WEBVTT - What are AI Chips? 0:00:04.480 --> 0:00:12.400 Welcome to tech Stuff, a production from iHeartRadio. Hey there, 0:00:12.400 --> 0:00:15.880 and welcome to tech Stuff. I'm your host, Jonathan Strickland. 0:00:15.920 --> 0:00:19.439 I'm an executive producer with iHeart Podcasts. And how the 0:00:19.520 --> 0:00:22.600 tech are you? Now? There's a pretty good chance that 0:00:22.800 --> 0:00:27.880 you've heard or read something about AI chips. But what 0:00:27.920 --> 0:00:31.960 the heck is an AI chip? Is it a microchip 0:00:32.000 --> 0:00:36.720 that actually has artificial intelligence incorporated directly into the semiconductor 0:00:36.800 --> 0:00:41.040 material somehow? And if so, what does that mean? I 0:00:41.040 --> 0:00:42.920 figured it would be a good idea to talk about 0:00:43.159 --> 0:00:48.159 microchips and processors and AI enabled chips in particular to 0:00:48.240 --> 0:00:53.000 help demystify everything because part of the problem, I think 0:00:53.200 --> 0:00:55.880 is that AI chips are are kind of becoming a 0:00:55.920 --> 0:01:00.360 marketing term. It's not just a way to describe technology. 0:01:00.440 --> 0:01:03.120 It's a way to try and set aside a product 0:01:03.400 --> 0:01:06.680 to try and you know, pose it as the new hotness. 0:01:07.000 --> 0:01:10.760 And yes, I know I'm ancient and I use outdated slang. 0:01:11.200 --> 0:01:14.600 So first, when I talk about microchips, I'm talking about 0:01:14.840 --> 0:01:19.600 integrated circuits. Jack Kilby invented the first integrated circuit back 0:01:19.640 --> 0:01:24.240 in nineteen fifty eight at Texas Instruments, and an integrated 0:01:24.280 --> 0:01:29.000 circuit is a collection of interconnected electronic components that happens 0:01:29.040 --> 0:01:32.959 to be built on top of a semiconductor material. Semiconductors, 0:01:32.959 --> 0:01:36.959 as the name suggests, our materials that under certain conditions 0:01:37.000 --> 0:01:41.760 will conduct electricity and under other conditions will insulate or 0:01:41.920 --> 0:01:45.959 block the flow of electricity. The invention of the transistor, 0:01:46.280 --> 0:01:51.080 in addition to the integrated circuit is what allowed for miniaturization. 0:01:51.520 --> 0:01:54.840 That's why computers no longer have to be huge, you know. 0:01:54.880 --> 0:01:58.200 I'm talking about like those old computers, those mainframes that 0:01:58.240 --> 0:02:01.520 would fill up entire rooms or even an entire floor 0:02:01.600 --> 0:02:05.360 of a building. Miniaturization would eventually allow for the production 0:02:05.480 --> 0:02:09.000 of powerful personal computers that were a fraction of the 0:02:09.080 --> 0:02:12.640 size of their predecessors, but just as powerful or even 0:02:12.680 --> 0:02:18.639 more so. The development of arithmetic logic units or ALUs, which, 0:02:18.639 --> 0:02:22.640 as the name suggests, are circuits designed to perform arithmetic 0:02:22.919 --> 0:02:28.600 or mathematical functions on inputs and then produce the relevant outputs. Those, 0:02:28.919 --> 0:02:32.959 in turn served as a building block for the development 0:02:33.200 --> 0:02:39.400 of central processing units or CPUs. The first CPU microprocessor, 0:02:39.919 --> 0:02:45.200 arguably was Intel's for zero zero four computer on a CHEP. 0:02:45.480 --> 0:02:49.360 So this was a fairly limited processor, particularly if we 0:02:49.480 --> 0:02:52.000 judge it by today's standards. I think it had like 0:02:52.040 --> 0:02:56.440 a four bitwidth bus that would allow for processing data, 0:02:56.480 --> 0:03:00.240 which means it could not handle very large values. But 0:03:00.600 --> 0:03:04.640 you have to start somewhere, and the four zero zero 0:03:04.760 --> 0:03:08.359 four was a stepping stone to Intel's eight zero zero 0:03:08.520 --> 0:03:12.280 eight processor. That was the processor that was found in 0:03:12.320 --> 0:03:15.600 a lot of the first commercial personal computers. They used 0:03:15.880 --> 0:03:20.600 the eight zero zero eight processor as their CPU. Now, 0:03:20.680 --> 0:03:25.520 a central processing unit's job is more complex than that 0:03:25.840 --> 0:03:30.840 of an ALU. In fact, ALUs are part of