How the Ohio Supercomputer Center Drives the Future of Computing - Ep. 213

0:11Before we get to today's guest, if you're interested in NVIDIA's GPU Technology Conference GTC this March in San Jose, California, and exploring the cutting-edge technologies shaping our future, we have a special offer. Use the discount code AIPODCAST, all one word, when you register for the conference at nvidia.com slash GTC to receive 20 % off your registration. That code again, AIPODCAST. Hello, and welcome to the NVIDIA AI Podcast. I'm your host, Noah Kravitz. Supercomputers are awesome, but not everyone has access to a supercomputer when they want it. At least not just yet. That's why organizations like the Ohio Supercomputer Center are so vital to our collective digital futures.

0:56The Center's open on-demand program empowers Ohio higher education institutions and private industry by providing accessible, reliable, and secure computational services, including access to GPUs and AI-related applications. But there's so much more to it than just that. Here to tell us all about the program, including how art students at Ohio University can use iPads to access supercomputers for design work, and I think maybe even a story about how NASCAR is shifting their workflows, is Alan Chalker. Alan is the Director of Strategic Programs at the Ohio Supercomputer Center, and he's also the Principal Investigator for the Open On Demand program.

1:35Alan, welcome, and thanks so much for joining the NVIDIA AI podcast. It's a pleasure to be here, Noah. Thank you. You know, we try to ram everything into a little succinct intro without spoiling the podcast, give users a taste and our listeners a taste, I should say. And in this case, I know for every word I was saying, I was like, I don't even know half the stories behind these words. So I just wanted to get through it. I'm so excited about this. Turn it over to you. Tell us about OSC and open on demand and, you know, start wherever you want to go, because I'm sure this It's going to be a great conversation.

2:08Sure, happy to. And I think I'll start back at the beginning, the history. And the history goes back to the mid-'80s, and these things called supercomputers were starting to become just part of the general tool set for researchers and scientists and engineers. The National Science Foundation realized that they needed to support those researchers. And so they put out a call to create some national supercomputer centers. There were a group of Ohio faculty members that responded to that call, and Ohio did not win. The national centers are places like the NCSA in Illinois and Pittsburgh Superhero Center in Pittsburgh, places like that.

2:50Those Ohio faculty members said, just because the federal government isn't going to fund us doesn't mean that's the end of the story. And so they went back to the state of Ohio and the governor and the General Assembly and said, well, how about you funding this? and the state said, yeah, it's a great idea. And so we were created as an act of the General Assembly signed by the governor back in 1987 to be a resource for all highlands. And as you said in the intro, it's not just an academic thing, it's also private industry on there. This is an aside trivia question kind of thing, but does that, are you the first state-funded supercomputer center in the US?

3:27Yeah, that's a great question. Yeah, we're one of the very few state-funded ones as well. Normally, state supercomputer centers are either federal-funded by NSF or Department of Energy, Department of Defense, or they're funded at the campus level. So, yeah, we definitely are the state one. We're in this little bit of an interesting niche there as a result of that. Yeah, yeah, very good. Now, before I get into the open-on demand, I want to just talk about what are supercomputers because that's something that a lot of people – Yeah. And I like to give a really good analogy here in that the first supercomputer that we bought, like I said, we were created in 1987.

4:02We took delivery of our first supercomputer in 1989. It was a$20 million Cray supercomputer system that had eight processors, 128 megawords of memory. The important thing is you can measure the speed of computers and capabilities of computers different ways. but it was capable of doing about 2 billion operations per second is what it could do. The cell phone right on my desk right here,$1 ,000 cell phone, has eight processors, way more. I mean, gigabytes and gigabytes of memory. You know, 500 gigabytes of memory. But more importantly, it's probably tens of trillions of operations per second that it could do.

4:43You know, a supercomputer is just something that's kind of at any given moment of time amongst the most powerful computers in the world. and that is constantly changing as a result of new and emerging technology from amazing companies like nvidia you know what they bring to to uh our community so our job at the ohio super computer center is just to provide the latest and greatest top-of-the-line technology uh as a kind of a shared service to the people of the state and the people of ohio that work outside of the state so So collaborators, folks like that. That kind of sets the stage a little bit.

