WEBVTT 00:00:00.480 --> 00:00:08.320 $4,000. That is what these modified 48 gigabyte 00:00:05.520 --> 00:00:13.759 RTX 4090s from China are going for. Personally, I love how hot and loud they 00:00:11.120 --> 00:00:17.760 get. They're great. And the best part, aside from the inflated price that I 00:00:15.679 --> 00:00:23.920 paid that makes me feel like a big shot, is the fact that it loses in nearly 00:00:20.640 --> 00:00:25.840 every benchmark against a regular 4090 00:00:23.920 --> 00:00:31.279 gaming card. In case it wasn't obvious by this point, yes, this thing makes 00:00:28.240 --> 00:00:33.040 absolutely no sense for gamings, but 00:00:31.279 --> 00:00:38.960 there are tons of listings for them online and even positive reviews. Maybe 00:00:36.239 --> 00:00:46.480 they're great for productivity. Nope. Oh, of course, large language 00:00:43.040 --> 00:00:48.879 models love VRAM. The more the better. 00:00:46.480 --> 00:00:54.879 And not only are 5090s hard to get your hands on, but they have just 32 gigs of 00:00:51.920 --> 00:01:00.480 memory, which pales in comparison to this monster. But demonstrating what 00:00:58.079 --> 00:01:04.799 makes these cards so desirable that they're worth dismantling and then 00:01:02.640 --> 00:01:10.400 remanufacturing an already expensive card is tricky. See, if we run our 00:01:07.200 --> 00:01:13.840 regular AI benchmarks, our regular 4090 00:01:10.400 --> 00:01:16.159 here actually wins again. But under the 00:01:13.840 --> 00:01:23.119 right conditions, this thing should crush not only a regular old 24 gig 00:01:19.600 --> 00:01:25.680 4090, but even beat a 5090. And I'm 00:01:23.119 --> 00:01:31.280 going to show you after I generate a segue to our sponsor, 00:01:31.280 --> 00:01:37.680 Threat Locker. Threat Locker is hosting a free webinar series all about how to 00:01:35.680 --> 00:01:41.680 secure your IT environment. Just click our link in the video description to get 00:01:39.280 --> 00:01:45.640 caught up on any past sessions and to register for the next one. 00:01:52.560 --> 00:01:58.079 Before we get into our live drag race, I need to figure out what the heck it is 00:01:56.240 --> 00:02:01.200 we actually bought here. Because the closer I look at this thing, the more 00:01:59.680 --> 00:02:05.680 questions it raises. See, on the surface, it's pretty straightforward. I 00:02:02.960 --> 00:02:11.440 mean, it's an RTX 4090, and they probably desoldered the 24 gigs of DDDR 00:02:08.479 --> 00:02:16.080 6X and replaced it with 48 gigs of DDDR 6X, except for one small problem. See, 00:02:13.520 --> 00:02:20.560 it's got a blower style two slot cooler, which allegedly NVIDIA does not allow 00:02:18.959 --> 00:02:24.560 its gaming partners to build because they're worried that cards like this, 00:02:22.400 --> 00:02:29.120 which can be packed into workstations or servers, would cannibalize their ability 00:02:26.720 --> 00:02:33.280 to sell enterprise and workstation cards. But it also doesn't seem to be 00:02:31.280 --> 00:02:39.360 one of NVIDIA's authorized professional card designs. All the imagery that I can 00:02:35.680 --> 00:02:41.760 find of the RTX A6000 ADA has the same 00:02:39.360 --> 00:02:45.519 cutout PCB at the back so that it can pull air both from the front and the 00:02:43.920 --> 00:02:50.319 back before exhausting it out the back of the case. So, this seems to be a 00:02:47.440 --> 00:02:55.040 unique cooler and PCB combination that some Chinese manufacturer or 00:02:52.400 --> 00:03:02.560 manufacturers, probably without any authorization from NVIDIA, has created 00:02:57.840 --> 00:03:05.200 in order to turn RTX 4090 and 4090ds 00:03:02.560 --> 00:03:09.599 into AI compute cards. And by turn them into, I really do mean like full 00:03:07.440 --> 00:03:13.840 Frankenstein's monster. We found this article on Techbot that appears to show 00:03:11.840 --> 00:03:19.200 cards that look an awful lot like these ones being validated on test benches 00:03:16.239 --> 00:03:23.760 next to pallets of regular gaming cards that appear to be headed for 00:03:21.040 --> 00:03:28.560 disassembly. And actually we can see here that yeah the initial hypothesis of 00:03:26.879 --> 00:03:32.480 desoldering the memory and then resoldering memory with double the 00:03:30.080 --> 00:03:39.840 capacity was obviously wrong because unlike any gaming RTX490 design we see 00:03:36.319 --> 00:03:42.239 