a CPU. 0:03:30.960 --> 0:03:34.560 They are a component that make up part of a CPU. 0:03:35.040 --> 0:03:40.160 The CPU's job is to accept incoming instructions from programs, 0:03:40.200 --> 0:03:45.720 to retrieve those instructions, to execute those instructions on data, 0:03:46.080 --> 0:03:50.800 and to produce the relevant outcomes. And they carry out 0:03:50.840 --> 0:03:54.760 logic operations. They send results to the appropriate destination. That 0:03:54.800 --> 0:03:59.960 destination might be feeding back into software to continue that process, 0:04:00.320 --> 0:04:02.960 or it might mean that you're feeding output to some 0:04:03.120 --> 0:04:06.480 sort of output device like a display or a printer 0:04:06.720 --> 0:04:11.200 or something along those lines. CPUs have two very broad 0:04:11.680 --> 0:04:15.720 categories of operations. Again, this is super super high level. 0:04:15.760 --> 0:04:18.880 I mean, we could get far more complicated than this, 0:04:19.200 --> 0:04:23.760 but they have logic functions and memory functions. Memory being 0:04:24.040 --> 0:04:27.000 that's where you store information so that you can reference 0:04:27.040 --> 0:04:30.280 it quickly in order to carry out these operations, and 0:04:30.360 --> 0:04:33.840 logic being the actual logic gates that end up defining 0:04:33.960 --> 0:04:37.800 how data gets processed. Those are the two big components, 0:04:38.120 --> 0:04:40.640 and there are a few different ways that we measure 0:04:40.720 --> 0:04:44.159 CPU performance. One is we measure it by clock speed. 0:04:44.560 --> 0:04:46.880 You can think of this as the number of instructions 0:04:46.920 --> 0:04:50.560 the CPU is able to handle every second. So the 0:04:50.680 --> 0:04:54.240 higher the clock speed, the more instructions the CPU is 0:04:54.279 --> 0:04:57.080 able to handle per second. That like three point six 0:04:57.120 --> 0:05:01.400 gigahertz would mean three point six billion operations or instructions 0:05:01.400 --> 0:05:05.000 I should say per second. You can also have operations 0:05:05.000 --> 0:05:08.440 that have multiple sets of instructions, so it's a little 0:05:08.480 --> 0:05:11.240 more complicated than just saying, oh, it can handle this 0:05:11.320 --> 0:05:14.200 number of operations per second. Now, you can also have 0:05:14.279 --> 0:05:18.360 CPUs that have multiple cores, and a core is essentially 0:05:18.839 --> 0:05:23.360 all the little individual components of a CPU compartmentalized so 0:05:23.400 --> 0:05:27.719 that you have almost like multiple CPUs on a single chip. 0:05:27.960 --> 0:05:32.680 A single core processor is like a really fast processor. 0:05:33.040 --> 0:05:37.640 A multicore processor is one that divides the processor capabilities 0:05:37.680 --> 0:05:40.400 into these individual cores, and you might wonder, well, why 0:05:40.400 --> 0:05:42.040 would you want to do that, Why would you want 0:05:42.080 --> 0:05:45.159 to take something that is typically very powerful and very 0:05:45.200 --> 0:05:49.320 fast and then divide that up into smaller units. Well, 0:05:49.320 --> 0:05:53.920 that's because some computational processes are able to be performed 0:05:53.960 --> 0:05:58.159 in parallel. This means you can divide up a task 0:05:58.560 --> 0:06:02.680 into smaller jobs and then assign those smaller jobs to 0:06:02.760 --> 0:06:06.360 individual cores. So for these kinds of processes, a multi 0:06:06.440 --> 0:06:11.320 core processor can sometimes be faster than a more powerful 0:06:11.440 --> 0:06:14.720 single core processor would, And that means it's time to 0:06:14.800 --> 0:06:17.920 use an analogy. I bust out every time I talk 0:06:17.960 --> 0:06:21.240 about parallel processing. Fans of tech stuff who have been 0:06:21.279 --> 0:06:23.520 around for years, you all know what's going to happen. 0:06:23.560 --> 0:06:26.120 Go ahead and make yourself a cup of coffee or something. 0:06:26.320 --> 0:06:30.080 So imagine you have a math class. You're a teacher. 0:06:30.320 --> 0:06:33.080 You've got a math class, and your math class has 0:06:33.440 --> 0:06:36.480 five students in it. It's very small focus group. One 0:06:36.520 --> 0:06:40.520 of those five students is like a super math genius. 