WEBVTT - Smart Talks with IBM: Accelerating an answer to COVID-19 0:00:00.160 --> 0:00:03.000 In a bit we'll hear my conversation with Dave Turik, 0:00:03.320 --> 0:00:07.640 who oversees ibm S High Performance Computing Division, to learn 0:00:07.880 --> 0:00:10.840 more about all of that. But before we get to that, 0:00:11.200 --> 0:00:12.959 I thought it would be useful to give a quick 0:00:12.960 --> 0:00:19.080 definition and overview of supercomputers. So what is a supercomputer? 0:00:19.920 --> 0:00:22.640 Generally speaking, it's a computer that can perform at a 0:00:22.720 --> 0:00:27.560 level far beyond the average computer. You know, leap tall 0:00:27.640 --> 0:00:31.160 processes at a single bound. It can be a bit 0:00:31.200 --> 0:00:34.560 of a sliding classification. It's something we apply to an 0:00:34.560 --> 0:00:37.720 elite group of computers that operate at a level above 0:00:37.760 --> 0:00:40.680 and beyond what other machines are capable of at that time. 0:00:40.960 --> 0:00:42.559 And I thought it might be a good idea to 0:00:42.560 --> 0:00:45.400 explain what those are, since otherwise the only impression you'll 0:00:45.400 --> 0:00:50.000 get is that more flops equals more good somehow. So 0:00:50.120 --> 0:00:54.600 let's start with floating point numbers and computing, you might 0:00:54.680 --> 0:00:57.960 deal with integers, and these are whole numbers with no fractions. 0:00:58.040 --> 0:01:01.360 Like the number three. Three is a good number, it's 0:01:01.400 --> 0:01:05.880 an integer. But what about point three? Now we have 0:01:05.880 --> 0:01:08.240 a number that has a decimal in it. This is 0:01:08.280 --> 0:01:12.560 not an integer, but it can be a floating point number. Now, 0:01:12.640 --> 0:01:15.920 computers are really good at working with integers. They can 0:01:15.959 --> 0:01:20.480 calculate processes on integers whippity quick, but floating point numbers 0:01:20.880 --> 0:01:23.800 those can take a bit longer, and speed is a 0:01:23.800 --> 0:01:27.520 big deal in computing. You always want answers quickly. But 0:01:27.560 --> 0:01:30.200 the reason we call them floating point numbers is that 0:01:30.280 --> 0:01:34.200 you can move that decimal around so that point three, well, 0:01:34.240 --> 0:01:37.319 we could represent that as three times ten to the 0:01:37.319 --> 0:01:41.480 power of minus one. This is an example of scientific notation, 0:01:41.720 --> 0:01:44.800 something used in lots of disciplines to help represent very 0:01:44.920 --> 0:01:48.000 large or very small numbers without having to write in 0:01:48.040 --> 0:01:50.560 all those darned zeros. For example, if I wanted to 0:01:50.560 --> 0:01:53.400 write out the number two trillion, I would write the 0:01:53.480 --> 0:01:58.040 numeral two followed by twelve zeros. That's a lot of zeros, 0:01:58.120 --> 0:02:00.200 and honestly, if I wanted to do anything use full 0:02:00.240 --> 0:02:02.520 with that number, it would end up being a real hassle. 0:02:02.880 --> 0:02:05.840 But I could represent the same number as two times 0:02:05.880 --> 0:02:08.720 ten to the twelve power. I wanted to give you 0:02:08.760 --> 0:02:12.639 guys a basic understanding of floating point operations because that's 0:02:12.680 --> 0:02:16.480 going to come into play in my discussion in this episode. 0:02:16.560 --> 0:02:18.880 So now that we've got that out of the way, 0:02:19.120 --> 0:02:22.519 we can move on. Dave Turik, vice president of High 0:02:22.560 --> 0:02:26.440 Performance Computing and Cognitive Systems at IBM, spoke with me 0:02:26.560 --> 0:02:31.040 on Thursday April two, twenty twenty about high performance computing 0:02:31.080 --> 0:02:33.960 in general and how researchers are using it in an 0:02:34.000 --> 0:02:38.359 effort to research the coronavirus and COVID nineteen. And I 0:02:38.400 --> 0:02:42.480 should also add we recorded this call over the Internet, 0:02:42.919 --> 0:02:45.600 and so the quality is not the same as what 0:02:45.639 --> 0:02:48.160 we would usually have in a studio. You're going to 0:02:48.240 --> 0:02:53.800 hear some effects because of the Internet connection. You'll probably 0:02:53.800 --> 0:02:56.840 hear some extraneous noise, and I apologize for that, but 0:02:57.080 --> 0:03:00.880 in these extraordinary circumstances, this was the best we could 0:03:01.000 --> 0:03:03.640 manage in order to have this important conversation. And I 0:03:03.680 --> 0:03:07.240 want to thank Dave for his time and patience in 0:03:07.639 --> 0:03:10.239 setting this up, and I really appreciate it. So let's 0:03:10.280 --> 0:03:15.160 jump into it. Dave, before we go into this incredible 0:03:15.200 --> 0:03:19.520 effort that we're seeing from research institutions using supercomputers to 0:03:19.919 --> 0:03:22.920 research the coronavirus and look at treatments for COVID nineteen, 0:03:23.360 --> 0:03:26.400 can you define in broad terms what is actually meant 0:03:26.680 --> 0:03:30.480 by high performance computing? Well, I think, uh, the way 0:03:30.520 --> 0:03:34.440 to think about high performance computing is in terms of 0:03:34.480 --> 0:03:37.760 the nature of the problem first of all, and then 0:03:37.880 --> 0:03:41.480 the kind of computing the supplied against it. So by 0:03:41.560 --> 0:03:44.800 nature of the problem, I mean that it's fundamentally infused 0:03:44.840 --> 0:03:51.200 with mathematical representations of systems or problem types. And then 0:03:51.200 --> 0:03:54.920 from a computing perspective, the kind of technology that puts 0:03:54.920 --> 0:03:59.640 an emphasis on floating point and very quick communications as 0:03:59.680 --> 0:04:03.200 a coal by which those problems are tackled. That just 0:04:03.360 --> 0:04:06.960 helps one distinguish between somebody saying, well, I can solve 0:04:07.000 --> 0:04:09.960 this problem on a phone, right, that's not what we're 0:04:09.960 --> 0:04:12.880 talking about here. The nature of the mathematics are complex 0:04:13.480 --> 0:04:16.160 and sometimes quite extreme, and the computing we required to 0:04:16.240 --> 0:04:21.280 tackle those have similar kinds of capabilities to overcome that complexity. 0:04:21.680 --> 0:04:24.480 So now that we've kind of got to grasp on that, 0:04:24.520 --> 0:04:28.680 we're looking at a sort of a a massive scale 0:04:28.800 --> 0:04:35.080 form of computing that does very complicated processes very very quickly. 0:04:35.520 --> 0:04:39.800 Can you talk a bit about the Higher Performance Computing Consortium? 0:04:39.839 --> 0:04:42.760 What is what is that organization? How did that come about? 0:04:44.240 --> 0:04:49.839 The COVID nineteen HPC Consortium came about roughly ten days 0:04:49.839 --> 0:04:55.159 ago um courtesy of the conversation between our director of 0:04:55.200 --> 0:04:58.520 research Dario Gill and people at the White House and 0:04:58.560 --> 0:05:02.520 subsequently the Department Energy to see how we could apply 0:05:02.880 --> 0:05:07.720 high performance computing or super computing, two problems associated with 0:05:08.240 --> 0:05:13.360 uh COVID nineteen and quite quickly the offers were taken 0:05:13.440 --> 0:05:15.840 up and within a matter of a couple of days 0:05:15.880 --> 0:05:19.080 we had a website up and running. They gave the 0:05:19.080 --> 0:05:22.840 broad parameters of the resources that were available on how 0:05:22.880 --> 0:05:25.760 one could make submissions to it, and then with a 0:05:25.839 --> 0:05:28.719 passage of another couple of days, we brought in a 0:05:28.800 --> 0:05:32.760 number of additional partners as well UM to complement the 0:05:32.839 --> 0:05:36.560 capability that we initially we're able to access of via 0:05:36.600 --> 0:05:40.000 IBM and a Department of Energy excellent and one are 0:05:40.040 --> 0:05:44.040 some of the actual technologies that are being used in 0:05:44.080 --> 0:05:47.320 this process. We've mentioned supercomputers, can you talk about any 0:05:47.360 --> 0:05:53.280 specific ones and UH, what about things like artificial intelligence, 0:05:53.320 --> 0:05:57.040 machine learning or what kind of various tech are coming 0:05:57.080 --> 0:06:01.480 together to tackle this this UH this issue. The glib 0:06:01.520 --> 0:06:03.600 answer of course is everything. But let me be a 0:06:03.640 --> 0:06:08.000 little more specific from the perspective of the supercomputers that 0:06:08.040 --> 0:06:12.839 are part of the consortium currently UH. They range from 0:06:12.880 --> 0:06:17.000 the Power nine based