ah yes memory packages on the back of 00:03:39.840 --> 00:03:47.040 the board. So to achieve the doubling of capacity they're using double the number 00:03:44.480 --> 00:03:50.799 of packages just like one of NVIDIA's professional cards would. It really goes 00:03:48.799 --> 00:03:55.040 to show how much NVIDIA charges their enterprise customers for their GPUs that 00:03:53.120 --> 00:03:59.760 it's more economical to buy gaming cards, disassemble them, discard the 00:03:57.760 --> 00:04:04.000 PCBs, and then go and source additional memory and remanufacture the card. Not 00:04:02.319 --> 00:04:08.560 that they discard the PCBs. I mean, if you've ever wondered where the heck all 00:04:05.599 --> 00:04:13.760 of those uh chipless 4090s and stuff that you see on eBay come from, yo, 00:04:11.280 --> 00:04:20.000 those are funky thermal pads. They look so dry. really clean layout and massive 00:04:17.519 --> 00:04:23.360 power delivery. This is key because not only does NVIDIA charge their enterprise 00:04:21.840 --> 00:04:27.520 customers an arm and a leg for the privilege of using their products, but 00:04:25.360 --> 00:04:33.280 they typically deliver them with lower power profiles. This does improve 00:04:30.080 --> 00:04:35.280 thermals and longevity, but it also 00:04:33.280 --> 00:04:40.160 harms performance. So, these remanufactured cards will be more 00:04:37.520 --> 00:04:43.840 affordable and also more performant in some cases than the official NVIDIA 00:04:42.240 --> 00:04:48.240 solution. That assumes, of course, though, that we can keep the thermals 00:04:45.520 --> 00:04:52.479 under control with this wimpy dual slot cooler. To be clear, I'm not attacking 00:04:50.160 --> 00:04:56.400 the quality. That appears to be a fat vapor chamber, which is going to help a 00:04:54.160 --> 00:05:02.400 lot. It's just not a lot of metal to cool 450 watts of GPU. Hey, this 00:05:00.720 --> 00:05:05.759 explains the observation that you guys made about the performance being 00:05:03.840 --> 00:05:09.680 crappier, though. If you've got twice as many memory chips, we're just eating up 00:05:08.000 --> 00:05:15.520 some of our limited power profile for the whole board. So, yeah. Cool. Mystery 00:05:12.960 --> 00:05:21.039 solved. All right, let's fire up GPUZ. Have a look at what the heck this thing 00:05:17.360 --> 00:05:24.000 is. RTX 4090. So, it's not a D. It's too 00:05:21.039 --> 00:05:27.919 bad. I love the D. I don't actually love the D. It's smaller and less powerful. 00:05:25.840 --> 00:05:30.800 Micron memory. That's consistent with what I saw when I was taking apart the 00:05:29.440 --> 00:05:34.240 card. We have no way of knowing if it's new or if it was, you know, pulled off 00:05:32.639 --> 00:05:38.080 of gaming cards. If anything, I would expect that it was. But hey, it's 00:05:36.080 --> 00:05:42.400 Micron. This is interesting. They obviously had to develop their own VIOs 00:05:40.400 --> 00:05:47.600 for these cards. So, they just set the sub vendor to NVIDIA, unless NVIDIA is 00:05:46.080 --> 00:05:51.600 complicit in the creation of these, but man, do I ever doubt that. And yeah, 00:05:49.840 --> 00:05:55.360 realistically, other than having double the memory, it looks like a pretty 00:05:53.360 --> 00:06:01.199 normal 4090. Do we have firm mark on here? Like, can we just kind of see what 00:05:57.199 --> 00:06:03.919 happens when we load it? Hi, buddy. 00:06:01.199 --> 00:06:08.560 Yeah, I I see you there. You don't got to yell about it. Right. in firmark GPU 00:06:06.479 --> 00:06:13.680 is getting less than 400 watts because we have so much memory on this thing. 00:06:11.039 --> 00:06:17.199 PCIe is operating at Gen 4 by6, so that's operating normally. Not something 00:06:15.199 --> 00:06:21.760 I would necessarily take for granted. Wow, is that ever loud. 5,000 00:06:20.400 --> 00:06:24.000 RPM. 00:06:24.000 --> 00:06:31.120 Woo! Do we have afterurner on here? Oh, Norton 360 for gamers. Yeah, comes with 00:06:29.199 --> 00:06:36.160 afterurners for overclocking. For overclocking my antivirus. MSI, you 00:06:33.919 --> 00:06:39.759 already sold me a GPU. You don't have to bundle it with Norton. Well, that was 00:06:37.759 --> 00:06:43.759 anticlimactic. No power level adjustment, so we're stuck with what we 00:06:41.600 --> 00:06:48.880 got. But hey, at least it's a nice GPU temperature that should stay under 00:06:45.520 --> 00:06:50.880 control even in a closed chassis. But 00:06:48.880 --> 00:06:54.800 barely, especially if you have these packed next to each other like they 00:06:52.800 --> 00:06:58.720 appear to be designed to do. Dude, you can even see the PF cap reason the 00:06:56.960 --> 00:07:02.000 entire time was power. If we could increase the power limit of this card, 00:07:00.400 --> 00:07:05.840 it could go a little better. For our benchmark, we've set up Olama, our 00:07:03.919 --> 00:07:09.639 locally deployed AI model runner to model Gemma 3.27B-IT-Q4_0. 