0:06:40.760 --> 0:06:43.480 They are leagues ahead of the other students. The other 0:06:43.560 --> 0:06:47.640 four students are good at math, they're great students, but 0:06:47.880 --> 0:06:51.040 they take a little more time than the genius does 0:06:51.120 --> 0:06:55.080 to work out Your typical math problem. So you decide 0:06:55.120 --> 0:06:58.080 you're going to give a pop quiz to your class. 0:06:58.200 --> 0:07:01.480 But this pop quiz is a race. It's a race 0:07:01.520 --> 0:07:06.960 that is pitting the super genius against the other four students. Now, 0:07:07.279 --> 0:07:11.440 if that pop quiz consisted of just one mathematical problem, 0:07:12.000 --> 0:07:14.880 or if it had a series of math problems, but 0:07:14.920 --> 0:07:18.440 those math problems were sequential, which means like the information 0:07:18.520 --> 0:07:21.240 you need to solve question two can be found in 0:07:21.280 --> 0:07:24.360 the answer of question one. If that were the case, 0:07:24.760 --> 0:07:28.400 your super genius is gonna win, right because they would 0:07:28.440 --> 0:07:31.680 be able to solve the problem or series of problems 0:07:31.800 --> 0:07:35.239 much more quickly than anyone in the rest of the class. 0:07:35.480 --> 0:07:37.760 And you can't divide that problem up. If it's a 0:07:37.800 --> 0:07:40.480 series that you know question two depends on the outcome 0:07:40.480 --> 0:07:43.040 of question one, you can't divide that up because you 0:07:43.040 --> 0:07:45.400 wouldn't have the information you need to work on the 0:07:45.440 --> 0:07:49.800 problem until the first part was solved. However, let's say 0:07:49.840 --> 0:07:54.200 instead you make a pop quiz that has four math 0:07:54.240 --> 0:07:57.600 problems on it. Each of these four math problems is 0:07:57.640 --> 0:08:00.600 self contained. They do not depend upon the outcomes of 0:08:00.680 --> 0:08:03.480 any of the other questions. So the super genius needs 0:08:03.520 --> 0:08:07.680 to finish all four problems. But for your other four students, 0:08:07.840 --> 0:08:11.120 they're given the option that they can each tackle a 0:08:11.280 --> 0:08:14.480 different problem on the quiz, and if all four of 0:08:14.520 --> 0:08:19.480 them finish whatever respective problem they've chosen first, then as 0:08:19.520 --> 0:08:22.720 a group they win. Now, in that case, the four 0:08:22.760 --> 0:08:25.160 students are far more likely to come out on top. 0:08:25.360 --> 0:08:28.360 The super genius could be as far as like question 0:08:28.480 --> 0:08:31.320 three or four, But each of the other students only 0:08:31.360 --> 0:08:34.160 has to solve a single problem in order to complete 0:08:34.200 --> 0:08:38.080 the pop quiz. That's like a multi core processor working 0:08:38.080 --> 0:08:43.439 on a parallel processing problem. For some subsets of computational operations, 0:08:43.440 --> 0:08:46.280 having multiple cores to work on things all at the 0:08:46.360 --> 0:08:49.240 same time is a big advantage, all right. So that's 0:08:49.280 --> 0:08:53.800 a super high level look at CPUs. Now let's turn 0:08:53.880 --> 0:08:58.880 to GPUs. These are graphics processing units. The name actually 0:08:58.880 --> 0:09:02.520 comes from the g Force two fifty six graphics card 0:09:02.559 --> 0:09:06.600 from Nvidia. So in the nineteen nineties we saw the 0:09:06.640 --> 0:09:11.920 introduction of new graphics intensive applications, particularly in things like 0:09:12.080 --> 0:09:16.439 video editing or in video games, and the CPUs of 0:09:16.480 --> 0:09:20.920 that era were not necessarily optimized to get the job done, 0:09:20.960 --> 0:09:24.120 like it was more work than the CPU could typically handle. 0:09:24.480 --> 0:09:29.400 So the performance of these kinds of programs would be substandards. 