{"video_id":"GLGK0GKexds","title":"NVIDIA Showed Me Their Supercomputer","channel":"Linus Tech Tips","show":"Linus Tech Tips","published_at":"2026-01-04T14:53:45Z","duration_s":805,"segments":[{"start_s":0.4,"end_s":7.359,"text":"If I were a shovel manufacturer during the gold rush, what sort of shovel would","speaker":null,"is_sponsor":0},{"start_s":4.72,"end_s":12.639,"text":"I build for myself? Well, that's exactly the question that I'm answering today as","speaker":null,"is_sponsor":0},{"start_s":9.44,"end_s":15.12,"text":"we tour the Nyx supercomput at NVIDIA","speaker":null,"is_sponsor":0},{"start_s":12.639,"end_s":22.0,"text":"headquarters here in Santa Clara. Behind this door are 1,192","speaker":null,"is_sponsor":0},{"start_s":18.56,"end_s":24.96,"text":"B200 GPUs. Absolute monsters rated for","speaker":null,"is_sponsor":0},{"start_s":22.0,"end_s":30.24,"text":"up to,200 watts of power each. And they've been used for everything from AI","speaker":null,"is_sponsor":0},{"start_s":27.199,"end_s":33.04,"text":"research to DLSS upscaling for gamers to","speaker":null,"is_sponsor":0},{"start_s":30.24,"end_s":36.079,"text":"good old-fashioned NLPF drag racing. You know, to put the other AI up starts in","speaker":null,"is_sponsor":0},{"start_s":34.88,"end_s":40.719,"text":"their place. So, what are we waiting for? Let's gear up and get inside. Uh","speaker":null,"is_sponsor":0},{"start_s":38.559,"end_s":49.76,"text":"oh, shoot. What was the password again? Oh, that's right. S E G [music] U E for","speaker":null,"is_sponsor":1},{"start_s":45.84,"end_s":52.32,"text":"our sponsor, MSI. Their Crosshair A16 HX","speaker":null,"is_sponsor":1},{"start_s":49.76,"end_s":57.92,"text":"AI is an 18-in high-end gaming laptop that comes with a 50 series card and a","speaker":null,"is_sponsor":1},{"start_s":54.719,"end_s":59.6,"text":"240 Hz display. Experience smooth gaming","speaker":null,"is_sponsor":1},{"start_s":57.92,"end_s":64.72,"text":"on the go with the link in the video description. Are you tired of boring","speaker":null,"is_sponsor":1},{"start_s":61.68,"end_s":66.799,"text":"screwdrivers? Boom. Four new colors of","speaker":null,"is_sponsor":1},{"start_s":64.72,"end_s":73.52,"text":"the brand new LT transparent screwdriver. Plasma purple, cryo teal,","speaker":null,"is_sponsor":1},{"start_s":70.56,"end_s":76.64,"text":"molten orange, and carbon black. They're see-through. They're awesome. and","speaker":null,"is_sponsor":1},{"start_s":75.119,"end_s":81.04,"text":"they're probably going to sell out because for a limited time only. If you","speaker":null,"is_sponsor":1},{"start_s":79.04,"end_s":85.6,"text":"buy through YouTube shopping at the link in the video description, you will get","speaker":null,"is_sponsor":1},{"start_s":82.96,"end_s":89.84,"text":"20% off your driver, but only until December 30th. Collect them all or maybe","speaker":null,"is_sponsor":1},{"start_s":88.0,"end_s":94.0,"text":"pair them with our brand new bid case and all over print hoodie. This","speaker":null,"is_sponsor":1},{"start_s":91.92,"end_s":98.56,"text":"particular data center space is used for a combination of R&D and NVIDIA internal","speaker":null,"is_sponsor":0},{"start_s":97.119,"end_s":102.4,"text":"production, meaning that it gets reconfigured on a pretty regular basis","speaker":null,"is_sponsor":0},{"start_s":100.871,"end_s":107.6,"text":"[music] as NVIDIA works through validation of new chip or server","speaker":null,"is_sponsor":0},{"start_s":104.96,"end_s":111.759,"text":"designs. That's why the raised floor has these plumbing access hatches for","speaker":null,"is_sponsor":0},{"start_s":109.52,"end_s":115.92,"text":"highdensity water cooled deployments, even though at the moment it's set up","speaker":null,"is_sponsor":0},{"start_s":113.28,"end_s":120.32,"text":"for air cooling to accommodate the DGX B200 racks that currently line the","speaker":null,"is_sponsor":0},{"start_s":118.24,"end_s":124.88,"text":"floor. We're in the first of two rooms that are linked by this fiber optic","speaker":null,"is_sponsor":0},{"start_s":122.159,"end_s":129.599,"text":"cabling, each of which contains two cats or cold air containment units. We're","speaker":null,"is_sponsor":0},{"start_s":127.92,"end_s":133.36,"text":"going to go in one of them in a second, but before we do, I want to talk a","speaker":null,"is_sponsor":0},{"start_s":131.68,"end_s":138.319,"text":"little bit about some of the details that would be easy to miss in here, like","speaker":null,"is_sponsor":0},{"start_s":136.0,"end_s":142.239,"text":"this decibel meter here that probably illustrates why it's better for us to","speaker":null,"is_sponsor":0},{"start_s":140.0,"end_s":146.72,"text":"talk out here. The deployment is absolutely peppered [music] with","speaker":null,"is_sponsor":0},{"start_s":143.76,"end_s":151.44,"text":"sensors. obvious ones like temperature on both sides of the