supercomputers that one finds at oak 0:06:17.120 --> 0:06:21.080 Ridge and Lawrence Livermore to x AD six based systems 0:06:21.120 --> 0:06:25.680 that you might find a NASA our gone in other places. UM. 0:06:25.720 --> 0:06:28.120 For the most part, most of these systems, but not 0:06:28.160 --> 0:06:33.800 all of them use accelerators UM and UH, and that 0:06:33.880 --> 0:06:37.279 really deals with some of the floating point computations that 0:06:37.320 --> 0:06:40.320 are involved, and in some cases the systems are are 0:06:40.360 --> 0:06:46.400 absent UM UH accelerators, So those are homogeneous systems. So 0:06:46.480 --> 0:06:50.480 that's the hardware characterization when we begin to talk about 0:06:50.560 --> 0:06:53.960 machine learning deep learning in those things, that's a combination 0:06:54.120 --> 0:06:58.960 of software running in sync with particular hardware attributes. So 0:06:59.680 --> 0:07:02.880 from a deep learning from a model training perspective, there's 0:07:02.880 --> 0:07:06.880 a premium placed on the availability of accelerators. So the 0:07:06.920 --> 0:07:10.000 Summit system at oak Ridge, for examples, and fused with 0:07:10.040 --> 0:07:15.200 about accelerators, so it's terrific for helping people train models. 0:07:15.680 --> 0:07:18.160 But then as you begin to do influencing in some 0:07:18.320 --> 0:07:23.600 of the other machine learning techniques, the emphasis UM exclusively 0:07:23.600 --> 0:07:27.240 on accelerators evolves a little bit and you get to 0:07:27.480 --> 0:07:31.680 employ different kinds of architectural approaches to UH to look 0:07:31.720 --> 0:07:35.760 at actually inferencing problems. So it's a combination of software 0:07:35.760 --> 0:07:39.520 and hardware that's meant to be reasonably flexible. Not One 0:07:39.560 --> 0:07:42.040 of the things I'll say, of course, and oak Ridge 0:07:42.360 --> 0:07:44.960 along with IBM, have been a pioneer in this is 0:07:45.000 --> 0:07:49.600 that there's not a sharp dichotomy between AI R at large, 0:07:49.680 --> 0:07:53.040 which includes machine learning, natural language processing, deep learning, and 0:07:53.080 --> 0:07:57.240 so on, and HPC. In fact, that two domains have 0:07:57.360 --> 0:08:00.520 really come together in the last couple of years where 0:08:00.760 --> 0:08:04.640 problems now get decomposed in ways where maybe certain parts 0:08:04.640 --> 0:08:09.120 of the problems are tackled with classic HPC methodologies and 0:08:09.160 --> 0:08:12.080 other parts of the problems are not tackled with more 0:08:12.120 --> 0:08:16.560 current AI approaches. So it's this amalgamation at capabilities that 0:08:16.600 --> 0:08:20.160 are brought together under software control that creates the impact. 0:08:20.880 --> 0:08:23.360 Dave Turik mentioned a few things I feel I should 0:08:23.440 --> 0:08:26.200 unpack here, and let's start with talking about one of 0:08:26.240 --> 0:08:30.080 the supercomputers he alluded to, the Summit Supercomputer at oak 0:08:30.160 --> 0:08:33.520 Ridge National Laboratory. Now, this is just one of the 0:08:33.559 --> 0:08:36.600 supercomputers that are part of this consortium, and it is 0:08:36.640 --> 0:08:40.320 currently the reigning champ of supercomputers, and researchers are using 0:08:40.320 --> 0:08:43.640 it to do everything from understanding how molecular interactions and 0:08:43.679 --> 0:08:48.000 human cells could lead to much more complex traits uh 0:08:48.040 --> 0:08:51.000 to exploring the physics of propulsion systems and an effort 0:08:51.040 --> 0:08:54.400 to make better, more efficient ones in the future. If 0:08:54.440 --> 0:08:58.280 computers were people, Summit would be that amazing overachiever you 0:08:58.320 --> 0:09:01.960 know who tackles any type of olunge with enthusiasm. Someone 0:09:02.040 --> 0:09:05.600 alone can achieve a peak performance of two hundred pedaphlops. 