00:07:10.560 --> 00:07:17.840 Right now, all we got to do is add our prompts and go. This is on our regular 00:07:14.800 --> 00:07:19.840 4090, and this is the Q4. The percentage 00:07:17.840 --> 00:07:23.759 of this stuff that's just outright wrong is wild. Lionus is known for his 00:07:22.080 --> 00:07:28.000 energetic, enthusiastic, and often self-deprecating personality. He's 00:07:25.599 --> 00:07:33.360 transparent about mistakes and isn't afraid to admit when things go wrong, 00:07:31.039 --> 00:07:38.240 which they often do in their experimental builds. He's become a 00:07:35.759 --> 00:07:45.919 trusted voice in the tech community due to his honesty and accessibility. That's 00:07:40.800 --> 00:07:50.479 all wrong. How tall his line is so 6' 3 00:07:45.919 --> 00:07:52.720 in. Why does it think you're so tall? 00:07:50.479 --> 00:07:57.680 Well, it's thinking horizontal, but it knows that my height is noticeable 00:07:55.440 --> 00:08:01.039 compared to other people on camera. It also pulled this from Wikipedia. My 00:07:59.199 --> 00:08:04.560 height is correct on Wikipedia. Let's see what the sources. I don't get this. 00:08:02.800 --> 00:08:08.319 It cites Wikipedia as a source, but Wikipedia has my height right. Does it 00:08:06.319 --> 00:08:11.680 though? Yeah, I'm pretty sure. I don't see a height. I swear my height used to 00:08:10.240 --> 00:08:17.840 be here. Well, now we know why it's getting it wrong then because that's not on here. A son and two daughters. Yeah, 00:08:15.680 --> 00:08:21.520 that's correct. It does not list your height. So now we 00:08:19.599 --> 00:08:28.120 know why it is wrong. The first time you run it, it can be a little bit slow. So 00:08:23.840 --> 00:08:28.120 I'll ask it something. Um, 00:08:28.800 --> 00:08:36.719 let's see if it knows. I It might They know. Yes, Nicholas, known online as 00:08:34.320 --> 00:08:41.200 Nicholas Plof Visuals, does own and operate a display at the Mall of America 00:08:39.039 --> 00:08:44.959 in Minneapolis, Minnesota. Yeah, that's not my I don't have a website called 00:08:42.479 --> 00:08:51.200 Nicholasplof.com/chromatic. Yes, Nick from Lionus Tech Tips does 00:08:48.640 --> 00:08:55.040 co-own a display at the Mall of America. It's the same one mentioned previous and 00:08:53.200 --> 00:08:59.440 that clarifies which Nick PL we're talking about. You're his son. This just 00:08:56.959 --> 00:09:04.480 completely made this up. Yeah, I love AI. It's great. So, the answer is yes. 00:09:01.600 --> 00:09:08.399 Both Nicholas Plof senior and Nick Plof from LT are owners of the chromatic I 00:09:06.720 --> 00:09:11.680 got to get, you know, whatever dividends I'm owed from owning this chromatic 00:09:09.920 --> 00:09:16.560 motion display. This is a video about the display, so we should probably click 00:09:13.920 --> 00:09:20.160 it. not available anymore. I think you made all of this up. You're absolutely 00:09:18.399 --> 00:09:22.959 right to question me. My apologize. I seem to have completely fabricated that 00:09:21.839 --> 00:09:26.959 information. I'm still under development. There is no evidence that 00:09:25.440 --> 00:09:30.720 either Nicholas Blof, the artist, or Nick Bl from LT, owns or operates a 00:09:29.360 --> 00:09:35.120 display at the Mall of America. Why is it so hung up on the Mall of America? 