WEBVTT 00:00:00.240 --> 00:00:06.560 We are here at the Enterprise Data 00:00:03.560 --> 00:00:12.960 Center/ Supercomputer Lab at AMD's Sunnyvale headquarters where they have 00:00:09.040 --> 00:00:15.280 got the definition of an all hands on 00:00:12.960 --> 00:00:20.640 deck project going on here. They've flown in their techs and partners from 00:00:17.600 --> 00:00:23.039 around the world. Even Mark Hurst, 00:00:20.640 --> 00:00:28.000 corporate vice president of Radeon Technologies Group, is getting in on the 00:00:25.199 --> 00:00:35.120 action, planning the power runs for one of the most exceptional supercomput 00:00:30.720 --> 00:00:38.800 systems ever created. When it's full of 00:00:35.120 --> 00:00:43.480 20 servers, this single rack is going to 00:00:38.800 --> 00:00:47.239 be capable of one path flop of full 00:00:43.480 --> 00:00:49.760 precision performance. And AMD 00:00:47.239 --> 00:00:54.719 estimates, it's an estimate, they're not done building it yet, that they'll be 00:00:51.520 --> 00:00:57.520 doing this at around half the cost of a 00:00:54.719 --> 00:01:04.720 competing solution, not to mention more than double the flops per 00:01:01.160 --> 00:01:06.360 watt. And how are they doing such a 00:01:04.720 --> 00:01:12.479 thing? Well, as crazy as it might 00:01:09.560 --> 00:01:20.560 sound, AMD is actually letting me assemble one of the systems. So, uh, 00:01:16.159 --> 00:01:20.560 well, I'm going to show you guys exactly 00:01:27.880 --> 00:01:33.680 how Synergy allows you to share your 00:01:31.040 --> 00:01:39.040 mouse and keyboard between multiple computers at once. Check it out now at 00:01:36.000 --> 00:01:41.200 the link in the video description. 00:01:39.040 --> 00:01:48.479 So, pretty much everything you guys see in this room arrived about 12 hours ago 00:01:44.479 --> 00:01:52.479 at 3:46 in the morning from AMD partners 00:01:48.479 --> 00:01:55.119 like Invente, Samsung, and Melanox and 00:01:52.479 --> 00:02:01.840 has been undergoing assembly and testing non-stop since then. So, it can be 00:01:57.360 --> 00:02:04.280 unveiled at Sigraph in less than a week. 00:02:01.840 --> 00:02:10.360 So, this video is as much a behindthescenes look at the hard work 00:02:06.880 --> 00:02:14.200 that goes into the glitzy polished show 00:02:10.360 --> 00:02:16.879 demonstrations as it is about this 00:02:14.200 --> 00:02:22.080 system. Who am I kidding? We're here for the beastly hardware. So, this in front 00:02:19.599 --> 00:02:27.440 of me is what AMD and Invente are calling the 00:02:24.040 --> 00:02:29.560 P47. And it's this that allows them to 00:02:27.440 --> 00:02:39.760 cram up to 640 CPU cores. That's,280 threads if you 00:02:34.000 --> 00:02:44.879 prefer up to 40 terabytes of memory and 00:02:39.760 --> 00:02:48.560 up to 80 Radeon Instinct Vega GPUs into 00:02:44.879 --> 00:02:52.160 a single rack with some room for 00:02:48.560 --> 00:02:54.920 high-speed networking like Infiniban. At 00:02:52.160 --> 00:03:01.280 a total of about 35,000 watts of power consumption, this 00:02:58.640 --> 00:03:07.200 is about as power dense as you would want to go with air cooling using a rear 00:03:04.640 --> 00:03:13.519 door cooling system like what we saw at the SFU Supercomput recently. So, let's 00:03:11.120 --> 00:03:18.480 take a closer look at it. At the front, you've got some storage bays like you'd 00:03:16.080 --> 00:03:24.239 expect to find on a server. So, I don't think AMD or Invecting anybody to use 00:03:21.680 --> 00:03:30.000 these for more than just like a basic scratch disc. And then on the back is 00:03:27.519 --> 00:03:36.400 the reason for it because that's where the 100 gigabit Infiniband connections 00:03:33.360 --> 00:03:40.080 go that handle job sharing between 00:03:36.400 --> 00:03:42.400 servers as well as access to high-speed 00:03:40.080 --> 00:03:48.640 dedicated storage machines that are elsewhere on the network. 00:03:44.720 --> 00:03:52.040 Moving inside, at the heart of every P47 00:03:48.640 --> 00:03:56.400 is the brand new AMD Epic server 00:03:52.040 --> 00:04:02.720 processor. This one is 32 cores on a 00:03:56.400 --> 00:04:06.239 single CPU with a 180 watt TDP. 32 cores 00:04:02.720 --> 00:04:09.920 at 180 watts. But that's actually not 00:04:06.239 --> 00:04:13.360 the whole story behind what gives P47 00:04:09.920 --> 00:04:16.560 its stunning efficiency. Each epic can 00:04:13.360 --> 00:04:18.320 handle up to two terabytes of DDR4 00:04:16.560 --> 00:04:24.880 memory running in an 8 channel configuration at speeds of up to 2666 00:04:21.600 --> 00:04:27.440 MHz. So that's great for virtualization 00:04:24.880 --> 00:04:36.000 uh rendering and machine intelligence. Our machines have 512 gigs of RAM each. 00:04:31.280 --> 00:04:38.560 But, and this is the key, each epic can 00:04:36.000 --> 00:04:48.800 also handle 128 PCI Express 00:04:44.120 --> 00:04:51.680 lanes. 