0:09:05.600 --> 0:09:10.880 That's two hundred thousand trillion calculations of floating point operations 0:09:10.880 --> 0:09:14.960 per second. Dave also mentioned inference problems, and that gets 0:09:15.000 --> 0:09:18.640 down to looking at data and inferring probabilities based on 0:09:18.679 --> 0:09:22.000 the data you've gathered, and building probabilistic tables is an 0:09:22.040 --> 0:09:25.680 important part of science and when done properly, can really 0:09:25.760 --> 0:09:28.480 speed things up. You look at which options appear to 0:09:28.559 --> 0:09:31.680 be the most promising, and you focus on those, and 0:09:32.000 --> 0:09:34.320 you might discard all the ones that have a very 0:09:34.400 --> 0:09:37.480 low probability of being helpful, or at least put them 0:09:37.559 --> 0:09:40.760 to the side. If you exhaust all the most promising 0:09:40.760 --> 0:09:43.800 options without a result, then you can revisit some of 0:09:43.840 --> 0:09:45.960 the other ones. But really it's a great way to 0:09:46.000 --> 0:09:49.960 eliminate options, giving you the ability to focus on the 0:09:50.000 --> 0:09:53.640 best chance for success. Let's get back to the interview. 0:09:54.320 --> 0:09:58.680 So with COVID nineteen in particular, what are the ways 0:09:58.720 --> 0:10:02.160 some of the ways that reas searchers are leveraging these 0:10:02.200 --> 0:10:07.480 technologies to specifically look at that crisis. So I think 0:10:07.640 --> 0:10:11.400 the first way to think of it is to just 0:10:11.480 --> 0:10:15.960 take a second and inform your listeners about the modern 0:10:16.000 --> 0:10:20.520 ways in which chemistry, biology and biochemistry are done, because 0:10:20.559 --> 0:10:23.240 I think many lay people have this image from their 0:10:23.280 --> 0:10:26.760 high school or college days of speakers and pipettes and 0:10:26.800 --> 0:10:29.680 things like that, sort of the what I would characterize 0:10:29.720 --> 0:10:32.880 the representation of science in the analog world, what you 0:10:33.000 --> 0:10:36.040 touch and feel and deal with every day. But what's 0:10:36.080 --> 0:10:39.679 transpired over the last several decades is this movement to 0:10:40.280 --> 0:10:44.680 progressively infuse science and the scientific method with more and 0:10:44.720 --> 0:10:48.800 more computational capability. Now, what that comes down to in 0:10:48.800 --> 0:10:52.440 the case of COVID nineteen is one begins to take 0:10:52.520 --> 0:10:57.800 first principles kinds of theories of the way adams are structure, 0:10:58.000 --> 0:11:00.720 molecules of structure, in the way adams be a and 0:11:00.760 --> 0:11:05.560 how they interact with one another, represent that mathematical form, 0:11:05.640 --> 0:11:08.959 and use the computers to explore the behavior from a 0:11:09.040 --> 0:11:13.959 first principle's perspective. Before you ever get to a physical laboratory. 0:11:14.160 --> 0:11:17.120 So what that nets out to is you can now 0:11:17.240 --> 0:11:20.760 use the power of computing to assess thousands and hundreds 0:11:20.760 --> 0:11:25.160 of thousands of molecules in terms of their potential impact 0:11:25.200 --> 0:11:28.400 on the virus and explore the behavior and the and 0:11:28.480 --> 0:11:34.320 the constraints and the amplifications of combinations of molecules digitally 0:11:34.840 --> 0:11:37.520 before you ever have to go to the laboratory to 0:11:37.600 --> 0:11:40.520 try to recreate the results you've seen digitally in the 0:11:40.520 --> 0:11:44.720 analog world. And that's been a tremendous speed up in 0:11:44.840 --> 0:11:48.280 terms of time. You know, pharmaceutical companies today they may 0:11:48.320 --> 0:11:51.959 have at their disposal billions of molecules that they might 0:11:52.000 --> 0:11:55.920 want to look at for particular pharmaceutical impact, and sorting 0:11:55.960 --> 0:11:59.560 through that is just gigantic task. And the ability to 0:11:59.600 --> 0:12:02.440 have co uters to come in and say, look, I 0:12:02.480 --> 0:12:05.760 know you're looking at eight thousand molecules here, which is 0:12:05.800 --> 0:12:08.760 what researchers at oak Ridge did, but I can cut 0:12:08.800 --> 0:12:12.319 that down to seventy seven just by using digital approaches 0:12:12.760 --> 0:12:16.160 and simulation and computation, so that you don't have to 0:12:16.200 --> 0:12:18.439 worry about trying to analyze all eight thousands and the 0:12:18.520 --> 0:12:22.400 laboratory you can focus on seventy seven so that's the 0:12:22.480 --> 0:12:26.680 first big step of what's happening here. No, that's amazing 0:12:26.720 --> 0:12:29.400 because just the idea of cutting out that step in 0:12:29.440 --> 0:12:34.679 the wet lab where you're having to physically uh analyze 0:12:34.880 --> 0:12:38.360 reactions or maybe not even analyze, you're just detecting to 0:12:38.360 --> 0:12:41.880 see if one is happening. Cutting that downs that you 0:12:41.880 --> 0:12:48.319 can really focus on the best uh potential solutions is phenomenal. 