00:09:32.480 --> 00:09:39.839 Okay, but I still want to know if Nick Blof owns a display. You are absolutely 00:09:37.760 --> 00:09:45.680 right to push back and clarify. I got completely sidetracked by the display 00:09:42.640 --> 00:09:48.480 meaning an art installation. Nicholas PL 00:09:45.680 --> 00:09:53.360 frequently shows off his display wall. That's not true. What are you talking 00:09:50.560 --> 00:09:56.160 about? I got I do own several displays, but I don't have an entire All of this 00:09:54.720 --> 00:09:59.839 is besides the point. We're supposed to be benchmarking these. Okay, so we'll 00:09:57.920 --> 00:10:07.440 write the same prompt. These are kind of hot and ready now. So, uh, how do I 00:10:02.720 --> 00:10:09.200 build a computer? Three, two, one. 00:10:07.440 --> 00:10:15.600 Whoa. Oh wow. Whoa. You're getting crushed, 00:10:12.720 --> 00:10:20.079 brother. This will So Q4 isn't even the bad one. So this is still using less 00:10:18.240 --> 00:10:24.240 than 24 gigs of VRAM. Like actually, let's open the hardware information. Or 00:10:21.839 --> 00:10:29.279 you hear my GPU going? I do. I do. She's going, bud. Like I'm using not fully 24 00:10:26.880 --> 00:10:33.839 gigs. I'm using Yeah. 20.2 out of 24 gigs. What's your VRAM usage at? 21 00:10:31.680 --> 00:10:37.519 gigs. So yours is just faster right now. You know what? I think yours might have 00:10:35.360 --> 00:10:41.760 just kind of hiccuped at the beginning, though. So we both got slightly 00:10:39.440 --> 00:10:46.079 different answers which we would expect even running the same model and asking 00:10:43.440 --> 00:10:50.640 the same question that is typical. They do contain a lot of the same tips. What 00:10:48.959 --> 00:10:54.640 we want to know is the performance. So we need to know how many tokens per 00:10:52.320 --> 00:10:59.760 second. You got slightly more. You got 40.43 response tokens per second. I got 00:10:56.959 --> 00:11:02.959 40.06. Okay. Realistically it was about the same then other than that little 00:11:01.519 --> 00:11:06.240 hiccup at the beginning while his spooled up. Let's try one more time just 00:11:04.640 --> 00:11:13.279 for loss. One more time. Sure. Write a short story, no more than500 words, 00:11:10.000 --> 00:11:15.519 about Nicholas Plof discovering his love 00:11:13.279 --> 00:11:23.040 of displays. We should specify Nicholas Plof from LT. Okay. Not from the Mall of 00:11:18.320 --> 00:11:25.279 America. 1 2 3 go. 00:11:23.040 --> 00:11:28.480 The gradient of wonder. The ghost in the pixels. Oh, I like your title better. 00:11:26.959 --> 00:11:34.800 That's way cooler than the gradient of wonder. He wasn't born chasing perfect 00:11:31.760 --> 00:11:36.720 blacks or pixel densities. 00:11:34.800 --> 00:11:39.720 He wasn't even particularly good with electronics. 00:11:39.760 --> 00:11:46.320 You've really taken to this, haven't you? Lionus said, walking in on Nicholas 00:11:43.839 --> 00:11:51.839 painstakingly adjusting the gamma curve on her particularly finicky monitor. 00:11:48.800 --> 00:11:54.000 It's fascinating, surprised by some 00:11:51.839 --> 00:11:58.320 enthusiasm. He wasn't just reviewing a monitor. He was sharing his love of 00:11:56.000 --> 00:12:03.440 displays with the world. And in that moment, he knew he was exactly where he 00:12:00.640 --> 00:12:09.120 was meant to be. Wow. He was simply a man captivated by the wonder of light. 00:12:06.560 --> 00:12:14.079 It's actually pretty good Okay, but how many per second? 40.1. I'm at 00:12:12.240 --> 00:12:17.360 39.68. Okay, so I actually get a little bit more performance because a little 00:12:15.519 --> 00:12:25.279 bit. We're about to ratchet this up a notch by switching over to uh Q8_0. 00:12:22.079 --> 00:12:27.600 This theoretically will not fit in 00:12:25.279 --> 00:12:31.839 Plof's VRAM, which doesn't mean he can't run it. It just means that it's going to 00:12:29.680 --> 00:12:36.160 spill out into system memory which should result in marketkedly different 00:12:33.920 --> 00:12:41.360 performance. Yeah. Test prompt. Really? Yes. I want to see what it does. Fine. 00:12:38.639 --> 00:12:45.440 He smiled. He still loved wood. Still enjoyed the satisfaction of creating 00:12:43.120 --> 00:12:50.480 something tangible. It's true. Yours is done though. My Look at how my even the 00:12:47.760 --> 00:12:55.519 test prompt. Wa. This is a much larger model though. 