SC24 | Camilo Buscaron, Alafia AI

Clips
News
More from SC24

Camilo Buscaron

Founder & CEO

ALAFIA

From theory to reality: AI-powered PCs transform precision medicine

Artificial intelligence is revolutionizing personal computing, with AI in healthcare driving a transformative shift toward purpose-built systems designed for specialized industries.AI-powered PCs are emerging as game-changers, leveraging cutting-edge technology to modernize infrastructure, improve efficiencies and unlock innovation in areas such as precision medicine. This trend reflects a broader move away from one-size-fits-all solutions, tailoring hardware and software to meet the unique demands of specific sectors while supporting the development of future algorithms, according to Camilo Buscaron (pictured), founder and chief executive officer of Alafia AI Inc

play_circle_outline AVAS Supercomputer: Revolutionizing Precision Medicine and Healthcare Infrastructure

play_circle_outline Transformation of computer architecture for AI algorithms

play_circle_outline Importance of AI in healthcare and high-performance computing

play_circle_outline Accelerating Product Development with Innovation Cycle and Hardware/Software Optimization for Enhanced Performance

Info
Transcript

Camilo Buscaron, Alafia AI

Camilo Buscaron

Founder & CEO ALAFIA

Day three of the coverage on theCUBE in Atlanta, Georgia. AVAS, an all-in-one personal supercomputer for precision medicine, is featured. The focus on healthcare and precision medicine is explained. The AIPC is discussed as a PC with advanced capabilities. The evolution of computer architecture and algorithms for AI is highlighted. The innovative Omics computing model is described to process biological data efficiently. The product's optimization for regulatory compliance in the medical industry is mentioned. The hardware specifications and software optimizat... Read more

explore Keep Exploring

What were the main reasons for focusing on healthcare and precision medicine with AVAS? add

What are some factors that have remained consistent in computer architecture but have seen advancements in technology and capacity over time? add

What is necessary in order for us to be able to run future learning algorithms on computers effectively in the field of HPC? add

What are the specifications and features of the product being showcased, and how has the software stack been optimized to leverage the hardware architecture? add

bolt Powered by CUBE AI

Camilo Buscaron, Alafia AI

search

>> Good morning high-performance computing fans and welcome back to Atlanta, Georgia. We are here kicking off day three of our coverage on theCUBE. My name's Savannah Peterson, here with John Furrier. It's our third year doing the show together. Day three, every time we get smarter.

>> Day three is always kind of like you got the energy, you're amped up. Your adrenaline's pumping. But this segment is to me super exciting because we have the first AIPC that we've seen out in the wild. A true IPC has got H-100s in it. I don't know how many GPUs, thousands of them. So it should be great.

>> It will be great. And we actually have something tangible, we get to look at it. Cam, thanks so much for hanging out with us this week.

>> Hey guys. Thank you. I'm honored to be here. If we go back to the first time I came up to theCUBE, I disrupted one of your interviews with a robot dog, so I figured this time watch for permission, you know?

>> Yeah.

>> You are always here to bring us the toys. I guess I should let you introduce the other guests on the stage as well. Cam, tell us about this computer.

>> This is AVAS. This is our all-in-one personal supercomputer, essentially an AI appliance for precision medicine.

>> That is awesome. So personal supercomputer. And well, we're going to really get into the cool stuff here, but I want to ask you a question because I think you've done something incredibly smart with your product development. And rather than market to the entire world, which is what you hear a lot of the companies trying to do here, you focused your efforts specifically on healthcare. Tell me why you made that decision.

>> I think there's three main reasons why we focus on healthcare and precision medicine. I think if you look at the healthcare ecosystem, they have been under-invested in their infrastructure for a long time. That's why we ended up building this product itself. Right?

>> That's a great point.

>> Yeah. They have under-invested infrastructure and then there is a desire to accelerate and improve the current efficiencies and the current processes. So ultimately, bringing that infrastructure up to the standards that we're used to in enterprise AI or any other enterprise system adds a lot of value. So that's led us to this product.