0:09:29.400 --> 0:09:32.320 Sometimes the programs wouldn't even run on a computer that 0:09:32.480 --> 0:09:35.600 just had a CPU, even a good CPU. So then 0:09:35.640 --> 0:09:39.239 you had companies like in Video that began to introduce 0:09:39.400 --> 0:09:43.200 graphics cards, and these graphics cards had integrated circuits that 0:09:43.240 --> 0:09:48.960 were better designed. They were optimized to handle graphics processing specifically, 0:09:49.240 --> 0:09:52.559 so that would let the CPU offload the graphics processing 0:09:52.679 --> 0:09:55.800 jobs to the graphics card. The CPU could then focus 0:09:55.840 --> 0:09:59.840 on other operations. The g Force two fifty six introduced 0:09:59.840 --> 0:10:03.440 a ton of new capabilities and features. And while the 0:10:03.480 --> 0:10:06.760 graphics processing unit name might have just started off as 0:10:06.840 --> 0:10:09.920 kind of a marketing strategy, you know, Nvidia gave the 0:10:09.920 --> 0:10:12.840 g Force two fifty six this designation of graphics processing 0:10:12.920 --> 0:10:15.280 unit to kind of set it apart from other graphics 0:10:15.320 --> 0:10:17.600 cards that were on the market. Well, it would turn 0:10:17.640 --> 0:10:21.920 out that the GPU name would have staying power, and 0:10:21.960 --> 0:10:26.080 today any self respecting gamer has a powerful GPU in 0:10:26.160 --> 0:10:29.200 their gaming rig. The GPUs would grow to be more 0:10:29.240 --> 0:10:32.640 important than CPUs, at least for some people. Though it 0:10:32.640 --> 0:10:35.560 would be reductive to say that gamers only need a 0:10:35.679 --> 0:10:38.200 powerful GPU and they don't have to worry about the 0:10:38.240 --> 0:10:40.960 CPU at all. It honestly depends a lot on the 0:10:41.000 --> 0:10:44.360 types of games they play. That is a big component. 0:10:44.440 --> 0:10:47.440 Sometimes having a really fast GPU isn't going to help 0:10:47.440 --> 0:10:49.600 you out that much. It all depends on the types 0:10:49.640 --> 0:10:51.880 of processing you're doing. If you're not doing a lot 0:10:51.880 --> 0:10:56.200 of parallel processing, then a really fast GPU isn't likely 0:10:56.280 --> 0:10:59.440 to boost your performance that much. But the real purpose 0:10:59.480 --> 0:11:02.760 of a GPU is to perform certain types of computational 0:11:02.800 --> 0:11:06.920 operations very quickly and efficiently, in order to do stuff 0:11:06.960 --> 0:11:11.680 like speed up image creation, video and animation. As it 0:11:11.679 --> 0:11:15.160 would turn out, GPUs would also be handy for other things. 0:11:15.440 --> 0:11:20.040 So your typical GPU consists of many specialized processor cores. 0:11:20.360 --> 0:11:23.200 These cores are not designed to do everything you know. 0:11:23.240 --> 0:11:26.360 They do a subset of operations really well, but if 0:11:26.360 --> 0:11:29.000 you ask a GPU to do something outside of that, 0:11:29.080 --> 0:11:32.000 it's not going to perform at you know, at the 0:11:32.000 --> 0:11:35.000 same level as your typical CPU would. But this does 0:11:35.080 --> 0:11:38.800 mean a GPU is a fantastic tool for specific operations 0:11:39.040 --> 0:11:42.560 and then less useful for others. Apart from processing graphics, 0:11:42.600 --> 0:11:45.400 GPUs have been found to be really useful in applications 0:11:45.480 --> 0:11:50.000 ranging from machine learning projects to proof of work cryptocurrency 0:11:50.320 --> 0:11:55.720 mining operations. Now to be clear, GPUs, at least until recently, 0:11:55.960 --> 0:12:00.000 occupied a kind of a sweet space in crypto minds. 0:12:00.679 --> 0:12:02.760 They are not the top of the heap when it 0:12:02.800 --> 0:12:07.120 comes to crypto mining integrated circuits. We'll get to the 0:12:07.200 --> 0:12:10.400 kind that are used in high end crypto mining in 0:12:10.520 --> 0:12:13.280 just a little bit. So for stuff like Bitcoin, which 0:12:13.320 --> 0:12:15.920 as I record this episode, is trading at around fifty 0:12:15.920 --> 0:12:18.480 eight thousand dollars per coin. In fact, I think it's 0:12:18.480 --> 0:12:23.000 like fifty eight point five thousand. That's a lot of money. Well, 0:12:23.160 --> 0:12:25.720 if you're using GPUs, you're not going to cut it. 0:12:25.920 --> 0:12:28.520 You're not going to compete in that space. GPUs just 0:12:28.600 --> 0:12:33.080 can't operate at a level that would make it feasible 0:12:33.120 --> 0:12:36.240 for you to use them for your mining operations. That's 0:12:36.280 --> 0:12:40.240 because the value of bitcoin is so high that it 0:12:40.400 --> 0:12:45.360 drives cryptocurrency miners to seek out the absolute top tier processors, 0:12:45.600 --> 0:12:50.160 and GPUs, while they're great, they're really more mid tier. 0:12:50.760 --> 0:12:52.880 Now it helps if you know what proof of work 0:12:53.000 --> 0:12:57.239 crypto mining is all about. So with proof of work systems, 0:12:57.520 --> 0:13:00.040 you have a network of machines that make up this 0:13:00.360 --> 0:13:04.120 cryptocurrency network, such as bitcoin. We'll use Bitcoin as the 0:13:04.200 --> 0:13:07.640 main example because that was sort of the progenitor of 0:13:07.679 --> 0:13:12.360 this space. So every so often the network issues a challenge, 0:13:12.360 --> 0:13:15.600 which is to solve a mathematical problem, and if you 0:13:15.679 --> 0:13:18.000 do solve it, if you're the first one to solve it, 0:13:18.080 --> 0:13:21.120 you will receive some crypto coins as a reward. The 0:13:21.160 --> 0:13:24.440 act of solving typically is tied to validating a block 0:13:24.520 --> 0:13:29.200 of crypto transactions, So the problem's complexity will depend upon 0:13:29.280 --> 0:13:33.400 a couple of different things. Typically, there's an ideal amount 0:13:33.440 --> 0:13:37.840 of time that it should take to solve this mathematical problem. 0:13:38.040 --> 0:13:41.720 For bitcoin, that time is ten minutes. The other thing 0:13:41.760 --> 0:13:44.440 that determines the complexity of the problem is how much 0:13:44.480 --> 0:13:48.000 computing power is being thrown at solving the problem in 0:13:48.040 --> 0:13:51.160 the first place. So let's go back to our classroom analogy. 0:13:51.280 --> 0:13:54.640 Let's say that you're creating a test, and for whatever reason, 0:13:54.720 --> 0:13:57.720 you have decided this test should take the students ten 0:13:57.800 --> 0:14:01.720 minutes to complete, so you're not really focused on any 0:14:01.760 --> 0:14:04.160 other outcome other than trying to make a test that's 0:14:04.200 --> 0:14:08.559 going to take ten minutes to complete. However, you've misjudged 0:14:08.559 --> 0:14:11.960 the difficulty. Maybe one of your students hands in their 0:14:12.040 --> 0:14:14.920 test six minutes in. Now you know you need to 0:14:14.960 --> 0:14:17.400 make the next test harder in order to hit this 0:14:17.440 --> 0:14:21.240 seemingly arbitrary goal of ten minutes. On the flip side, 0:14:21.400 --> 0:14:23.560 let's say the first student to solve the test took 0:14:23.640 --> 0:14:26.680 fifteen minutes to complete it. Then you know your test 0:14:26.720 --> 0:14:29.240 is too hard and you need to ease up a 0:14:29.240 --> 0:14:32.040 little bit for the next test. When the value of 0:14:32.080 --> 0:14:35.520 cryptocurrency goes up, there's a greater incentive to be the 0:14:35.600 --> 0:14:39.440 first to solve the mathematical problem because the reward is larger. 0:14:39.800 --> 0:14:43.840 That drives miners to buy more processors and to network 0:14:43.880 --> 0:14:47.320 them together, and these are processors that are particularly good 0:14:47.360 --> 0:14:50.480 at solving the types of problems that you get when 0:14:50.480 --> 0:14:55.000 you're crypto mining. For a while, that meant GPUs they 0:14:55.000 --> 0:14:58.760 were the best. But the value of bitcoin went up 0:14:58.840 --> 0:15:03.160 and up and up, and there were other options besides GPUs. 0:15:03.240 --> 0:15:06.400 There were options that were more expensive than GPUs, so 0:15:06.440 --> 0:15:09.480 it require a bigger investment, but then on top of that, 0:15:09.520 --> 0:15:11.800 you were looking at bigger rewards, so it made that 0:15:11.840 --> 0:15:17.120 investment worthwhile. So the integrated circuit that typically replaces GPUs 0:15:17.160 --> 0:15:21.080 for high end cryptocurrency mining, those would be application specific 0:15:21.160 --> 0:15:25.200 integrated circuits or AASAC ASIC. We'll get to those in 0:15:25.520 --> 0:15:28.920 just a bit, so you could if you wanted to 0:15:29.080 --> 0:15:32.920 still run mining rigs using GPUs, nothing would stop you 0:15:33.120 --> 0:15:35.560 from doing that, but you'd be going up against people 0:15:35.600 --> 0:15:39.960 with networks and machines running much more streamlined optimized processors, 0:15:40.240 --> 0:15:44.040 so you would be unlikely to beat them. Okay, we 0:15:44.160 --> 0:15:46.560 got a lot more to cover, but let's take a 0:15:46.600 --> 0:15:58.720 quick break to thank our sponsors. Okay, we're coming back 0:15:58.720 --> 0:16:00.920 to talk a little bit more about crypt currency mining 0:16:00.920 --> 0:16:06.120 in GPUs. So for a while, people who were crypto 0:16:06.240 --> 0:16:11.040 mining ethereum would stick with GPUs. The reason for this 0:16:11.120 --> 0:16:14.920 is ethereum had a lower value, much lower than bitcoin. 