rack like this one","speaker":null,"is_sponsor":0},{"start_s":148.64,"end_s":158.4,"text":"right here and also less obvious ones like humidity and air pressure. Humidity","speaker":null,"is_sponsor":0},{"start_s":155.44,"end_s":163.36,"text":"management is very important. Too low and static electricity becomes a","speaker":null,"is_sponsor":0},{"start_s":160.16,"end_s":165.519,"text":"problem. Too high and your servers start","speaker":null,"is_sponsor":0},{"start_s":163.36,"end_s":172.239,"text":"to sweat and then you're going to start to sweat. Air pressure or maybe more","speaker":null,"is_sponsor":0},{"start_s":169.44,"end_s":177.599,"text":"accurately air flow comes down to cooling. The DGX V200 systems here are","speaker":null,"is_sponsor":0},{"start_s":176.0,"end_s":181.92,"text":"of course equipped [music] with their own cooling fans and the same goes for","speaker":null,"is_sponsor":0},{"start_s":179.92,"end_s":188.239,"text":"the accompanying advanced tablet networking equipment. But for we are","speaker":null,"is_sponsor":0},{"start_s":184.56,"end_s":190.4,"text":"talking about up to 14,000","speaker":null,"is_sponsor":0},{"start_s":188.239,"end_s":194.8,"text":"watts of power consumption in each of these units.","speaker":null,"is_sponsor":0},{"start_s":192.239,"end_s":200.0,"text":"Let's just say that these fans can use all the help that they can get. So,","speaker":null,"is_sponsor":0},{"start_s":197.04,"end_s":204.959,"text":"fresh air gets actively blown up through the floor, creating positive pressure","speaker":null,"is_sponsor":0},{"start_s":202.64,"end_s":212.239,"text":"inside this room where it gets pushed through our DGXs and out into the hot","speaker":null,"is_sponsor":0},{"start_s":208.239,"end_s":215.04,"text":"aisle where uh things definitely get a","speaker":null,"is_sponsor":0},{"start_s":212.239,"end_s":220.64,"text":"lot hotter. I mean, I was getting kind of chilly in there, but here I would be","speaker":null,"is_sponsor":0},{"start_s":218.0,"end_s":224.56,"text":"sweating in a matter of minutes. But in spite of the challenges, the systems are","speaker":null,"is_sponsor":0},{"start_s":222.64,"end_s":229.2,"text":"running perfectly. And NVIDIA pointed out that some of these nodes are running","speaker":null,"is_sponsor":0},{"start_s":226.319,"end_s":234.0,"text":"full tilt on undisclosed workloads as we speak. Now, because NVIDIA uses spaces","speaker":null,"is_sponsor":0},{"start_s":231.92,"end_s":238.319,"text":"like these for validation, it's important that they get all the little","speaker":null,"is_sponsor":0},{"start_s":235.76,"end_s":242.72,"text":"details right so customers can just take this blueprint, copy it, and paste it","speaker":null,"is_sponsor":0},{"start_s":240.64,"end_s":246.08,"text":"into their own facility and trust that it's going to run at scale. The","speaker":null,"is_sponsor":0},{"start_s":244.319,"end_s":250.48,"text":"networking racks, for instance, are strategically positioned to minimize the","speaker":null,"is_sponsor":0},{"start_s":248.4,"end_s":255.36,"text":"length of fiber optic cabling between the four cats that make up this cluster.","speaker":null,"is_sponsor":0},{"start_s":253.2,"end_s":259.519,"text":"Two in this room and two in the next one over. This is both to maintain signal","speaker":null,"is_sponsor":0},{"start_s":257.44,"end_s":264.72,"text":"integrity. They found any higher than 100 to 150 m can be problematic at these","speaker":null,"is_sponsor":0},{"start_s":262.4,"end_s":269.12,"text":"speeds, but it's also to maintain the integrity of the ceiling. See, it turns","speaker":null,"is_sponsor":0},{"start_s":266.8,"end_s":273.759,"text":"out that when you start bundling up hundreds of fiber optic cables, they can","speaker":null,"is_sponsor":0},{"start_s":271.759,"end_s":278.639,"text":"get pretty heavy. And the longer the runs, the more strain they put on the","speaker":null,"is_sponsor":0},{"start_s":275.759,"end_s":284.16,"text":"ceiling. It also helps save on cost. This I don't have much to say about","speaker":null,"is_sponsor":0},{"start_s":280.32,"end_s":285.68,"text":"other than Oh my god, isn't it","speaker":null,"is_sponsor":0},{"start_s":284.16,"end_s":289.919,"text":"beautiful? Okay, I lied. I do have some stuff to","speaker":null,"is_sponsor":0},{"start_s":287.44,"end_s":295.28,"text":"say. This cable management is not just for looks, but rather to maintain air","speaker":null,"is_sponsor":0},{"start_s":292.32,"end_s":300.24,"text":"flow. When you have this many cables, you actually do need to bundle them. And","speaker":null,"is_sponsor":0},{"start_s":297.759,"end_s":305.919,"text":"it also helps facilitate maintenance in the event of a broken or damaged cable.","speaker":null,"is_sponsor":0},{"start_s":303.12,"end_s":311.52,"text":"They run extra fiber in each bundle that can be terminated as needed, and it is