5:25What are supercomputers? It's nothing magical. If you walked into our data center, it just racks of, happen to be Dell computer, Dell nodes right now. It's nothing special on there. Now, to get to the story of the open on-demand and where we're going with some of these things is that supercomputer, scary word, scary thing happening. You go back to the mid-aughts And I already mentioned cell phones and iPhone comes out and consumers and general public became familiar with just using handheld devices and using their laptops and doing online banking and online shopping, all that stuff. And they started coming to the supercomputer centers around the country and around the world and saying, well, hey, if I can do online banking, why can't I do it for the supercomputers?

6:16I need to use the supercomputers. I've got to use terminals, and it's scary. There's, you know, like the hackers in the movies with green text scrolling by, right? So we just, we took that to heart, and we kind of happened to be in the right place at the right time, where we started to develop a web-based interface to the resources that we support and maintain. And because we are engineers and computer scientists and not very creative, we called that OSC on demand. That was the name of it. Yeah. It works. Yeah, it works. It works. And it was a very rudimentary, just you could log into a web page and move files around, upload, download, things that we take for granted right now that we do all the time.

7:02Right. So what happened then? No good deed goes unpunished. We started showing this to our peers and colleagues at other centers around the country at the various conferences and events we go to. And they were like, this is amazing. We need this, too. Can we get a copy of that code? We're like, well, we didn't think about that. This was put together with a baleen wire and chewing gum. The equivalent of that from a code standpoint behind the scenes. So we went to the National Science Foundation and have been privileged to now receive a series of four different funding awards from National Science Foundation to take what we had originally developed.

7:41and because we weren't creative, we said, well, it was called OSC On Demand. What are we going to call it? Well, let's call it Open On Demand. There you go. And so we... It was sort of presciently trendy, right? Yes. And open is open. Open's a good thing. Yeah, yeah. So we've just last year, last summer, we celebrated our 10th anniversary of the platform. Open On Demand is a fully open source. You can go to GitHub. You can download the entire source code platform that basically provides a computer resource or a computer center a web-based access to their resources. Last count, last summer, it's in use about 700 computing centers, 62 countries, every continent except Antarctica.

8:33I'm really hoping to get Antarctica. If anybody's listening, you can help me. I know there's data centers in Antarctica, so we'll get there someday. But it's just kind of become that standard interface that facilitates access. We like to say the motto we have is that open on-demand, it's any device anywhere. And it's true. You can, as a result of this, because it's standard web technologies, you can access these high-performance computers, supercomputer resources from any device anywhere. Now, I have seen pictures of graduate students here in Ohio sitting in a pub with a pint of beer in one hand and their cell phone logged into OSC's version of On Demand in the other hand.

9:21Now, at least here in Ohio, it's not illegal to drink and compute, but we don't necessarily condone it. Yeah, no, we're just telling stories. Yeah. Yeah. So I think that hopefully gives a pretty good context of what our mission at the Ohio Supercomputer Center is and what the purpose and what we're trying to do with the open on demand platform. Absolutely. And as eager as I am to have you jump into what's going on now in the current, let alone talking about what may be coming in the future for supercomputing in general and NOSC, of course, I'm curious. It's been about 40, 35 years maybe that the center has been around.

10:00And you talked about in the aughts going, you know, the web interface and people saying, you know, we've been coming here, but you have supercomputers. And now that I can online bank. what what were people doing with supercomputers in you know the the mid-80s versus the mid-90s versus the aughts um even versus 10 years ago how how have the the uses changed over time uh that's a great question um so the the very early it was i would say just a handful of what we'll call the traditional core sciences. So things like astronomy, obviously, they're modeling all kinds of things with black holes, things like that.

10:46Computational chemistry has always been a pretty big one, trying to figure out. So things are really big, really small. You can't really do an experiment on a galaxy. It's kind of hard to do some of the experiments you want to do on molecules and atoms. So you do that. So those are the types of things that have always been the bread and butter of supercomputer kind of the prior things. Now, what's happened in recent years, the past 10 years, is because we've lowered that barrier to entry by having easy-to-use web access, it's just blossomed across the board in terms of the scientific disciplines that use it.