0:16:15.200 --> 0:16:17.280 All right, We're talking about a few thousand dollars as 0:16:17.280 --> 0:16:20.880 opposed to tens of thousands of dollars per coin, and 0:16:20.920 --> 0:16:26.240 this meant that it would be impractical to use high 0:16:26.400 --> 0:16:31.000 end integrade circuits like AASC circuits for mining ethereum because 0:16:31.040 --> 0:16:34.680 the cost of doing so would be such that you 0:16:34.680 --> 0:16:37.760 wouldn't be making up that cost in the profit you 0:16:38.280 --> 0:16:43.040 gained from mining the cryptocurrency. So sticking with GPUs made 0:16:43.080 --> 0:16:47.120 more sense, right because from an economic standpoint, that was 0:16:47.680 --> 0:16:52.560 the sweet spot. However, then Ethereum switched to proof of 0:16:52.760 --> 0:16:56.360 steak instead of proof of work. Proof of steak doesn't 0:16:56.360 --> 0:16:59.400 do that whole math problem thing at all, and the 0:16:59.440 --> 0:17:02.960 demand for GPUs and crypto mining plummeted as a result. 0:17:03.200 --> 0:17:05.679 There are other cryptocurrencies out there, some of which that 0:17:06.440 --> 0:17:09.119 still do use proof of work, but they're not as 0:17:09.200 --> 0:17:14.120 sought after as Bitcoin or ethereum are. So this meant 0:17:14.200 --> 0:17:17.439 that the demand for GPUs in the crypto space began 0:17:17.560 --> 0:17:20.640 to diminish, and that became really good news for people 0:17:20.680 --> 0:17:23.679 who wanted a GPU for something else, like for a 0:17:23.680 --> 0:17:27.280 gaming rig for example. Now, I would say for the 0:17:27.359 --> 0:17:31.640 majority of people out there, like your average consumers, CPUs 0:17:31.720 --> 0:17:34.680 and GPUs are the beginning and end of it when 0:17:34.720 --> 0:17:38.159 it comes to processors or microchips that are meant to 0:17:38.280 --> 0:17:41.280 act like processors, But there are a couple of other 0:17:41.359 --> 0:17:45.080 varieties out there that we use for special purposes. And 0:17:45.119 --> 0:17:48.880 the special purpose thing is the important part to keep 0:17:48.920 --> 0:17:52.400 in mind. A CPU, by necessity, has to be able 0:17:52.400 --> 0:17:55.200 to do a bit of everything right. Because a CPU 0:17:55.600 --> 0:18:00.240 is the control center of your typical computer. It needs 0:18:00.280 --> 0:18:02.600 to be able to handle operations from a variety of 0:18:02.600 --> 0:18:05.320 different programs and that kind of thing. It is a 0:18:05.400 --> 0:18:08.080 jack of all trades master of none. It needs to 0:18:08.119 --> 0:18:10.239 be able to handle whatever you throw at it, but 0:18:10.280 --> 0:18:14.840 that means it cannot be optimized for any specific task. 0:18:15.320 --> 0:18:19.760 So what it lacks in efficiency, it makes up for inversatility. 0:18:20.240 --> 0:18:24.240 GPUs are more specialized and so they can handle certain 0:18:24.280 --> 0:18:29.240 processes better than a CPU typically can, But a GPU 0:18:29.400 --> 0:18:32.199 might not be so good at executing all the different 0:18:32.240 --> 0:18:35.080 tasks that a CPU has to handle, So while it 0:18:35.240 --> 0:18:39.520 is faster with some stuff, it's slower with other stuff. Now, 0:18:39.560 --> 0:18:42.240 the next two types of semiconductor devices I want to 0:18:42.280 --> 0:18:46.280 mention are even more specialized than GPUs, and then we'll 0:18:46.400 --> 0:18:52.159 end with one that is specialized specifically for the AI field. 0:18:52.720 --> 0:18:57.560 So next up is the field programmable gait arrays or 0:18:57.720 --> 0:19:02.879 FPGA's now a definition from XLinks dot com because x 0:19:02.920 --> 0:19:07.920 links is what introduced this technology back in the mid 0:19:08.080 --> 0:19:13.479 nineteen eighties. So x links defines this as FPGA's quote. 