>> A lot of people have been talking about the AIPC. You actually call it the personal supercomputer, which basically has-

>> Which I love.

>> NVIDIA love because NVIDIA is in it. So talk about what's in it because this is again a PC, the old PC was you did stuff on an office, email, you got a laptop, you got a device with it, you got a mobile phone. iMessage kind of connects to that, but WhatsApp comes out. So you started to see now the connected experience with the user. This is like a power workstation slash like the old PC model. Tell us something big.

>> Nearly 75 years ago as the Allied Forces were liberating the camps of Dachau and Auschwitz, the greatest computer scientists of all time asked a very simple question. Do you know what that question was now?

>> No.

>> And we're talking about Alan Turing, right?

>> Oh yeah, I was going to say it's obviously Alan.

>> Alan Turing. But he asked a very simple question, which was can machines think? And to ask that question at that time, it's unheard of because there were no machines, right? If you fast-forward, that inspired the pioneering computer architects to create our computer model that has evolved over the last 50 years. But that model, the Binomial architecture model, it hasn't changed very much. The memory architecture, the compute modules, how it interfaces across buses, that has pretty much stayed the same. We have increased our capacity. We have increased our technology. We have improved levels of the technology, but that model hasn't changed. If you look at the algorithms, we're talking about the algorithms that are both today for AI, there were algorithms that were developed 30, 40 years ago that did not work at the time that were conceived, but now are starting to work because we have a modern computer architecture led by the GPU, right? If we talk about HPC, in this space, I think it's safe to say today that AI is the killer app for HPC. If you would've said that in this conference 20 years ago, it would've been heresy, right?

>> Right.

>> So with that thinking of how do we improve the computer itself to be able to run not the algorithms that we have now, but the algorithms that are yet to come, right? For us to unlock the next stage, we have to have software engineers and computer scientists create new learning algorithms, and we have to sign a computer from the ground up with the best technology that we have in our hands today just to support that next movement of technology.

>> And that's why NVIDIA's earnings we were talking about before we came on camera, is selling so much product because people are preparing for that infrastructure for the killer app, AI, which has more killer apps on top of it, search today, thinking tomorrow. Okay. Knowing that, okay, let's believe that. What's the approach to building the first PC supercomputer? I don't know. I mean, I call it AI, super computer, I like that name, but it's essentially a PC. It looks like a PC.

>> I mean, ultimately we're creating a computer model and we call it Omics computing. Why do we call it Omics? Omics is processing biological data from a human, and the amount of data that you get from a human to be able to do a digital twin of having your, let's say, all your medical imaging, your blood samples, your genomics data, any sequencing data, that amount of data, it completely breaks the current computer architecture model. Our computers were created with parts that we had available for basic data structures. So our ultimate model is to translate the biological language, digital biology, into data structures that are processed very well on our computer. So what we've done in this architecture is we've increased bandwidth across every layer. We have the highest memory capacity with the highest amount of compute possibly that you can possibly have. So that processing is state-of-the art. As we learn those data structures that map from Omics to a computing data structure, we can then create and innovate a computing model for the next generation of architects.

>> . So I mean, I love this, the medical industry, having done medical device, it's incredibly hard to get approval, compliance, HIPAA, there's a lot of different factors that you're dealing with. You didn't exactly go into an unregulated niche with this.

>> Absolutely.

>> So tell me a bit about how you've designed this to be able to operate in that environment.

>> Absolutely, and that is actually a tell-win for our business at the moment. Actually the regulatory environment has gotten much better over the last 10 years. I started my career building medical devices, cardiology, surgical tools. It took 10 years. We were very successful because we would sell the IP to the large manufacturers. But getting the technology done, getting it to regulatory, getting it to the hands of a doctor took at least a ten-year cycle. I think that today, that regulatory environment is much better, is shown over the last year as well. The amount of software as a medical device, applications and approvals, and I think the outlook over the next four or five years is going to be a... I mean, I think the economic boom we're going to have partially led by a pro-business regulatory environment, I think it's going to be very helpful.