a","speaker":null,"is_sponsor":0},{"start_s":308.639,"end_s":316.32,"text":"lot easier to find that when things are organized. [music] Careful thought goes","speaker":null,"is_sponsor":0},{"start_s":313.68,"end_s":319.919,"text":"into rack serviceability, too. While the networking racks use traditional","speaker":null,"is_sponsor":0},{"start_s":318.16,"end_s":325.44,"text":"vertical PDUs for their power distribution, the DGX racks use three","speaker":null,"is_sponsor":0},{"start_s":322.8,"end_s":328.88,"text":"top-mounted PDUs that help balance the three [music] phases of power coming in.","speaker":null,"is_sponsor":0},{"start_s":327.039,"end_s":332.479,"text":"On the subject of power, we skipped over these at the SFU data center, but uh","speaker":null,"is_sponsor":0},{"start_s":331.12,"end_s":336.72,"text":"have you ever wondered what a >> 415 volt 100 amp","speaker":null,"is_sponsor":0},{"start_s":334.24,"end_s":340.24,"text":">> power plug looks like? Wonder no longer. Another key advantage of putting all the","speaker":null,"is_sponsor":0},{"start_s":338.479,"end_s":345.6,"text":"PDUs up top is it makes it a little easier to get the DGX units in and out.","speaker":null,"is_sponsor":0},{"start_s":342.96,"end_s":349.28,"text":"They uh turn this one off and agreed to pull it out of the deployment just to","speaker":null,"is_sponsor":0},{"start_s":347.12,"end_s":352.639,"text":"show us how it's done. But when I asked if we could tear it down, they were","speaker":null,"is_sponsor":0},{"start_s":350.88,"end_s":356.32,"text":"like, \"Huh?\" Not because they have anything to hide,","speaker":null,"is_sponsor":0},{"start_s":354.24,"end_s":360.4,"text":"but just cuz they thought this soon to be decommissioned and recycled","speaker":null,"is_sponsor":0},{"start_s":357.919,"end_s":367.12,"text":"engineering unit that we can really get our greasy meat mitts into would be a","speaker":null,"is_sponsor":0},{"start_s":362.8,"end_s":369.52,"text":"lot more [music] fun. Woo!","speaker":null,"is_sponsor":0},{"start_s":367.12,"end_s":378.639,"text":"If you ever wondered what the heat sink would look like on a 1200 W GPU, wonder","speaker":null,"is_sponsor":0},{"start_s":374.88,"end_s":380.72,"text":"no more. How many heat pipes is this?","speaker":null,"is_sponsor":0},{"start_s":378.639,"end_s":385.28,"text":"It's like a forest of heat pipes in there. Oh, they're flat. That makes so","speaker":null,"is_sponsor":0},{"start_s":383.6,"end_s":389.12,"text":"much sense cuz you'll get more air flow through the fins. And I mean, that's","speaker":null,"is_sponsor":0},{"start_s":386.88,"end_s":395.44,"text":"some pretty hard working air dealing with not one but two rows of these GPUs.","speaker":null,"is_sponsor":0},{"start_s":392.8,"end_s":400.96,"text":"Not to mention this middle heat sink here that seems to be for the Envy Link","speaker":null,"is_sponsor":0},{"start_s":398.4,"end_s":408.319,"text":"switching equipment. I mean, I knew you need some beefy to allow these GPUs to","speaker":null,"is_sponsor":0},{"start_s":403.84,"end_s":410.88,"text":"pull the 192 gigs of HBM 3 memory that","speaker":null,"is_sponsor":0},{"start_s":408.319,"end_s":415.199,"text":"they have, but still, I'd never seen the cooling for it. Another thing I've never","speaker":null,"is_sponsor":0},{"start_s":413.199,"end_s":419.84,"text":"gotten to see, or at least never gotten to film, is NVIDIA's proprietary SXM","speaker":null,"is_sponsor":0},{"start_s":417.919,"end_s":423.28,"text":"interface. I know because I was looking for a B-roll shot of one of these a","speaker":null,"is_sponsor":0},{"start_s":421.599,"end_s":427.68,"text":"little while ago, and I realized I've never actually held one up on camera.","speaker":null,"is_sponsor":0},{"start_s":425.12,"end_s":432.639,"text":"So, now I've done it. And they also gave us this to show what Okay, not the same","speaker":null,"is_sponsor":0},{"start_s":430.08,"end_s":436.88,"text":"generation, but a similar GPU might look like without the heat sink. Oh, we've","speaker":null,"is_sponsor":0},{"start_s":435.28,"end_s":440.24,"text":"also got to look at the SXM interface that goes back to the rest of the system","speaker":null,"is_sponsor":0},{"start_s":438.56,"end_s":447.28,"text":"and also carries power for these [music] GPUs. These are 1,200 W each, but these","speaker":null,"is_sponsor":0},{"start_s":444.0,"end_s":450.08,"text":"cards go as high as 1,400 W. I mean, how","speaker":null,"is_sponsor":0},{"start_s":447.28,"end_s":454.56,"text":"high would the cooler be at that point? JK, it would be water cooled. We were","speaker":null,"is_sponsor":0},{"start_s":452.8,"end_s":459.28,"text":"hoping to see some water cooled machines today, but I guess NVIDIA has to save","speaker":null,"is_sponsor":0},{"start_s":457.195,"end_s":464.319,"text":"[music] something for next time. I got to say though, if the new stuff is