0:19:13.640 --> 0:19:18.480 Are based around a matrix of configurable logic blocks CLBs 0:19:18.680 --> 0:19:23.840 connected via programmable interconnects. End quote that sounds like gibberish 0:19:23.920 --> 0:19:29.200 to some folks. It's definitely got some barriers there from 0:19:29.240 --> 0:19:33.040 easy understanding, but the idea is pretty simple. When you 0:19:33.040 --> 0:19:37.000 boil it down to what's basically happening. So imagine that 0:19:37.080 --> 0:19:40.479 you have a microchip and you're able to reconfigure the 0:19:40.640 --> 0:19:44.360 individual components on that microchip so that they're optimized for 0:19:44.440 --> 0:19:46.960 whatever it is you need to do. So you can 0:19:47.080 --> 0:19:50.879 reprogram this chip, in other words, so that it is 0:19:50.960 --> 0:19:54.880 better aligned with the task you have at hand. As 0:19:54.920 --> 0:19:57.640 I said, x links first introduced this type of integrated 0:19:57.640 --> 0:20:00.199 circuit back in nineteen eighty five, and the aim us 0:20:00.240 --> 0:20:02.800 to make an integrated circuit that could potentially fit the 0:20:02.840 --> 0:20:07.240 needs of different specific use cases, not by being a 0:20:07.320 --> 0:20:10.639 jack of all trades that could do anything, but do 0:20:10.800 --> 0:20:13.440 so at a kind of a mediocre level, but rather 0:20:13.600 --> 0:20:18.840 by being configured to work best for that specific application. Moreover, 0:20:19.560 --> 0:20:23.800 you can at least sometimes reconfigure without having to change 0:20:23.840 --> 0:20:27.439 the actual physical architecture of the chip itself. This is 0:20:27.480 --> 0:20:30.640 important because not everyone has access to a clean room 0:20:30.840 --> 0:20:35.439 with incredibly precise and computer operated tools. That's exactly what 0:20:35.480 --> 0:20:37.800 you would need if you wanted to perform surgery on 0:20:37.880 --> 0:20:43.679 a microchip. Instead, an FPGA has these CLBs that x 0:20:43.720 --> 0:20:47.439 links talked about, the configurable logic blocks. These can be 0:20:47.480 --> 0:20:50.679 programmed to act like simple logic gates, and these gates 0:20:50.680 --> 0:20:54.800 follow specific rules. Essentially, they either allow electrical current to 0:20:54.800 --> 0:20:58.640 flow through or they block it from flowing through, and this, 0:20:59.040 --> 0:21:01.600 when you look at it a macro level, is what 0:21:01.800 --> 0:21:06.720 allows operations on a processor. The field and field programmable 0:21:06.840 --> 0:21:10.560 gate array means you can actually do this reprogramming after 0:21:10.640 --> 0:21:14.840 the FPGA has shipped from its manufacturer. So instead of 0:21:14.960 --> 0:21:18.439 working with a manufacturer to specialize a chip from the 0:21:18.520 --> 0:21:21.760 design phase and then go all the way through to manufacturing, 0:21:22.119 --> 0:21:26.600 the manufacturer makes this FPGA that can potentially be one 0:21:26.720 --> 0:21:30.159 of thousands of different configurations, and then you program it 0:21:30.200 --> 0:21:33.639 once you receive it. Now, some of these FPGAs are 0:21:33.680 --> 0:21:37.399 limited to kind of a one time only configuration, so 0:21:37.600 --> 0:21:40.720 you can program them once you get them, but then 0:21:40.760 --> 0:21:45.760 they're set in that particular configuration from that point forward. 0:21:45.880 --> 0:21:49.200 But others are designed so that they can be reprogrammed 0:21:49.440 --> 0:21:52.359 multiple times, which obviously makes them very useful. If you 0:21:52.440 --> 0:21:57.399 wanted to prototype a technology and you aren't really sure 0:21:57.520 --> 0:22:00.320 which configuration is going to be best for whatever it 0:22:00.359 --> 0:22:02.359 is you're trying to do, it's great to have a 0:22:02.480 --> 0:22:07.399 chip you can reprogram so you can try different configurations 0:22:07.440 --> 0:22:09.760 to find the one that makes the most sense for 0:22:09.880 --> 0:22:13.159 whatever it is you're trying to achieve. One issue with 0:22:13.400 --> 