11:25But last year, I think here at the Ohio Supercomputer Center, we counted 42 different fields of science that our clients were from. We had about 7 ,000 individuals that logged into our resources over the course of the year from several hundred different organizations. And it would be amazing. The types of fields that are using us now, I wouldn't even imagine before. Sure, of course. Horticulture and crop scientists are using us. Yep. No, and it makes me think of the, you know, six years or so now I've been hosting the podcast. And just that, yep, we had somebody, you know, back when with an agricultural robot using computer vision to detect, is it a weed or not a weed?

12:13And astronomers. And yeah, yeah, it's just, it's wonderful. It absolutely is. Absolutely is. Yep. So tell us about Open On Demand Now. And I mentioned at the top the involvement with the Ohio University Art Design Program. There's a NASCAR story. You've got your hands in. And as you said, it's important to note, it's not just higher ed. It's industry. It's all kinds of uses. Exactly. Exactly. Yeah. So let me give the Ohio University story. And I'll spoil the lead a little bit here, but then we'll get into detail here. And there's a stream of words that are going to come out of my mouth here that are going to make perfect sense to us and many of the people that are listening that just a few years ago would not have made any sense at all.

12:57I want to point that out here. So right now, as we're talking— Is one of them Riz? Because I don't know. No, no. But right now, right now, as we're talking, there are art design students in Athens, Ohio, at Ohio University, taking an art design class from Professor Basil Masarazada, who he's having them with their iPads connect to OSC's on-demand instance. they are tapping a few buttons and they are launching stable diffusion which is a generative ai uh application out there they type in some prompts say i want to you know uh design a i don't know a an image of an astronaut riding a unicorn on mars who knows something like that They hit go, and behind the scenes, that fires off onto our supercomputers a Kubernetes partition where we have some third-generation old – these are Volta NVIDIA processors.

14:04So it's three-generation old processor or GPU cards that they are time-slicing sharing. A few seconds later, they get back a generative AI image of an astronaut riding a unicorn or whatever they prompt. Now, they have no idea the amazingness of what just happened behind the scenes and that was facilitated. And they don't need to, but they're benefiting from this amazing technology and this whole workflow and stream of stuff on there. It's really quite impressive. And again, that stream of words would have made any sense just a few years ago. No, it's so much is so much has changed so fast and is still changing.

14:48So since we're talking about an educational use, does the center offer, you know, programs, educational programs, training programs, that kind of thing? Or is it kind of strictly I don't mean strictly in a negative way, but is it kind of strictly a compute on demand resource, if you will? That's a great question. And we absolutely have education and training as part of our mission, because one of the things that we realize we need to do is, again, our mission is to provide support for Ohioans. And in particular, right now, we've been called the Silicon Heartland. I'm not sure if you've heard that, but that's a phrase.

15:32There's a certain other company that made a large investment recently in terms of big fab in Ohio. And so we're the Silicon Heartland. And we, part of our mission at the Ohio Supermature Center is to help ensure Ohioans stay at the forefront of being that kind of tech savvy, you know, and benefit from technology as a society across all aspects, education, business, you name it on there. So a good example of a training program that some colleagues of mine did just over the past year or so is, again, another NSF-funded thing. But we did a whole bunch of AI workshops where we were bringing in people to teach them generally about, okay, AI.

16:18AI is moving so fast, obviously. And just the stuff they covered even a year and a half ago, some of it has been supplanted. But we're trying to constantly teach them that. So this is a great example. Let me go back to the art design students here. Now, we don't have necessarily a lot of in-house experience with stable diffusion. It's evolving. But as we're talking to them and as we're going through with them, we're realizing we need to have more experience. We recently just hired a dedicated machine learning engineer, my colleague Evan Jaffe, who has experience in this. And he needs to kind of provide that expertise.

16:59How big is your staff? Oh, great question. About 30 people, about 30 employees in the core staff. Now, I work at The Ohio State University, so we rely upon the broader institution for things like HR and business and stuff like that. So that's really just the core technical staff there. And as you were kind of alluding to with your new colleague, I was imagining that over the years, kind of the shape and sort of disciplinary focus of a staff like yours must kind of, I don't know, kind of ebb and flow like an amoeba or something, adapt to the times, right? Absolutely, yeah. So, you know, we have an awful lot of cybersecurity experts and engineers now that we did not have 15 years ago.