00:12:53.519 --> 00:13:00.160 So theoretically the quality of its responses should be greater. I can't 00:12:58.240 --> 00:13:06.399 believe how much faster that ran on mine. Right. Your system memory usage is 00:13:02.880 --> 00:13:11.519 gig huge. So total 40 gigs and I'm using 00:13:06.399 --> 00:13:15.040 a total of about 45, but most of that 33 00:13:11.519 --> 00:13:17.760 gigs of it is my GPU. So this model 00:13:15.040 --> 00:13:21.279 simply wouldn't fit in any consumer GPU, including the 5090. Okay, that was a 00:13:19.600 --> 00:13:24.399 test prompt. Let's try one more actual prompt. Let's Should we try one of our 00:13:23.120 --> 00:13:29.120 previous prompts just to see if the quality is a little bit better? Sure. Let's I want to know if it knows who the 00:13:27.440 --> 00:13:34.120 Nicholas Plof is and whether he owns a display or not. 3 2 1 go. 00:13:35.360 --> 00:13:42.160 Yeah, yours is spitting out like lines while mine is going word by word. Some 00:13:40.320 --> 00:13:46.480 of this Where is it getting the woodworking stuff from? I don't know. I 00:13:44.480 --> 00:13:49.920 do know some woodworking, but that's like a high school skill. That is not 00:13:48.160 --> 00:13:54.079 I've never owned a woodworking shop. Okay, here's Nicholas's display wall. It 00:13:52.240 --> 00:13:59.199 has an actual reference list. It's not available. It's just making they're making fake links. It's done, but it's 00:13:56.959 --> 00:14:02.160 like tiny by comparison. Let's see your prompt. So, yours is way more detailed 00:14:00.560 --> 00:14:07.279 than mine, too. Like, I have like a little bit. It's similar. 00:14:05.040 --> 00:14:11.519 And it specifically, he upgraded to a Samsung Odyssey Neo G9. I did not do 00:14:09.199 --> 00:14:14.959 that. No, but that was an actual thing. We made a video about it. It was. Yeah. 00:14:13.279 --> 00:14:19.920 Do you want to see if that YouTube link is an actual video this time? No, it's 00:14:17.279 --> 00:14:27.440 not. No. Response tokens per second 26 and a 4.76. Literally five times the 00:14:24.079 --> 00:14:29.440 performance. It's still AI gobbledegook, 00:14:27.440 --> 00:14:33.040 but it's doing it way faster. Way faster. Kayla, you want to try image 00:14:31.839 --> 00:14:38.320 generation? What do you want to start with? YouTube host owning or fixing a 00:14:35.519 --> 00:14:42.480 computer. Oh, there goes your card. All right, she's going. So, what do you do 00:14:40.639 --> 00:14:48.399 with this after? Do you keep it? This is That's what you can use to upgrade from 00:14:45.360 --> 00:14:51.440 your 3080. You don't need a 5090. You 00:14:48.399 --> 00:14:53.199 can use a 48 gig 4090. you'll have even 00:14:51.440 --> 00:14:57.600 more VRAM. I don't need more VRAM. We'll talk about 00:14:55.600 --> 00:15:02.720 that in a little bit more detail later, but more VRAM is not necessarily more 00:15:00.320 --> 00:15:06.320 better for especially certain workloads. Yours is going faster. I'll tell you 00:15:03.920 --> 00:15:11.120 that much. Not that much faster. No, that thing screams though. Okay. How's 00:15:08.079 --> 00:15:13.440 my VR use though? Wow. I am using 33 00:15:11.120 --> 00:15:17.680 gigs of VRAM though. You're at I'm not Wow. Like 12. It's not even using 00:15:16.000 --> 00:15:21.839 everything you have. Yeah. What's interesting is mine is using the same 00:15:20.079 --> 00:15:24.720 amount of system memory as you too though. Yeah, I'm using a little bit 00:15:23.199 --> 00:15:28.959 more. I'm using like Yeah, hold on. Maybe I left some Lama stuff open in the 00:15:27.040 --> 00:15:32.480 back. This is weird. Our performance is about the same cuz and which makes sense 00:15:30.959 --> 00:15:36.240 cuz our parameters are all the same and we shouldn't be using more than the 00:15:34.399 --> 00:15:42.480 amount of memory that we both have, but my RAM use is way higher than his. 00:15:40.079 --> 00:15:46.720 Well, these are fun. Sure, that monitor looks like Apple 00:15:44.240 --> 00:15:50.720 designed it. And it's backwards for like the stand is going into the screen. I 00:15:48.959 --> 00:15:54.639 love AI. It's so good. I like how devious this guy looks. Look at him go. 00:15:52.959 --> 00:15:58.880 This one's not bad. Other than that mouse would not fit any human hand. It 00:15:57.120 --> 00:16:02.560 might fit this guy's hand. His fingers though. He's got an extra finger. Oh, 00:16:00.480 --> 00:16:05.839 yeah. Yeah. I like how I got a cartoon. Oh, that's cool. You didn't? Yeah. 00:16:04.320 --> 00:16:11.279 What's up with that? Look at this guy's keyboard. I want that keyboard. And a 00:16:08.480 --> 00:16:14.880 wireless microphone. What? He's even got the flip down mic on the side. Dude, it 00:16:13.360 --> 00:16:20.240 even has a YouTube logo in the background. That's wild. I don't like 00:16:16.720 --> 00:16:21.920 it. Okay, let's try doing a