even","speaker":null,"is_sponsor":0},{"start_s":461.44,"end_s":468.24,"text":"just equally cool to the last gen H100 water cooled machines we saw at the SFU","speaker":null,"is_sponsor":0},{"start_s":466.319,"end_s":473.039,"text":"Fur Supercomput, I'm sure it's absolutely mind-blowing. All of which is","speaker":null,"is_sponsor":0},{"start_s":470.88,"end_s":476.879,"text":"really cool, but what do they need all of this for? Well, the folks who are","speaker":null,"is_sponsor":0},{"start_s":475.36,"end_s":480.639,"text":"responsible for getting us access to everything today are actually from the","speaker":null,"is_sponsor":0},{"start_s":478.4,"end_s":484.96,"text":"GeForce gaming team. So, one of their big uses for these internal compute","speaker":null,"is_sponsor":0},{"start_s":482.479,"end_s":490.479,"text":"resources is, of course, deep learning superers sampling or DLSS. In a","speaker":null,"is_sponsor":0},{"start_s":487.599,"end_s":495.919,"text":"nutshell, DLSS allows your GPU to render your game at a lower resolution, then","speaker":null,"is_sponsor":0},{"start_s":492.96,"end_s":499.759,"text":"use deep learning or AI to upscale each output frame to your monitor's","speaker":null,"is_sponsor":0},{"start_s":497.44,"end_s":503.44,"text":"resolution. The benefit of this is that you can run at a higher frame rate for","speaker":null,"is_sponsor":0},{"start_s":501.52,"end_s":507.759,"text":"improved animation smoothness. But the drawback is that you can't create","speaker":null,"is_sponsor":0},{"start_s":505.68,"end_s":511.84,"text":"something from nothing and upscaled images struggle to achieve the same","speaker":null,"is_sponsor":0},{"start_s":509.28,"end_s":516.24,"text":"fidelity as a native rendered image. With that said, DLSS has improved a lot","speaker":null,"is_sponsor":0},{"start_s":514.479,"end_s":520.0,"text":"over the years and I got a chance to sit down and chat with Edward Leu from that","speaker":null,"is_sponsor":0},{"start_s":518.159,"end_s":523.839,"text":"team who talked us through some of the processes that they use to develop new","speaker":null,"is_sponsor":0},{"start_s":521.68,"end_s":529.2,"text":"features and new fixes. The first thing he pointed out is that while this hot","speaker":null,"is_sponsor":0},{"start_s":526.56,"end_s":534.16,"text":"new AI model took a thousand GPUs two weeks to train or whatever, makes for a","speaker":null,"is_sponsor":0},{"start_s":531.519,"end_s":538.399,"text":"good headline, it overlooks most of the actual cost and time, which is in the","speaker":null,"is_sponsor":0},{"start_s":536.72,"end_s":542.88,"text":"test training runs that take place before the hero run. His team is","speaker":null,"is_sponsor":0},{"start_s":541.04,"end_s":546.32,"text":"constantly iterating on the data they're feeding into DLSS and [music] the","speaker":null,"is_sponsor":0},{"start_s":544.56,"end_s":551.6,"text":"waiting, and they're evaluating new innovations in the AI space. Sometimes","speaker":null,"is_sponsor":0},{"start_s":549.12,"end_s":555.68,"text":"it's more surgical, like, \"Oh, hey, uh, we noticed Cyberpunk has an issue with","speaker":null,"is_sponsor":0},{"start_s":553.36,"end_s":558.88,"text":"cars having like three or four bumpers as you're driving around. How can we","speaker":null,"is_sponsor":0},{"start_s":557.519,"end_s":564.56,"text":"address this with the current model?\" That kind of thing can be turned around sometimes in a matter of weeks or months","speaker":null,"is_sponsor":0},{"start_s":562.56,"end_s":569.04,"text":"using the resources that we just saw. Though, uh, he was quick to point out","speaker":null,"is_sponsor":0},{"start_s":566.64,"end_s":573.6,"text":"that Jensen, if you're watching, there's no limit to how many GPUs his team could","speaker":null,"is_sponsor":0},{"start_s":571.279,"end_s":577.36,"text":"use to speed up the process. I told him I'd say that because the faster they can","speaker":null,"is_sponsor":0},{"start_s":575.44,"end_s":583.36,"text":"test a new data set, the faster they can tune it and ship it. Other times, the","speaker":null,"is_sponsor":0},{"start_s":581.04,"end_s":586.959,"text":"changes are more transformative, pun intended, like the move from a","speaker":null,"is_sponsor":0},{"start_s":584.959,"end_s":591.36,"text":"convolutional neural network to the more accurate transformer model that runs","speaker":null,"is_sponsor":0},{"start_s":588.959,"end_s":595.6,"text":"best on NVIDIA's newest cards. This can require basically a complete tear down","speaker":null,"is_sponsor":0},{"start_s":593.36,"end_s":600.0,"text":"and doover of the entire pipeline and can take a year or more. But I mean,","speaker":null,"is_sponsor":0},{"start_s":598.08,"end_s":603.36,"text":"hey, NVIDIA has bet pretty much their entire future on being able to stay at","speaker":null,"is_sponsor":0},{"start_s":601.76,"end_s":607.76,"text":"the forefront of their 21st century shovel