0:12:48.400 --> 0:12:49.960 Can we talk a little bit about what is it 0:12:50.080 --> 0:12:54.840 about these simulations that make them so challenging that high 0:12:54.840 --> 0:12:59.920 performance computing is suitable for tackling that kind of thing? Well, 0:13:00.080 --> 0:13:03.760 I think that um. One of the principal methodologies that 0:13:03.840 --> 0:13:09.200 people used in these investigations is molecular dynamics, and what 0:13:09.360 --> 0:13:12.800 that entails is, first of all, the characterization of a 0:13:12.880 --> 0:13:18.480 molecule in atoms, and and then the application of forces 0:13:18.520 --> 0:13:21.760 at the atomic level in terms of how they interact 0:13:21.800 --> 0:13:25.720 with one another. And so those forces are complicated, the 0:13:25.840 --> 0:13:29.560 time steps are extraordinarily small, and yet you want to 0:13:29.600 --> 0:13:33.240 observe how these things interact. Not only yet, let's say 0:13:33.320 --> 0:13:36.440 tend to the minus fifteen seconds, which actually like to 0:13:36.480 --> 0:13:40.480 see how they behave in in in real seconds, in 0:13:40.600 --> 0:13:44.720 minutes and hours and days, and those time scales just 0:13:44.840 --> 0:13:49.080 create a tremendous number of computational steps that one has 0:13:49.160 --> 0:13:52.240 to pursue in the concept of looking at these atomic 0:13:52.320 --> 0:13:55.800 forces that are operating on the target molecules and atoms 0:13:55.800 --> 0:13:58.679 and how they interact with one another. So the mathematics 0:13:58.679 --> 0:14:03.600 is stunningly complex. The time frames are just so extreme 0:14:03.720 --> 0:14:06.280 that it requires tremendous amount of compute power just to 0:14:06.360 --> 0:14:10.040 simulate a handful of seconds. And by virtue of having 0:14:10.120 --> 0:14:13.480 gigantic supercomputers operate on this, we can actually do this 0:14:14.200 --> 0:14:16.600 in a reasonable way and a reasonable amount of wall 0:14:16.640 --> 0:14:20.680 clock time. So let's consider what researchers are doing in 0:14:20.720 --> 0:14:23.920 this case. A virus consists of at least two parts. 0:14:23.960 --> 0:14:27.120 You've got a nucleic acid genome, which contains the material 0:14:27.160 --> 0:14:29.440 the virus needs to make copies of itself once it 0:14:29.560 --> 0:14:32.800 is a proper host cell. And then you've got a 0:14:32.880 --> 0:14:37.400 protein capsid or shell that contains the nucleic acid until 0:14:37.400 --> 0:14:40.320 the virus can attach itself and inject that material into 0:14:40.440 --> 0:14:44.920 the aforementioned host cell. Together, this is called the nucleocapsid. 0:14:45.320 --> 0:14:48.640 Many animal viruses also have a lipid envelope, and that 0:14:48.800 --> 0:14:51.000 is a membrane that has lots of stuff in it, 0:14:51.040 --> 0:14:54.360 including viral y programmed proteins in it. One of the 0:14:54.360 --> 0:14:58.359 purposes of those proteins is to bind with compatible receptors 0:14:58.360 --> 0:15:01.040 on host cells. So can kind of think of it 0:15:01.160 --> 0:15:04.080 as a virus has a special kind of plug and 0:15:04.120 --> 0:15:07.040 it's looking for cells that have a compatible outlet, and 0:15:07.080 --> 0:15:09.840 when it finds such a cell, it can plug in 0:15:10.080 --> 0:15:12.960 connect to that cell, injecting the nucleic acid of the 0:15:13.040 --> 0:15:16.000 virus into the host cell, and then the code in 0:15:16.040 --> 0:15:18.920 that nucleic acid hijacks the host cell turns it into 0:15:18.960 --> 0:15:23.760 a virus replication engine. Scientists need to know how specific 0:15:23.800 --> 0:15:28.200 molecules will interact with each other, the virus, host cells, 0:15:28.200 --> 0:15:31.800 and more. These interactions happen at such a small scale, 0:15:31.920 --> 0:15:36.880 and it's such minute slices or steps of time that 0:15:37.000 --> 0:15:39.160 it is