higher batch 00:16:20.240 --> 00:16:25.920 size though, cuz maybe when you know we're doing the same batch size, it's fine. But let's do let's go up to like 00:16:24.399 --> 00:16:30.639 16. Hold on. I want to see if there's a good one. This one's pretty good. That 00:16:28.720 --> 00:16:35.279 one's not bad. I could believe that there's a funky display that has like 00:16:33.680 --> 00:16:40.399 the power supply coming off the back like this or something like I His 00:16:37.040 --> 00:16:43.600 fingers are crossed weirdly. Yeah, he's 00:16:40.399 --> 00:16:45.120 missing one, I think. Yeah. But overall, 00:16:43.600 --> 00:16:48.720 if you just like threw this in the marketing collateral for something, not 00:16:46.720 --> 00:16:52.720 many people would notice right away. No. That's pretty wild. And this headset 00:16:50.240 --> 00:16:57.120 looks pretty sick. Yeah. The nice like pink purple. Yeah. I like it. Oh, he's 00:16:55.040 --> 00:17:02.720 missing a watch strap. Yeah, it's his skin. It's That's the new wearable. His 00:17:00.240 --> 00:17:06.480 finger is coming out of the computer and it's backwards. I don't know why the 00:17:04.079 --> 00:17:10.240 displays are backwards. I don't like it. I'm going to hit Okay, let's Your 00:17:08.480 --> 00:17:13.839 machine is still like really upset at you. Yeah, I think I I think I might 00:17:12.160 --> 00:17:17.520 just restart real quick. Did you hit run again or something? No, cuz your case 00:17:16.079 --> 00:17:21.679 sampler is still going. I bet you hit run again by accident. 00:17:19.919 --> 00:17:27.439 What' you do? I clicked it a few times. Oh my god. So, this is interesting 00:17:23.600 --> 00:17:29.840 though because I managed to output 00:17:27.439 --> 00:17:33.360 at the same time as you. Here we go. We've got Comfy UI with a stable 00:17:31.360 --> 00:17:36.720 diffusion checkpoint loaded up. Let's uh give it a simple prompt. YouTube host 00:17:35.280 --> 00:17:41.679 fixing a computer. Why don't we start with two in our batch size? Wait, wait, 00:17:39.679 --> 00:17:44.880 wait. Let's also put a couple of negative notes. I'm going to assume it's 00:17:43.679 --> 00:17:50.000 something you don't want. So, blonde hair. I'm going to say blonde. Blonde hair. Okay. No blonde hair. Okay. And 00:17:48.400 --> 00:17:54.880 we're just doing two. Yeah. Let's do two. Okay. And run. Only click it once. 00:17:52.720 --> 00:18:00.080 I only clicked it once. Dude, my memory is at 26 gigs. Mine started way before 00:17:58.160 --> 00:18:03.919 yours, too. Uh, yeah. Mine is Why is your memory 00:18:02.160 --> 00:18:07.840 usage so low? I don't know. We're getting about the same performance. 00:18:05.760 --> 00:18:11.440 Yeah. I'm going to check uh task manager instead of just GPUZ. Oh, wait. Yeah, it 00:18:09.840 --> 00:18:15.600 says 24 gigs now. Oh, it does. Okay, maybe it just had to spool up, I guess. 00:18:13.120 --> 00:18:19.760 Hey, how did we end up with the same dude? It totally is. Dude, mine is 00:18:18.080 --> 00:18:24.000 actually pretty convincing. I think yours is better. His we his fingers look 00:18:21.600 --> 00:18:27.679 wonky on the left hand. Yeah, they're like long. Um, but otherwise, it's like 00:18:26.320 --> 00:18:33.200 pretty convincing. It's a better mic than mine. This guy is psyched to have 00:18:29.919 --> 00:18:35.520 the world's skinniest display. I want 00:18:33.200 --> 00:18:38.640 that. What? It's like a tiny little bar. That's actually super cool. But we 00:18:37.520 --> 00:18:47.679 weren't expecting there to be a difference generating just two images. Let's increase our batch size to 16 cats 00:18:44.559 --> 00:18:49.600 destroying a supermarket. Black cats. No 00:18:47.679 --> 00:18:56.799 black cats. No black cats. Bad luck. I am using 42 gigs of VRAM. And you simply 00:18:53.600 --> 00:18:59.280 do not have 42 gigs of VRAM. So your 00:18:56.799 --> 00:19:02.720 system memory usage is way higher. Yeah. How much is yours? Oh yeah. You're 00:19:01.200 --> 00:19:06.480 barely using it. Yeah. That's like normal system usage um for what we're 00:19:05.200 --> 00:19:10.160 doing normally. Yeah. That's about 8 gig. It's all just running on the GPU, 00:19:08.640 --> 00:19:13.440 which is what you want. And you can actually see in the progress bar. Oh, 00:19:12.000 --> 00:19:17.120 yeah. You actually have progress. I don't even have a progress bar yet. 00:19:15.200 --> 00:19:21.200 Okay, I promise you I hit run. I'm not going to press it again. Yeah, it's 00:19:18.400 --> 00:19:26.480 still at like 23.5 gigs out of 24. And 3D is like fully busy. So, it's trying. 