technology. So, I guess that's","speaker":null,"is_sponsor":0},{"start_s":605.12,"end_s":611.04,"text":"the price you pay. But what does all this mean for traditional rendering?","speaker":null,"is_sponsor":0},{"start_s":609.36,"end_s":615.707,"text":"Well, the folks here believe very strongly that in time DLSS will not only","speaker":null,"is_sponsor":0},{"start_s":613.44,"end_s":619.44,"text":"be as good, it will be better than [music] traditional. In fact, Edward","speaker":null,"is_sponsor":0},{"start_s":617.6,"end_s":623.76,"text":"pointed to these slides from a talk that he gave a number of years ago, showing","speaker":null,"is_sponsor":0},{"start_s":621.36,"end_s":629.279,"text":"that what we think of as native rendering is already a pretty imperfect","speaker":null,"is_sponsor":0},{"start_s":626.64,"end_s":633.2,"text":"approximation. Here it is compared to a ground truth image, which is rendered at","speaker":null,"is_sponsor":0},{"start_s":631.519,"end_s":638.959,"text":"a much higher resolution than downsampled to 1080p. Native, it's","speaker":null,"is_sponsor":0},{"start_s":636.48,"end_s":644.32,"text":"clearly worse. And as you can see in this comparison, even back then with","speaker":null,"is_sponsor":0},{"start_s":640.959,"end_s":646.24,"text":"DLSS 2.0, So there were situations where","speaker":null,"is_sponsor":0},{"start_s":644.32,"end_s":652.0,"text":"a model trained on these higher resolution or more accurately higher","speaker":null,"is_sponsor":0},{"start_s":649.04,"end_s":656.972,"text":"sampled ground truth images could result in the GPU reconstructing an output with","speaker":null,"is_sponsor":0},{"start_s":654.8,"end_s":661.44,"text":"DLSS that's closer to ground truth [music] than the native image was. I","speaker":null,"is_sponsor":0},{"start_s":659.44,"end_s":664.56,"text":"asked, by the way, how are these ground truth images created? [music] and he","speaker":null,"is_sponsor":0},{"start_s":662.959,"end_s":669.519,"text":"said that typically it's actually just on run-of-the-mill gaming hardware, but","speaker":null,"is_sponsor":0},{"start_s":666.72,"end_s":673.839,"text":"instead of running at 60 or 120 frames per second, they'll sometimes be running","speaker":null,"is_sponsor":0},{"start_s":671.44,"end_s":679.839,"text":"tens or thousands of pixel samples to the point where we're talking more like","speaker":null,"is_sponsor":0},{"start_s":675.839,"end_s":682.48,"text":"60 to 120 seconds per frame or even","speaker":null,"is_sponsor":0},{"start_s":679.839,"end_s":686.32,"text":"more. Man though, if DLSS could consistently reconstruct that ground","speaker":null,"is_sponsor":0},{"start_s":684.8,"end_s":690.399,"text":"truth image from a lower resolution input every time, I'm sure no one would","speaker":null,"is_sponsor":0},{"start_s":688.48,"end_s":694.64,"text":"ever turn it off. Unfortunately for Edward and his team though, no one","speaker":null,"is_sponsor":0},{"start_s":692.32,"end_s":698.72,"text":"notices when DLSS is working perfectly. It's when it","speaker":null,"is_sponsor":0},{"start_s":696.64,"end_s":704.56,"text":"trips over itself that we tend to notice, and it still does do so","speaker":null,"is_sponsor":0},{"start_s":702.16,"end_s":708.72,"text":"fairly regularly. However, with that said, from my own experience, it has","speaker":null,"is_sponsor":0},{"start_s":706.88,"end_s":713.04,"text":"continued to improve at a pretty solid clip since its debut, and the rest of","speaker":null,"is_sponsor":0},{"start_s":710.64,"end_s":717.839,"text":"the industry has pretty much accepted that AI accelerated image enhancement is","speaker":null,"is_sponsor":0},{"start_s":715.76,"end_s":723.839,"text":"the future, whether every gamer wants it or not. So sales of shovels will likely","speaker":null,"is_sponsor":0},{"start_s":721.279,"end_s":727.76,"text":"continue until gamer morale improves. Good luck everyone. I hope it's not a","speaker":null,"is_sponsor":0},{"start_s":725.76,"end_s":733.2,"text":"bubble. This is a nice office. I'd hate for something to happen to it. Just like","speaker":null,"is_sponsor":0},{"start_s":730.16,"end_s":735.12,"text":"this is a nice segue to our sponsor,","speaker":null,"is_sponsor":1},{"start_s":733.2,"end_s":739.04,"text":"Squarespace. If you're building a brand or a business, it's important to have a","speaker":null,"is_sponsor":1},{"start_s":737.12,"end_s":743.279,"text":"website. Designing your own site can feel like a bit of a daunting task, but","speaker":null,"is_sponsor":1},{"start_s":741.12,"end_s":747.04,"text":"it really doesn't have to. Squarespace makes it easy to get your message across","speaker":null,"is_sponsor":1},{"start_s":744.959,"end_s":751.76,"text":"to potential customers and subscribers in a clean and digestible way. Their","speaker":null,"is_sponsor":1},{"start_s":749.36,"end_s":755.92,"text":"design intelligence tool utilizes AI