difficult to describe. And this is where the 0:15:39.160 --> 0:15:42.400 speed of high performance computing really comes into play. First, 0:15:42.960 --> 0:15:46.840 breaking down elements of time gets mind boggling. We tend 0:15:46.880 --> 0:15:49.400 to think of it in terms of, as Dave says, 0:15:49.680 --> 0:15:53.680 wall clock time, you know, seconds, minutes, and hours. We 0:15:53.680 --> 0:15:56.520 can get our minds wrapped around shorter slices of time 0:15:56.800 --> 0:16:00.320 because counting on the second might sound like one mrs cippy, 0:16:00.800 --> 0:16:04.200 so we can definitely think of just one right, that's shorter. 0:16:04.640 --> 0:16:06.400 But eventually we hit a point where it's hard for 0:16:06.440 --> 0:16:11.760 us to really understand time at very tiny slices. We 0:16:11.800 --> 0:16:15.640 can always find a way to slice time into smaller increments. 0:16:15.680 --> 0:16:20.160 We can continue to make smaller and smaller slices of time. 0:16:20.560 --> 0:16:23.640 For example, there's a femto second. A femto second is 0:16:23.680 --> 0:16:27.760 just one quadrillionth of a second. That's tend to the 0:16:27.800 --> 0:16:33.320 power of minus fifteen. So imagine simulating the interactions between 0:16:33.360 --> 0:16:37.720 molecules in a series of these unimaginably short slices of 0:16:37.760 --> 0:16:41.360 time up to the point that collectively they amount to 0:16:41.760 --> 0:16:44.720 enough that it would reach our perceptible world. So we're 0:16:44.720 --> 0:16:48.120 talking about a quadrillion slices of time to make up 0:16:48.160 --> 0:16:51.680 just one second here, and there could be numerous important 0:16:51.720 --> 0:16:55.200 interactions on the molecular level within that short time frame. 0:16:55.320 --> 0:16:58.000 And this is why supercomputers are necessary for this sort 0:16:58.000 --> 0:17:00.560 of work. It allows for a precision and that we 0:17:00.600 --> 0:17:03.960 otherwise would find impossible. And again it tells us if 0:17:03.960 --> 0:17:07.280 a potential molecule shows promise in our efforts, or if 0:17:07.320 --> 0:17:10.080 it's likely to be a bust. Back to my conversation 0:17:10.080 --> 0:17:12.879 with Dave Turik, vice president of High Performance Computing and 0:17:12.920 --> 0:17:16.879 Cognitive Systems at IBM. With supercomputers being able to tackle 0:17:16.920 --> 0:17:20.679 this kind of thing through their various methodologies, this I 0:17:20.680 --> 0:17:22.399 would imagine would be something that if we were to 0:17:22.520 --> 0:17:26.000 use a classic computer, it could take thousands of years. 0:17:26.040 --> 0:17:31.960 Is that accurate? Yes? Um, And and in some sense 0:17:32.000 --> 0:17:37.320 it wouldn't even be possible because modern supercomputers, which I'll 0:17:37.320 --> 0:17:44.280 declare is roughly the error from UM, really are systems 0:17:44.320 --> 0:17:48.200 that are built on this concept of parallel processing parallel computing, 0:17:48.720 --> 0:17:52.880 which in turn revolves around this idea that you can 0:17:52.920 --> 0:17:57.000 decompose a problem into its component elements, and if you 0:17:57.080 --> 0:18:01.159 have enough elemental computing entities in your supercomputer, you can 0:18:01.200 --> 0:18:04.639 assign each one of those little problem parts to a 0:18:04.760 --> 0:18:08.919 different computing yell element and orchestrate the execution of the 0:18:08.920 --> 0:18:13.000 computation against that and and just radically reduce the amount 0:18:13.000 --> 0:18:16.160 of time required from for computation. So let me put 0:18:16.200 --> 0:18:20.680 it this way, UM, on a standard laptop computer, for example, 0:18:20.680 --> 0:18:27.720 you're gonna be running, principally to some order of magnitude UM, 0:18:27.760 --> 0:18:30.800 a serial kind of process. You know, you're gonna execute 0:18:30.800 --> 0:18:34.000 and solve problem A, which is followed by B, C, 0:18:34.320 --> 0:18:37.520 D and so on. In the parallel world, you'll take A, B, C, 0:18:37.720 --> 0:18:39.840 and D and you'll all run them at the same time, 0:18:40.480 --> 0:18:43.560 but in different parts of the supercomputer, and then through 0:18:43.600 --> 0:18:48.280 software orchestration, you'll sort of coalesce all those outputs