0:22:16.840 FPGA's is that they are not cost efficient when you're 0:22:16.880 --> 0:22:21.000 looking at mass production. They're great if you are prototyping, 0:22:21.440 --> 0:22:23.760 but if you plan to make a whole bunch of them, 0:22:23.840 --> 0:22:27.439 it gets time consuming and expensive because not only do 0:22:27.480 --> 0:22:28.919 you have to have them made, then you have to 0:22:28.920 --> 0:22:32.440 have them programmed. Plus sometimes you may have an application 0:22:32.520 --> 0:22:36.200 in mind that an FPGA cannot accommodate even with all 0:22:36.200 --> 0:22:39.639 the reconfiguring. So think of an FPGA as having a 0:22:39.680 --> 0:22:42.760 limited number of configurations and it turns out that what 0:22:42.840 --> 0:22:46.040 you need is outside of this range. That would mean 0:22:46.080 --> 0:22:49.240 you would need to add additional integrated circuits to your 0:22:49.320 --> 0:22:54.359 system to accommodate these limitations of the FPGA itself, which 0:22:54.359 --> 0:22:57.359 means you're adding more complexity to your system, and that 0:22:57.400 --> 0:23:01.440 in turn also means you're adding more costs to your system. 0:23:01.680 --> 0:23:04.919 Next up, we have the one I mentioned earlier, the 0:23:05.040 --> 0:23:11.600 Application specific integrated circuit or AZC ASIC, as the name indicates, 0:23:11.640 --> 0:23:16.320 These chips are made to operate for specific applications, and 0:23:16.480 --> 0:23:21.639 as such, they are highly optimized from the hardware level 0:23:21.840 --> 0:23:25.080 up for that purpose. They are not meant to be 0:23:25.200 --> 0:23:28.720 general purpose processors like a CPU. So if you put 0:23:28.760 --> 0:23:31.320 an ASK to work on a task that it was 0:23:31.480 --> 0:23:34.240 not designed to handle, you are not going to get 0:23:34.240 --> 0:23:36.480 a good result. In fact, it may not work at all. 0:23:36.800 --> 0:23:39.760 But when it's integrated into a system that's meant to 0:23:39.840 --> 0:23:43.439 do that one thing it was designed to do, it 0:23:43.520 --> 0:23:46.840 does it really well. And AAK can be a speed 0:23:47.000 --> 0:23:51.320 demon and operate at an efficiency that's much more desirable 0:23:51.320 --> 0:23:55.120 than your typical CPU or even GPU. So unlike an 0:23:55.200 --> 0:24:00.920 FPGA and ASK cannot be reconfigured. It is a high tech, 0:24:01.560 --> 0:24:05.280 highly specialized chip. There are a few different approaches to 0:24:05.600 --> 0:24:09.320 create that specialization during the manufacturing process, but I feel 0:24:09.359 --> 0:24:11.840 like that's beyond the scope of this episode. I'll save 0:24:11.880 --> 0:24:14.360 it for a time when I do a full episode 0:24:14.480 --> 0:24:19.600 about AZK chips. Now, the design process for AZIK is complicated. 0:24:19.680 --> 0:24:22.880 So imagine you're building a chip intended to do one 0:24:22.960 --> 0:24:25.720 thing extremely well. You would have to do a lot 0:24:25.720 --> 0:24:27.680 of work to make sure that the chip you were 0:24:27.720 --> 0:24:31.400 designing met that purpose. So that means there's a lot 0:24:31.400 --> 0:24:34.520 of R and D and there's a lot of testing. However, 0:24:34.960 --> 0:24:38.320 once you do arrive at this final design, one big 0:24:38.359 --> 0:24:42.280 advantage of AZK over FPGA is that it can then 0:24:42.400 --> 0:24:47.000 go into large volume production. So while the development process 0:24:47.119 --> 0:24:50.520 of an AZK is typically longer and more expensive than 0:24:51.040 --> 0:24:55.360 using an FPGA, once you do get to the production stage, 0:24:55.640 --> 0:24:58.880 the AZIC chips become more cost effective. So if you're 0:24:58.880 --> 0:25:02.400 doing a one off, FPGA makes the most sense financially. 0:25:02.800 --> 0:25:05.600 If your goal is to make something that you're going 0:25:05.680 --> 0:25:09.560 to mass produce, AZIC makes far more sense. ASIC chips 0:25:09.560 --> 0:25:12.600 also tend to be more power efficient than FPGA's, so 0:25:12.720 --> 0:25:16.359 by their nature, an FPGA needs to have components that 0:25:16.400 --> 0:25:21.439 aren't necessary for all applications because the whole point of 0:25:21.440 --> 0:25:25.040