and works with your own creativity to create","speaker":null,"is_sponsor":1},{"start_s":753.6,"end_s":760.399,"text":"a theme and style for your site that matches your personality. You can also","speaker":null,"is_sponsor":1},{"start_s":758.32,"end_s":766.079,"text":"use Squarespace to directly invoice your customers with payment options like a","speaker":null,"is_sponsor":1},{"start_s":762.56,"end_s":768.0,"text":"direct debit, Apple Pay, CLA, and more.","speaker":null,"is_sponsor":1},{"start_s":766.079,"end_s":773.2,"text":"And use their analytic tools to track sales, strategize, and continue to build","speaker":null,"is_sponsor":1},{"start_s":770.399,"end_s":776.959,"text":"your brand. There's millions of URLs available, and Squarespace's domain tool","speaker":null,"is_sponsor":1},{"start_s":775.279,"end_s":781.2,"text":"will help you search for the right address just for you. We've even used","speaker":null,"is_sponsor":1},{"start_s":779.279,"end_s":785.04,"text":"Squarespace for some of our own websites here. Start building your website today,","speaker":null,"is_sponsor":1},{"start_s":783.2,"end_s":790.16,"text":"and you'll receive 10% off your first purchase by visiting squarespace.com/LTT.","speaker":null,"is_sponsor":1},{"start_s":788.399,"end_s":795.519,"text":"If you guys enjoyed this video, maybe go check out the time we I don't know,","speaker":null,"is_sponsor":1},{"start_s":793.279,"end_s":800.079,"text":"something invidated. Let's do a throwback video. How about","speaker":null,"is_sponsor":0},{"start_s":797.519,"end_s":806.48,"text":"when I checked out the launch of G-Sync? The production values were lower, but","speaker":null,"is_sponsor":0},{"start_s":803.04,"end_s":806.48,"text":"hey, it was fun.","speaker":null,"is_sponsor":0}],"full_text":"If I were a shovel manufacturer during the gold rush, what sort of shovel would I build for myself? Well, that's exactly the question that I'm answering today as we tour the Nyx supercomput at NVIDIA headquarters here in Santa Clara. Behind this door are 1,192 B200 GPUs. Absolute monsters rated for up to,200 watts of power each. And they've been used for everything from AI research to DLSS upscaling for gamers to good old-fashioned NLPF drag racing. You know, to put the other AI up starts in their place. So, what are we waiting for? Let's gear up and get inside. Uh oh, shoot. What was the password again? Oh, that's right. S E G [music] U E for our sponsor, MSI. Their Crosshair A16 HX AI is an 18-in high-end gaming laptop that comes with a 50 series card and a 240 Hz display. Experience smooth gaming on the go with the link in the video description. Are you tired of boring screwdrivers? Boom. Four new colors of the brand new LT transparent screwdriver. Plasma purple, cryo teal, molten orange, and carbon black. They're see-through. They're awesome. and they're probably going to sell out because for a limited time only. If you buy through YouTube shopping at the link in the video description, you will get 20% off your driver, but only until December 30th. Collect them all or maybe pair them with our brand new bid case and all over print hoodie. This particular data center space is used for a combination of R&D and NVIDIA internal production, meaning that it gets reconfigured on a pretty regular basis [music] as NVIDIA works through validation of new chip or server designs. That's why the raised floor has these plumbing access hatches for highdensity water cooled deployments, even though at the moment it's set up for air cooling to accommodate the DGX B200 racks that currently line the floor. We're in the first of two rooms that are linked by this fiber optic cabling, each of which contains two cats or cold air containment units. We're going to go in one of them in a second, but before we do, I want to talk a little bit about some of the details that would be easy to miss in here, like this decibel meter here that probably illustrates why it's better for us to talk out here. The deployment is absolutely peppered [music] with sensors. obvious ones like temperature on both sides of the rack like this one right here and also less obvious ones like humidity and air pressure. Humidity management is very important. Too low and static electricity becomes a problem. Too high and your servers start to sweat and then you're going to start to sweat. Air pressure or maybe more accurately air flow comes down to cooling. The DGX V200 systems here are of course equipped [music] with their own cooling fans and the same goes for the accompanying advanced tablet networking equipment. But for we are talking about up to 14,000 watts of power consumption in each of these units. Let's just say that these fans can use all the help that they can get. So, fresh air gets actively blown up through the floor, creating positive pressure inside this room where it gets pushed through our DGXs and out into the