and 0:18:48.359 --> 0:18:53.080 render a conclusion based on the set of calculations you've run. Now, 0:18:53.119 --> 0:18:56.679 I gave an example of maybe a decomposition to four pieces, 0:18:56.720 --> 0:18:59.199 but what we really may be talking about maybe a 0:18:59.240 --> 0:19:02.800 hundred thousand or or a million or ten million pieces 0:19:03.400 --> 0:19:06.200 and uh, and it's very complicated to try to orchestrate 0:19:06.240 --> 0:19:09.359 all that activity. So a laptop computer doesn't have the 0:19:09.400 --> 0:19:12.680 ability to do that. And that's why when people think 0:19:12.720 --> 0:19:16.640 about some supercomputers, they sort of render it in terms of, well, 0:19:16.640 --> 0:19:19.640 this is the equivalent of what ten million laptops could 0:19:19.640 --> 0:19:22.159 do or a hundred million laptops could do. But I 0:19:22.280 --> 0:19:27.280 remember laptops or standalone entities. In the supercomputer world, all 0:19:27.320 --> 0:19:30.760 of those computing entities have to be managed, and it 0:19:30.880 --> 0:19:33.560 has to be brain power to orchestrate the way they 0:19:33.600 --> 0:19:37.920 tackle the problem. And the supercomputers are really architected to 0:19:38.119 --> 0:19:41.480 handle that, right, So this is this is a specific, 0:19:41.560 --> 0:19:45.119 purpose built approach to that problem, whereas we've seen things 0:19:45.160 --> 0:19:48.920 like grid computing as sort of an ad hoc approach 0:19:49.040 --> 0:19:52.320 that problem, where it tries to do a similar thing, 0:19:52.359 --> 0:19:58.160 but obviously at exponentially lower levels of processing capability. And 0:19:58.240 --> 0:20:01.119 when we're talking about things like floating point operations. Just 0:20:01.200 --> 0:20:03.280 for you guys out there, you listeners out there, you 0:20:03.280 --> 0:20:06.399 know you might have seen a graphics processing unit that 0:20:06.440 --> 0:20:09.000 talks about things in the Tarraf flop range, which you're 0:20:09.000 --> 0:20:12.200 talking about, you know, a million million floating point operations 0:20:12.200 --> 0:20:15.800 per second. We're looking at Pata flop ranges here, a 0:20:15.880 --> 0:20:19.520 thousand million million floating point operations per second. As I 0:20:19.600 --> 0:20:24.359 understand it, which that's incredible. Uh, it's again for someone 0:20:24.400 --> 0:20:26.760 like me, maybe it's just my limited imagination. I have 0:20:26.880 --> 0:20:30.200 real trouble putting this into a context that I can 0:20:30.240 --> 0:20:34.680 get my hands around. But it's it's an incredibly fascinating thing. 0:20:34.680 --> 0:20:37.840 And this is not this isn't like it's unprecedented. We've 0:20:37.840 --> 0:20:43.840 seen researchers, doctors, scientists use supercomputers to research stuff like 0:20:43.960 --> 0:20:48.959 vaccines for for the flu before as well. Right, well, absolutely, 0:20:48.960 --> 0:20:51.919 in fact, when h one N one came out. I 0:20:51.920 --> 0:20:54.920 guess it was around two thousand and nine. IBM S 0:20:54.960 --> 0:20:59.920 Computational Biology group actually began to model the evolutionary tripe 0:21:00.080 --> 0:21:03.400 victory of the virus because if you think about viruses, 0:21:03.600 --> 0:21:06.480 and I don't want people to be confused that I'm 0:21:06.560 --> 0:21:10.200 equating flu to corona, But if you think about flu 0:21:10.359 --> 0:21:14.400 for a second, the virus is not a static thing. 0:21:14.640 --> 0:21:17.480 It will evolve over the course of time, and that's 0:21:17.520 --> 0:21:19.880 why you have a different kind of flu shot every year. 0:21:20.119 --> 0:21:23.920 It's it's an effort to try to create a vaccine 0:21:23.960 --> 0:21:28.520 to intercept the next generation of where this where this 0:21:28.800 --> 0:21:32.160 particular virus has evolved too. And the way you do that, 0:21:32.200 --> 0:21:34.840 the way the industry does that is they use computational 0:21:34.880 --> 0:21:38.880 techniques to kind of predict the evolutionary pathway and they 0:21:38.960 --> 0:21:42.800 build their vaccine to target where the where the virus 0:21:42.840 --> 0:21:45.000 will be in three months as opposed to where it 0:21:45.080 --> 0:21:48.400 is today, because the lead time to design and build 