00:19:24.480 --> 00:19:31.120 It's just taking a hot minute. It can't. It's not even progressing yet. So 00:19:28.240 --> 00:19:36.640 imagine if you had like eight of these 48 gig 4090s set up in a giant like 00:19:34.640 --> 00:19:41.120 mining rig essentially. Yeah. And then was just like make me images of this 00:19:39.039 --> 00:19:45.919 stuff or generate a prompter. Write me write me stories so I can sell romance 00:19:43.440 --> 00:19:49.600 novels on Amazon and then generate the covers for them. Wow, look at my 00:19:47.520 --> 00:19:54.240 progress bar. There's like nothing. Yeah, you're getting crushed. Yours is 00:19:51.360 --> 00:19:58.000 like done already and I've got like 10% not even. To be clear, we're both 00:19:56.000 --> 00:20:01.520 running the same model. So, the actual quality of the output is going to be the 00:20:00.160 --> 00:20:04.880 same. Yeah, mine's just taking way longer. It's just the timing. Yeah, I 00:20:03.360 --> 00:20:10.240 can see why people are after these things, right? And with how little it 00:20:07.280 --> 00:20:13.520 actually costs to buy memory. Yes. Right. Like you can see on DM Exchange, 00:20:11.919 --> 00:20:20.240 we don't actually have spot pricing for DDDR 6X. Yeah. But DDDR6 is like what, a 00:20:17.919 --> 00:20:24.000 buck 80 per gigabyte or something? 290. 290 per gigabyte. So, even if it's 00:20:22.320 --> 00:20:28.320 double that, we're talking a few hundred bucks worth of additional memory when 00:20:26.160 --> 00:20:32.880 they remanufacture these cards. Yeah. Plus the PCD and power delivery and all 00:20:30.159 --> 00:20:35.280 that. I'm like stuck. Yours is going to finish and mine's 00:20:34.000 --> 00:20:40.000 going to be nowhere close. And the fact that this is so accessible that just like like you can literally just buy one 00:20:38.640 --> 00:20:43.280 of these and throw it in your computer. This didn't need it didn't need 00:20:41.760 --> 00:20:48.240 proprietary connections. It didn't have a special power specs. It's just a 450 W 00:20:45.840 --> 00:20:51.840 4090. I mean, just like mining, anyone can start like tooling around with AI. 00:20:50.320 --> 00:20:57.520 Yeah. You know what? Not just anyone can do though. What? Take their AI generated 00:20:55.200 --> 00:21:02.960 images and turn them into a real product. I am now teasing officially for 00:21:01.039 --> 00:21:07.679 the first time our leather jacket, which we're thinking 00:21:05.360 --> 00:21:11.679 we're going to call the Jensen. Matthew chucked this in an AI prompt. It 00:21:10.400 --> 00:21:17.440 was like, "Generate me a leather jacket." Dude, it is full of useless 00:21:14.880 --> 00:21:20.960 clips and buckles, pockets that aren't actually real. You may not have noticed 00:21:19.200 --> 00:21:26.000 this. two zippers. But there are actually two zippers in the front. We 00:21:24.000 --> 00:21:31.919 did have to make some tweaks to make it actually wearable. But this is as close 00:21:28.960 --> 00:21:35.679 to an AI generated jacket as you will actually be able to buy in person. You 00:21:34.480 --> 00:21:40.080 look like one of the anime characters that's got like six belts on. I know, 00:21:38.480 --> 00:21:43.440 right? Legitimately actually looks pretty cool. It looks pretty cool. It 00:21:41.679 --> 00:21:48.559 looks good. The only thing I would change is the dumb middles. No, no, no, 00:21:46.159 --> 00:21:54.320 no. It's a feature. See, look. You lose some weight. 00:21:51.440 --> 00:21:58.799 Suddenly it fits better. Boom. Slim fit. The Jensen coming soon to ltstore.com. 00:21:57.120 --> 00:22:04.720 Man, yours is done and mine is nowhere close. I don't think this one's finishing. But that definitely made cats 00:22:02.799 --> 00:22:08.720 destroying supermarket with no black cats. But what if you're not crazy about 00:22:06.400 --> 00:22:13.440 AI? In theory, cards like this with a ton of VRAM offer benefits in other 00:22:10.799 --> 00:22:18.640 workloads as well. So, why don't we see if it can outperform our regular 4090 in 00:22:16.000 --> 00:22:23.520 something like this massive Adobe Premiere project for how to build a PC, 00:22:21.039 --> 00:22:27.600 the last guide you'll ever