hot aisle where uh things definitely get a lot hotter. I mean, I was getting kind of chilly in there, but here I would be sweating in a matter of minutes. But in spite of the challenges, the systems are running perfectly. And NVIDIA pointed out that some of these nodes are running full tilt on undisclosed workloads as we speak. Now, because NVIDIA uses spaces like these for validation, it's important that they get all the little details right so customers can just take this blueprint, copy it, and paste it into their own facility and trust that it's going to run at scale. The networking racks, for instance, are strategically positioned to minimize the length of fiber optic cabling between the four cats that make up this cluster. Two in this room and two in the next one over. This is both to maintain signal integrity. They found any higher than 100 to 150 m can be problematic at these speeds, but it's also to maintain the integrity of the ceiling. See, it turns out that when you start bundling up hundreds of fiber optic cables, they can get pretty heavy. And the longer the runs, the more strain they put on the ceiling. It also helps save on cost. This I don't have much to say about other than Oh my god, isn't it beautiful? Okay, I lied. I do have some stuff to say. This cable management is not just for looks, but rather to maintain air flow. When you have this many cables, you actually do need to bundle them. And it also helps facilitate maintenance in the event of a broken or damaged cable. They run extra fiber in each bundle that can be terminated as needed, and it is a lot easier to find that when things are organized. [music] Careful thought goes into rack serviceability, too. While the networking racks use traditional vertical PDUs for their power distribution, the DGX racks use three top-mounted PDUs that help balance the three [music] phases of power coming in. On the subject of power, we skipped over these at the SFU data center, but uh have you ever wondered what a >> 415 volt 100 amp >> power plug looks like? Wonder no longer. Another key advantage of putting all the PDUs up top is it makes it a little easier to get the DGX units in and out. They uh turn this one off and agreed to pull it out of the deployment just to show us how it's done. But when I asked if we could tear it down, they were like, \"Huh?\" Not because they have anything to hide, but just cuz they thought this soon to be decommissioned and recycled engineering unit that we can really get our greasy meat mitts into would be a lot more [music] fun. Woo! If you ever wondered what the heat sink would look like on a 1200 W GPU, wonder no more. How many heat pipes is this? It's like a forest of heat pipes in there. Oh, they're flat. That makes so much sense cuz you'll get more air flow through the fins. And I mean, that's some pretty hard working air dealing with not one but two rows of these GPUs. Not to mention this middle heat sink here that seems to be for the Envy Link switching equipment. I mean, I knew you need some beefy to allow these GPUs to pull the 192 gigs of HBM 3 memory that they have, but still, I'd never seen the cooling for it. Another thing I've never gotten to see, or at least never gotten to film, is NVIDIA's proprietary SXM interface. I know because I was looking for a B-roll shot of one of these a little while ago, and I realized I've never actually held one up on camera. So, now I've done it. And they also gave us this to show what Okay, not the same generation, but a similar GPU might look like without the heat sink. Oh, we've also got to look at the SXM interface that goes back to the rest of the system and also carries power for these [music] GPUs. These are 1,200 W each, but these cards go as high as 1,400 W. I mean, how high would the cooler be at that point? JK, it would be water cooled. We were hoping to see some water cooled machines today, but I guess NVIDIA has to save [music] something for next time. I got to say though, if the new stuff is even just equally cool to the last gen H100 water cooled machines we saw at the SFU Fur Supercomput, I'm sure it's absolutely mind-blowing. All of which is really cool, but what do they need all of this for? Well, the folks who are responsible for getting us access to everything today are actually from the GeForce gaming team. So, one of their big uses for these internal compute resources is, of course, deep learning superers sampling or DLSS. In a nutshell, DLSS allows your GPU to render your game at a lower resolution, then use deep learning or AI to upscale each output frame to your monitor's resolution. The benefit of this is that you can run at a higher frame rate for improved animation smoothness. But the drawback is that you can't create something from nothing and upscaled images struggle to achieve the same fidelity as a native rendered image. With that said, DLSS has improved a