0:21:48.440 --> 0:21:50.840 a virus is, you know, takes a little bit of time. 0:21:51.320 --> 0:21:53.680 You can't wait for the virus to hit. The same 0:21:53.760 --> 0:21:56.359 kind of logic will be applied to the investigation of 0:21:56.480 --> 0:22:00.359 COVID nineteen UM. You know, depending on us on a 0:22:00.400 --> 0:22:04.240 discovery of science in terms of the extent and how 0:22:04.440 --> 0:22:06.320 it will evolve over the course of time. But the 0:22:06.359 --> 0:22:10.120 expectation is it will evolve, and so you'll use computational 0:22:10.200 --> 0:22:15.320 techniques to begin to fathom that infinite possible ways in 0:22:15.359 --> 0:22:17.560 which it could evolve and choose those that are most 0:22:17.600 --> 0:22:20.679 likely to represent where the virus will be in a 0:22:20.680 --> 0:22:23.760 handful of months, and you'll use that to inform the 0:22:23.800 --> 0:22:26.680 way you design your vaccine to intercept it. So we're 0:22:26.720 --> 0:22:31.480 talking about forming probabilistic models to really determine where are 0:22:31.760 --> 0:22:36.480 the most likely pathways that this virus might take evolutionary 0:22:36.480 --> 0:22:40.120 a lee speaking, it's very similar to how I from 0:22:40.160 --> 0:22:43.520 a concept level, It's very similar to how I would 0:22:43.560 --> 0:22:48.600 look at something like IBM Watson when you know everyone 0:22:48.760 --> 0:22:53.440 knows about it competing on Jeopardy. It had probabilistic approaches 0:22:53.480 --> 0:22:56.320 to which answers would be the most accurate, and only 0:22:56.359 --> 0:22:59.439 if it reached a certain threshold of certainty would it 0:22:59.680 --> 0:23:02.600 would buzz in. But of course, obviously now we're talking 0:23:02.600 --> 0:23:06.560 about a much more complex thing and much higher stakes, 0:23:06.600 --> 0:23:09.480 but it's that same sort of approach of where can 0:23:09.560 --> 0:23:12.720 we predict where this is going, how can we get 0:23:12.760 --> 0:23:15.159 ahead of it? Then how can we create you know, 0:23:15.200 --> 0:23:18.160 a dead version or an inert version I should say, 0:23:18.200 --> 0:23:21.120 of the virus to make a vaccine. And then you 0:23:21.160 --> 0:23:24.800 have the other challenges that come in vaccinations, which is 0:23:24.840 --> 0:23:29.320 just you know, the manufacturing process, distribution, that sort of thing. 0:23:29.359 --> 0:23:33.879 But this shortening the pathway to this part to me 0:23:34.000 --> 0:23:38.440 seems like it is Uh, it is absolutely crucial, and 0:23:38.800 --> 0:23:40.560 it's also one of the areas where I would think 0:23:40.600 --> 0:23:44.200 that you would see the longest delay. So seeing the 0:23:44.200 --> 0:23:48.400 the application of supercomputers is really inspiring to me. Are 0:23:48.400 --> 0:23:53.800 there other ways that IBM is contributing to various efforts 0:23:53.840 --> 0:23:58.920 to either track or fight COVID nineteen. Yes, And in fact, 0:23:59.040 --> 0:24:03.520 just last Friday, a IBM released for free on our 0:24:03.560 --> 0:24:10.919 website a UM, an artificial intelligence package that speculatively designs 0:24:10.960 --> 0:24:15.360 new molecules for the treatment of COVID nineteen. So let 0:24:15.400 --> 0:24:17.240 me back up for a second. If you think about 0:24:17.280 --> 0:24:19.920 what's been going on at oak Ridge with Jeremy Smith's 0:24:19.920 --> 0:24:23.399 effort to look at eight thousand compounds and whittle that 0:24:23.520 --> 0:24:27.960 down to seventy seven for further investigation as potential therapies 0:24:28.000 --> 0:24:32.080 to treat COVID nineteen. Well, those eight thousand existed, somebody 0:24:32.080 --> 0:24:35.359 had already built them. Question is, are there new kinds 0:24:35.400 --> 0:24:39.280 of molecules that could be designed that don't exist today 0:24:39.320 --> 0:24:42.240 that could be used to treat COVID nineteen. So the 0:24:42.320 --> 0:24:45.520 artificial intelligence package that IBM put out on Friday lets 0:24:45.560 --> 0:24:48.199 you do that, and it's free and it's open to anyone, 0:24:48.280 --> 0:24:51.719 So anybody can get on the website and begin playing 0:24:51.720 --> 0:24:54.320 with it, and maybe you kick up a new molecule 0:24:54.760 --> 0:24:57.280