WEBVTT 00:00:00.000 --> 00:00:04.400 Although transistors still keep on shrinking, it's getting more and more difficult to pack 00:00:04.400 --> 00:00:09.200 as many of them on a chip as we'd like. Partial solutions to this, such as using chiplets 00:00:09.200 --> 00:00:13.840 to reduce the rate of manufacturing defects and stacking transistors on top of each other 00:00:13.840 --> 00:00:17.360 have been in vogue for a while now, but it might not be too surprising 00:00:17.360 --> 00:00:22.400 that some manufacturers have decided to simply make the chip themselves bigger, 00:00:22.400 --> 00:00:27.120 when in doubt, supersize. Now, I'm not saying that your next computer 00:00:27.120 --> 00:00:30.160 might have a CPU that's so big it'll take up half the motherboard, 00:00:30.160 --> 00:00:33.360 but when you get away from personal computers and start looking at chips 00:00:33.360 --> 00:00:39.040 that we might see in data centers in the near future, you start seeing some pretty eye-watering stuff. 00:00:39.040 --> 00:00:43.040 We're talking about designs like the Wafer Scale Engine 2 from Cerebrus, 00:00:43.040 --> 00:00:47.280 currently the largest chip in the world. Built on a seven nanometer process, 00:00:47.280 --> 00:00:51.840 it contains 850,000 cores 00:00:51.840 --> 00:00:56.800 and is a whopping 21.5 centimeters or 8.5 inches long. 00:00:56.800 --> 00:01:00.640 That's more total area than 25 Ryzen desktop CPUs. 00:01:00.640 --> 00:01:04.880 Perhaps unsurprisingly, a chip this big and with this many transistors, 00:01:04.880 --> 00:01:09.280 2.6 trillion, to be exact, requires a lot of power. 00:01:09.280 --> 00:01:12.800 The Wafer Scale Engine 2 sucks down 15 kilowatts, 00:01:12.800 --> 00:01:16.800 so if you were somehow able to drop this into your PC, 00:01:16.800 --> 00:01:20.000 you'd need 15 1000 watt power supplies 00:01:20.000 --> 00:01:23.440 just to keep it fit. And that's not even counting the rest of the system. 00:01:23.440 --> 00:01:28.480 But despite this, the new design should actually result in power savings. 00:01:28.480 --> 00:01:32.800 You see, data centers and supercomputers that do artificial intelligence processing 00:01:32.800 --> 00:01:38.400 often have to use lots of separate chips, such as GPUs, spread across a large facility. 00:01:38.400 --> 00:01:43.760 Having the same amount of computing power on just one physical chip is far more power efficient, 00:01:43.760 --> 00:01:48.800 even if the power consumption rating of that chip is a lot higher than a typical GPU. 00:01:48.800 --> 00:01:52.880 But there are other advantages to this approach, besides just saving energy. 00:01:52.960 --> 00:01:57.760 You might be wondering why we aren't simply just sticking a bunch of chiplets onto one package instead 00:01:57.760 --> 00:02:02.000 to make something like a really big version of an AMD Epic processor. 00:02:02.000 --> 00:02:06.880 So as versatile as chiplets have been, they still suffer from having more latency 00:02:06.880 --> 00:02:11.200 than one big monolithic processor. The little interconnects that move data 00:02:11.200 --> 00:02:16.160 between chiplets as quick as they may be, and they are fast, are still slower 00:02:16.160 --> 00:02:21.600 than if you physically put computing units directly adjacent to each other to form one big chip. 00:02:21.600 --> 00:02:26.560 Ultimately, this means that huge monolithic chips can process more data than a system 00:02:26.560 --> 00:02:30.400 with the same number of transistors spread out among multiple chips. 00:02:30.400 --> 00:02:34.720 And when you consider just how much data has to be processed for AI applications 00:02:34.720 --> 00:02:38.400 and scientific research, it makes a difference. 00:02:38.400 --> 00:02:41.760 Wave for scale technology has already drawn interest from diverse industries, 00:02:41.760 --> 00:02:46.320 including national intelligence and healthcare, but though it has some obvious advantages, 00:02:46.320 --> 00:02:51.200 that doesn't necessarily mean that it's the silver bullet to large-scale compute challenges. 00:02:51.200 --> 00:02:56.720 For instance, one big issue is the fact that these processors are designed to handle lots of data, 00:02:56.720 --> 00:03:02.160 so they also need access to a lot of memory, and designs with chiplets and larger amounts of memory 00:03:02.160 --> 00:03:05.360 built onto the same package may end up being more popular. 00:03:05.360 --> 00:03:09.600 This is similar to what Tesla has done with their new D1 chip, which was developed in part 00:03:09.600 --> 00:03:12.960 to help propel Tesla's self-driving AI technology. 00:03:12.960 --> 00:03:16.800 The D1 chip itself is much smaller than the Wave for Scale engine, 00:03:16.800 --> 00:03:23.680 but Tesla has included over 11 gigabytes of high-speed SRAM in an arrangement of 25 D1 chips 00:03:23.680 --> 00:03:28.400 connected together to make a training tile that's bigger than your head. 00:03:28.400 --> 00:03:33.760 And of course, making smaller chips reduces the amount of silicon you'll waste due to manufacturing errors, 00:03:33.760 --> 00:03:37.600 as we mentioned earlier. But regardless of whether a particular company 00:03:37.600 --> 00:03:40.640 is using Wave for Scale or an arrangement more like Tesla's, 00:03:40.640 --> 00:03:45.200 putting a large amount of silicon on one plane may end up becoming an industry trend. 00:03:45.200 --> 00:03:50.720 Because if America has taught us anything, it's that there's a deep human need to supersize. 00:03:50.720 --> 00:03:54.320 And if you feel the need, like the video or dislike the video, 00:03:54.320 --> 00:03:57.440 check out our other videos and comment below with video suggestions. 00:03:57.440 --> 00:04:00.640 We make videos here. Get your videos here at TechQuickie. 00:04:00.640 --> 00:04:02.320 Don't forget to subscribe and follow. 00:04:04.320 --> 00:04:05.040 See you later.