lot over the years and I got a chance to sit down and chat with Edward Leu from that team who talked us through some of the processes that they use to develop new features and new fixes. The first thing he pointed out is that while this hot new AI model took a thousand GPUs two weeks to train or whatever, makes for a good headline, it overlooks most of the actual cost and time, which is in the test training runs that take place before the hero run. His team is constantly iterating on the data they're feeding into DLSS and [music] the waiting, and they're evaluating new innovations in the AI space. Sometimes it's more surgical, like, \"Oh, hey, uh, we noticed Cyberpunk has an issue with cars having like three or four bumpers as you're driving around. How can we address this with the current model?\" That kind of thing can be turned around sometimes in a matter of weeks or months using the resources that we just saw. Though, uh, he was quick to point out that Jensen, if you're watching, there's no limit to how many GPUs his team could use to speed up the process. I told him I'd say that because the faster they can test a new data set, the faster they can tune it and ship it. Other times, the changes are more transformative, pun intended, like the move from a convolutional neural network to the more accurate transformer model that runs best on NVIDIA's newest cards. This can require basically a complete tear down and doover of the entire pipeline and can take a year or more. But I mean, hey, NVIDIA has bet pretty much their entire future on being able to stay at the forefront of their 21st century shovel technology. So, I guess that's the price you pay. But what does all this mean for traditional rendering? Well, the folks here believe very strongly that in time DLSS will not only be as good, it will be better than [music] traditional. In fact, Edward pointed to these slides from a talk that he gave a number of years ago, showing that what we think of as native rendering is already a pretty imperfect approximation. Here it is compared to a ground truth image, which is rendered at a much higher resolution than downsampled to 1080p. Native, it's clearly worse. And as you can see in this comparison, even back then with DLSS 2.0, So there were situations where a model trained on these higher resolution or more accurately higher sampled ground truth images could result in the GPU reconstructing an output with DLSS that's closer to ground truth [music] than the native image was. I asked, by the way, how are these ground truth images created? [music] and he said that typically it's actually just on run-of-the-mill gaming hardware, but instead of running at 60 or 120 frames per second, they'll sometimes be running tens or thousands of pixel samples to the point where we're talking more like 60 to 120 seconds per frame or even more. Man though, if DLSS could consistently reconstruct that ground truth image from a lower resolution input every time, I'm sure no one would ever turn it off. Unfortunately for Edward and his team though, no one notices when DLSS is working perfectly. It's when it trips over itself that we tend to notice, and it still does do so fairly regularly. However, with that said, from my own experience, it has continued to improve at a pretty solid clip since its debut, and the rest of the industry has pretty much accepted that AI accelerated image enhancement is the future, whether every gamer wants it or not. So sales of shovels will likely continue until gamer morale improves. Good luck everyone. I hope it's not a bubble. This is a nice office. I'd hate for something to happen to it. Just like this is a nice segue to our sponsor, Squarespace. If you're building a brand or a business, it's important to have a website. Designing your own site can feel like a bit of a daunting task, but it really doesn't have to. Squarespace makes it easy to get your message across to potential customers and subscribers in a clean and digestible way. Their design intelligence tool utilizes AI and works with your own creativity to create a theme and style for your site that matches your personality. You can also use Squarespace to directly invoice your customers with payment options like a direct debit, Apple Pay, CLA, and more. And use their analytic tools to track sales, strategize, and continue to build your brand. There's millions of URLs available, and Squarespace's domain tool will help you search for the right address just for you. We've even used Squarespace for some of our own websites here. Start building your website today, and you'll receive 10% off your first purchase by visiting squarespace.com/LTT. If you guys enjoyed this video, maybe go check out the time we I don't know, something invidated. Let's do a throwback video. How about when I checked out the launch of G-Sync? The production values were lower, but hey, it was fun."}