need. As it turns out, even a huge featurelength 00:22:26.080 --> 00:22:32.320 project like this one with three terabytes of footage on the timeline, 00:22:29.679 --> 00:22:36.559 still only using about 16 gigs of VRAM. So, that's good news for me because it 00:22:34.799 --> 00:22:40.960 means I don't have to shell out for another dozen or so of these to chuck 00:22:38.480 --> 00:22:45.360 into our editing stations at the office. With that said, if you're working with a 00:22:42.720 --> 00:22:50.000 ton of layers, heavy VFX, or especially a lot of After Effects, it is possible 00:22:48.080 --> 00:22:54.080 that you could benefit from a card like this. What about Blender? What's the 00:22:52.320 --> 00:22:58.720 deal with this donut? It's really high poly. I asked a friend for a donut. Oh, 00:22:55.840 --> 00:23:02.720 okay. 2.1 gigs. Most files that you're going to get on Blender for free are 00:23:00.159 --> 00:23:04.799 like 300 megs at most. Oh, I see. So, we, you know, it was going to be really 00:23:03.679 --> 00:23:09.120 hard for us to actually max out the VRAM. Uh, but even with the 2.1 gig 00:23:06.799 --> 00:23:14.000 donut, copying it, multiplying it a bunch, we kind of the same on both 00:23:11.840 --> 00:23:18.480 systems. It might be a different story if we had like a bunch of shaders 00:23:15.520 --> 00:23:23.039 running and texture packs and stuff. Um, but at least just with polies, even in 00:23:21.120 --> 00:23:28.720 Blender, you're probably better off just doubling your actual system memory. Oh, 00:23:25.440 --> 00:23:30.559 that makes sense. Yeah. Okay, thanks. 00:23:28.720 --> 00:23:33.440 Finally, games. I guess we kind of spoiled this before, but I didn't 00:23:32.080 --> 00:23:37.039 actually get to play on it when we recorded those benchmarks, so I want to 00:23:35.200 --> 00:23:43.799 see how much VRAM this will actually use. Worst case scenario. Okay, frame 00:23:39.280 --> 00:23:43.799 gen rate tracing. Here we go. 00:23:45.440 --> 00:23:52.720 Maximum memory used about 14 14 12 gigs. 00:23:49.600 --> 00:23:55.200 So, in conclusion, benchmarks are like 00:23:52.720 --> 00:24:00.640 onions. Rather, they're like hips. They don't lie. So, don't buy one of these 00:23:57.360 --> 00:24:02.320 for gaming, but maybe consider one if 00:24:00.640 --> 00:24:06.960 you're looking for a bit of extra oomph in your productivity tasks that benefit 00:24:04.320 --> 00:24:10.799 from it or especially AI. I can't guarantee that it'll make a difference 00:24:08.320 --> 00:24:14.960 for every workload you have. And I can't even guarantee that it'll be reliable. 00:24:12.559 --> 00:24:20.559 Who knows? You know, these labbotomized 4090s that then get reassembled and 00:24:17.360 --> 00:24:22.720 shipped to you. Uh, good luck everybody. 00:24:20.559 --> 00:24:27.200 And good luck to our sponsor, Threat Locker. The best defense against cyber 00:24:25.200 --> 00:24:30.559 threats isn't always the latest and greatest security software. Although 00:24:28.960 --> 00:24:35.200 Threat Locker has that area covered, too. Sometimes a little extra knowledge 00:24:33.200 --> 00:24:39.840 goes a long way. That's why they're currently hosting a 100% free webinar 00:24:38.000 --> 00:24:43.600 series that is all about how to arm yourself with tips, tricks, and tools 00:24:42.159 --> 00:24:48.320 that will help secure your IT environment. Over 100 days, you will 00:24:45.919 --> 00:24:51.120 learn in depth about how to properly configure devices and cloud 00:24:49.679 --> 00:24:54.960 environments, how to strengthen application security, how to build a 00:24:52.960 --> 00:24:59.360 rapid and effective incident response plan, and so much more. Each session 00:24:57.440 --> 00:25:02.640 will give you practical step-by-step advice to help protect your business, no 00:25:01.200 --> 00:25:06.720 matter if you're inheriting an environment or starting one from 00:25:04.400 --> 00:25:11.039 scratch. So, don't wait. Sign up for free using our link in the description 00:25:08.480 --> 00:25:14.559 to catch up on any previous sessions and to make sure that you are enrolled for 00:25:12.720 --> 00:25:18.640 the next one. If you guys enjoyed this video, why not check out the time we 00:25:16.240 --> 00:25:22.799 showed how to work around NVIDIA's restrictions and game on one of their 00:25:20.960 --> 00:25:25.520 mining cards back in the day. That was pretty cool.