17:45That was not even a concern there. I want to real quick go back to the Stable Diffusion app for the Ohio University students because I want to give you a bit of a peek of where we're going and what's happening there. So, you know, we stood up this application and the students are using it and it's great. Now, one of the things that's happening with Stable Diffusion and lots of the generated eyes is they're constantly evolving. They're constantly new versions coming out. And so we're having to now go and deploy new iterations of the model, the StereoDefusion model, for the students to use. Well, that means that the students over time are going to be getting perhaps a very different results from the same prompt, depending upon which version.

18:35They hadn't thought about that before, right? And so one of the things that we're incorporating into the application is the ability to select which version of the model they actually are running against. And it's educating them about the fact that, hey, these models change over time. Of course. So it's kind of this cyclic nature of, okay, what do we need to do to provide better resources to you so you can accomplish what you're trying to get done in your educational goals? But also, oh, we're exposing new things that you hadn't thought about and had to consider. No, it's fascinating to think, and it reminds me of, you know, the bubbles we all live and work in and such, and that I'm thinking, oh, right, well, I keep hearing 2023 was the year of the LLM.

19:182024 is going to be about kind of pitting LLMs and other models against each other to get them kind of, you know, not quite GANs, but getting them like working against each other to hone in on the most accurate or the most, you know, kind of the best response to your prompt kind of thing. And then I kind of go back a few steps and I'm like, right, if I wasn't hosting the AI podcast and I was an artist, I wouldn't, that wouldn't, I'm making my art. And so that whole idea of just the tools, not just becoming digitized, but as you said, kind of back in the aughts, I did a lot of work with cell phones.

19:54And we talked about it being kind of the age of beta because products would get released, but then they were built to be updated. Now it's alpha. What's before alpha? Yeah, something like that. Now, if you don't mind, there's another aspect of this that I want to speak to in terms of the future. So currently right now, centers like ours have many more traditional CPU resources than we have GPUs. Right. I was wondering, you say it's just a data center, but I was kind of curious what's going on in there. Yeah, so we have about 55 ,000 cores across our various systems that we provide, and about 400 GPUs, to give you the order of magnitude of that.

20:42Now, these Ohio University students, there's like 100, 150 of them that are in class and hammering on those GPUs, and we only have 400 or something, right? So that was part of the motivation for us going to Kubernetes with the time slicing, where they are sharing a GPU because they're not doing something that's that intensive on there. But you better believe it. It's coming. More and more of these classes are going to be asking for things like this. And they're going to, once they get beyond their undergrad into their graduate, they're going to start saying, wait a minute, I want to do much more intensive stuff.

21:20And so we're having to look now, how are we going to invest in more of the GPU technology? We've got some of that coming down the pipeline. But then how do we also make it accessible via scheduling limits and restrictions? Because it is a resource that has to be shared. Right, right. No, it's interesting because I was thinking about when you talked about the Silicon Heartland. And way back in the day, I lived in Ohio as a kid and I've been out on the West Coast for a long time. But I was thinking about things like broadband access and rural communities and wondering if, you know, someone maybe not in a pub on campus, but somebody a little further afield on a cell phone accessing the center is the bottleneck, you know, with the Internet access is the bottleneck with everybody wanting GPU access.

22:11You know, where do these things come from and how do you plan for the future? So it's interesting. A lot of moving parts. Yeah. And just to answer that, really, the network connectivity is not the bottom. It's not an issue. It's not at all. No. I mean, it's amazing, be it cell phone, be it Wi-Fi, be it whatever. It doesn't require a ton of bandwidth because it just... It's just an ant-text and an image. Yeah, exactly. Pushing some images across there. There's some wonderful technology for VNC and things like that. Right, of course, of course. My guest today is Alan Chalker. Alan is the Director of Strategic Programs at the Ohio Supercomputer Center, where he's also the PI on a program called Open On Demand, which is giving Ohioans in higher education and private industry access to supercomputing resources remotely through a web interface that, as Alan was talking about, was developed.

23:10going on 15, 20 years ago now almost. We talked before we started recording, and I think I mentioned it earlier about some of your other clients. One of them is a household name, I believe. You guys have been working with NASCAR. Yeah. Yeah. It's a great story. So NASCAR, the company, not necessarily a particular team, NASCAR company has been a long-term client of OSC. And then they actually are a client of OSC through a small Columbus-based consulting company called TotalSim, who their focus is in what's called computational fluid dynamics. And so that's just modeling airflow over things. Now, what TotalSim has developed over time are apps within the open on-demand platform that allow the NASCAR engineers to do whole simulations of race cars.