128 lanes. So, that's why a 00:04:48.800 --> 00:05:02.240 single Epic can be hooked up to up to six or even seven GPUs with a full PCIe 00:04:58.400 --> 00:05:05.040 16X Gen 3 interface. dedicated to 00:05:02.240 --> 00:05:10.880 everyone without a need for a second processor or a PLX chip. And that's with 00:05:08.479 --> 00:05:17.280 plenty of bandwidth left over for high-speed NVMe storage or high-speed 00:05:14.280 --> 00:05:19.600 networking. And all that's critical 00:05:17.280 --> 00:05:26.479 because the real star of the show today is the Radeon Instinct 00:05:22.520 --> 00:05:28.800 Mi25 based on AMD's Vega architecture. 00:05:26.479 --> 00:05:37.280 These puppies right here are actually responsible for the bulk of that pedlop 00:05:32.639 --> 00:05:40.639 of processing power. Every MI25 is a 00:05:37.280 --> 00:05:42.880 full fat Vega consuming 300 watts. 00:05:40.639 --> 00:05:48.639 Actually, similar to the Frontier Edition that already exists and the RX 00:05:45.520 --> 00:05:53.120 Vega that's coming soon. But unlike 00:05:48.639 --> 00:05:56.560 those, it lacks video outputs entirely 00:05:53.120 --> 00:05:59.440 and a fan since it was designed to be 00:05:56.560 --> 00:06:05.759 installed in specialized machines like this one where the cooling is 00:06:01.680 --> 00:06:10.000 integrated. And this is sick. The entire 00:06:05.759 --> 00:06:13.039 16 gigs of HBM2 memory can be accessed 00:06:10.000 --> 00:06:14.960 by another device on the PCI Express 00:06:13.039 --> 00:06:23.520 bus. This allows high-speed communication between discrete Radeon 00:06:18.880 --> 00:06:25.680 Instinct MI25s or even other PCI Express 00:06:23.520 --> 00:06:31.600 devices that are connected to the system like even an NVMe storage device. And 00:06:28.319 --> 00:06:35.039 this is without a proprietary connection 00:06:31.600 --> 00:06:37.759 like ND link and all this high-speed 00:06:35.039 --> 00:06:44.880 communication enabled by Epic, which you can kind of think of as both a CPU and 00:06:41.520 --> 00:06:48.319 the world's biggest PCI Express switch 00:06:44.880 --> 00:06:51.360 is at the core of a new opensource 00:06:48.319 --> 00:06:54.800 architecture that AMD is championing 00:06:51.360 --> 00:06:59.120 called Rockum. In a nutshell, the idea 00:06:54.800 --> 00:07:03.560 is that every part of an all AMD server 00:06:59.120 --> 00:07:06.800 system is linked at extreme speeds in a 00:07:03.560 --> 00:07:09.840 nonproprietary manner, so the user can 00:07:06.800 --> 00:07:11.080 customize and scale up their supercomput 00:07:09.840 --> 00:07:17.479 as they see fit. Want more performance? Just add 00:07:14.479 --> 00:07:21.560 more racks. And while 00:07:17.479 --> 00:07:24.639 $50,000 might seem like a lot for a 00:07:21.560 --> 00:07:26.960 P47, when you consider that AMD's 00:07:24.639 --> 00:07:30.639 initial estimates put them comfortably in a position to claim better 00:07:28.880 --> 00:07:36.800 performance per dollar for modern workloads like financial modeling, 00:07:33.199 --> 00:07:39.599 climate science, and AI, not to mention 00:07:36.800 --> 00:07:46.479 better performance per watt. You can see where their confidence about uh really 00:07:43.120 --> 00:07:48.599 shaking up the top 100 supercomputer 00:07:46.479 --> 00:07:53.319 list comes from and I wish them the absolute best 00:07:52.560 --> 00:07:59.199 of luck. Do you have two computers for some 00:07:56.319 --> 00:08:04.400 reason? Maybe one of them's PC and one of them's Mac or Linux or whatever. 00:08:01.680 --> 00:08:09.280 Well, Synergy lets you solve the problem of having separate keyboards and mice 00:08:06.319 --> 00:08:14.960 for them once and for all. You can share one mouse and one keyboard between two 00:08:12.160 --> 00:08:20.000 or even more computers, so you'll never get confused again. They've got basic 00:08:17.440 --> 00:08:23.840 and pro options for synergy with a one-time payment. And features include 00:08:22.000 --> 00:08:28.479 things like clipboard sharing between the computers, dragging and dropping 00:08:26.479 --> 00:08:33.839 files between the computers, the ability to set up hotkeys, and more. Use our 00:08:31.680 --> 00:08:38.880 link in the video description and get 50% off Synergy 00:08:36.440 --> 00:08:44.959 today. So, thanks for checking out our video here at AMD Supercomput Lab. And 00:08:43.279 --> 00:08:48.320 uh if you guys dislike this video, you can hit that button. But if you liked 00:08:46.320 --> 00:08:51.399 it, hit the like button, get subscribed, maybe consider checking out where to buy 00:08:50.000 --> 00:08:55.279 the stuff we featured at the link in the video 00:08:53.519 --> 00:08:58.480 description. Also down there, we've got a link to our t-shirt store, which has 00:08:56.720 --> 00:09:02.000 cool shirts like this one, as well as our community forum, where you can go 00:09:00.240 --> 00:09:05.360 and talk about all the cool stuff you saw today.