,880 a real opportunity to gain actual market share here, as long as your cards aren't buggy and don't 96 00:09:20,880 --> 00:09:26,720 melt. But hey, good news. AMD claims that this launch will be the best driver support they've 97 00:09:26,720 --> 00:09:32,160 ever had. And considering how much they've been delayed, apparently, I actually kind of believe 98 00:09:32,160 --> 00:09:36,560 that they've given the software team some extra time to cook. And they even say they've 99 00:09:36,560 --> 00:09:42,080 also improved their processes. New AI-optimized internal testing, increased number of hardware 100 00:09:42,080 --> 00:09:46,720 test configurations in the lab, and a new bug report tool that's loaded directly into the driver 101 00:09:46,720 --> 00:09:52,960 will hopefully lead to a smoother launch experience. But will it be a little smoother or a lot 102 00:09:52,960 --> 00:09:58,000 too smoother? We can't know until thousands of credit cards have been swiped and thousands of 103 00:09:58,000 --> 00:10:05,760 GPUs have played thousands of games, and also output thousands of LTT segues to our sponsor. 104 00:10:05,760 --> 00:10:11,520 If you guys liked this video, maybe check out our 9800X3D review. Those aren't on obtaining 105 00:10:11,520 --> 00:10:19,680 any more, and they should pair very nicely with any GPU you buy. Maybe even a red one.