24:04And the reason NASCAR does that is because NASCAR, they really have two kind of competing goals in general. They are an entertainment business. They want fans to be entertained, which means they want to ensure that the races are exciting, they're competitive, they're neck-to-neck, all kinds of those. But at the same time, they want to ensure that there's safety and that they know exactly what's going to happen in the performance. So they are constantly tweaking the rules just all the time. They're tweaking the rules because there's new developments, new changes, things like that. In the before times, I'll say they would tweak a rule and they would have to make a modification to a vehicle and go take a test driver out on the track and run it around and see what happens.

24:58And remember, sometimes these tweaks are so minor that it's only going to show up after a few hundred miles of racing. Yes. So what do they do now? Well, what they do is they do this all virtually first. They make a design tweak, something that might be a structural design to a car chassis. They create a model, a virtual model of it. And logging into OnDemand, they put it onto our systems and see what the result is in terms of they can look at very minute details in computational fluid dynamics. Okay. Now, here's the cool part of the story and what's happening. They've been doing this for years.

25:38And let's just say that they, roughly a typical stock car model, I don't know, 10, if not 100 million little elements. It's breaking it down into a mesh of elements. And to do a traditional simulation using traditional CPUs, 1 ,000 cores it might take for half an hour to run that. There's lots of dials and you can turn. but let's just give you an idea of that. And so 1 ,000 cores, half an hour, what you're talking about there from just a pure hardware investment standpoint is quarter of a million dollars potentially of hardware that they're using for half an hour. And they're starting to do some testing, exact same procedures, but benefiting from just some of the amazing optimizations that NVIDIA has been working with with some of the software manufacturers to get some of these CFD software to run on the GPUs.

26:39And they can do exactly what I just described. Same timeframe, three NVIDIA A100 cards instead of a thousand cores. So instead of a quarter of a million dollars, you're talking something like$30 ,000,$40 ,000 worth of hardware there. That's an amazing story. And that's also showing now that, again, our investment, where our investments need to go, because that means those thousand cores aren't constantly being tied up and other clients can use them for other things. So they're looking at that transition and starting to move towards that. It's a great story. The efficiencies, the cost savings, the additional resources, it's, yep, it's great.

27:25It's great. Alan, you mentioned one of the things that you're, I don't know what your timeframe is, but thinking about, you know, bringing some more GPUs into the pipeline and that kind of thing to meet growing demand, but kind of bigger picture, what's on tap for open on demand this year, or even kind of looking ahead, you know, sort of longer term. I mean, if you want to pontificate about the future of supercomputing, I'm all ears and I know the audience. Well, let me, let me, yeah, let me talk first about the open on demand. Then I'll talk about a little broader field here. So I mentioned that we've had a series of National Science Foundation awards.

28:00And the most recent one, thank you, the most recent one actually just started in September. And it's a very interesting award because it is not a technical award. It is a business award. What the National Science Foundation has realized is that over the years, they have funded a variety of cyber infrastructure programs like Open On Demand that are getting really large communities and being rather successful. And they want to make sure that these don't kind of wither and die on the vine if there isn't future NSF funding for them. So the program that I got is part of what's called their Pathways to Open Source Ecosystem program, POSE.

28:43And the idea is they want to facilitate the transition of something like open on-demand into think of any of the really successful open source ecosystems out there, be it Linux, Jupyter, Python, you know, Android even. So what we're doing for the next two years under this award is really more of a kind of a governance, a business model thing to make it so that open on demand, there's a good community. We have contributors from around the world there. But how do we make that ecosystem? How do we make it so that there's various people contributing to the roadmap and the pathways of features that we're developing, all of that?

29:27That sounds great. Yeah, so very excited about that. Now, there's all kinds of technical features I could talk to that we're going through. But really, that's the best. We're going to continue to meet the needs of the community and make sure that we're addressing that. Now, when it comes to the kind of supercomputing in general, I want to hark back to what I said at the start of supercomputing is just at any given moment of time, the most powerful computers in the world. Sure. Constantly evolving. And so I've lived through actually a couple different interesting transitions in terms of our first supercomputer was a Cray computer.