>> So Cam, talk about the product because we talked before you came out, you've been here all day and you're showcasing it here in the booth. It's our first ever cool feature of theCUBE where we present something we think is innovative. I asked you what your vision was and you said, "Hey, we built the bot because no one else was," but you're a software company. Now, people might not know, but you worked at NVIDIA on DGX, you know the playbook, software and hardware together. NVIDIA claims they're a software, although they sell a lot of hardware. The software's the key. So no one has this, just I want to point that out. So how did you do it and what's in it? Give us the speeds and the feeds and talk about the software because certainly this is going to be a template for others to follow, I guess immediately.

>> Absolutely. I mean, quickly on the speeds and feeds, we have 256 CPU cores, over 47,000 GPU cores, H-100s, MB-linked, four terabytes of RAM. Just at the high level, over a thousand gigabits per second of memory bandwidth. That's actually what limits a lot of the AI algorithms today is actually memory bandwidth, between the CPU, the memory, and the GPU. So we've optimized, we have an optimal architecture for that type of problem. But like you said, really what we've done, so that's at the hardware level, but what we've also done is optimize every layer of the software stack to leverage that architecture from the kernel to the operating system to the machine learning frameworks ultimately so the applications that we're working with, which that's what I like to talk about, because that's what I'm most excited about, perform the best they can in any environment. I mean at the kernel level, just one quick example, the typical kernel operating system has a four-page kernel. We have L1 cache on a processor that's 64 kilobytes. So we can have a different operating system that runs at 64K page, that's 16 times faster than the regular operating system. So just we got that level. And then there's hundreds of others.

>> It's a dream machine basically.

>> And then all those little things add up and then they give you the performance that we're looking at.

>> And that's the biology, that's the whole taking the neural network concept of the brain, but from biology, human, building the data structures to make it better for the application. Now, what are the target applications that you're using right now? What's the use case? Who's the user and what are they doing?

>> So a lot of our use cases are digital pathology. I don't know how much you know about a pathology, but traditionally a pathologist or a hematologist would take a blood sample or a tissue sample, look at it on the microscope. That's not going to happen anymore, right? Now there is this big scanners where you put the sample, the machine takes a very large picture over 100,000 pixels about, 100,000 pixels, and then that image gets processed by algorithms. And that's what we're doing and that's a use case that we provide performance no one else can today. Another use cases, neurology. Neurology research. We're working with some fantastic researchers at Stanford Medical School, Harvard Martinos, and they're doing what something called TMS, Transcranial Magnetic Stimulation. So they have a magnet treatment mostly for depression. There's a few other indications by the FDA. And what they do, that problem is actually very exciting because they have matrices that are a million by 1 million by 1 million, and you can't process on any regular computer. That's why they need the memory capacity that we have to be able to simulate and produce a numerical result necessary to have an effective treatment.

>> We fundamentally believe that AI is going to save lives. I think your AI hardware here is also proving that. Where are you in the product development journey and when is this going to be everywhere so that this can happen in all the hospitals and medical facilities around the world?

>> Absolutely. So if you look at the latest wave of enterprise AI adoption that probably started fall 2022 with the general purpose transformer going mainstream, right? Most enterprises already had a budget set for 2023. So most companies that were actually trying to sell products, AI solutions and AI product, they didn't have the entitlement to sell to enterprises. We just happened to be at the right place at the right time because we started building our product in the end of summer 2023. We started shipping early pilots that fall, early 2024. When we started shipping now this past summer-

>> .

>> Most health systems, enterprises, imaging centers already had a budget for AI adoption. That's how we've been able to build a nine figure pipeline in less than 18 months.

>> I was just going to say, hold on. You've only been working on this for 18 months?

>> Not quite, less than 18 months.

>> Less than 18 months, okay. That's impressive. In hardware land, for the record, that is... Hardware, the fastest you can get out is basically like a baby is nine months with supply chain.

>> It's a two-year cycle.

>> Yeah, minimum. I mean, I think a lot of the time you're inventing a personal supercomputer in a short period of time. This isn't your first rodeo in technology. You have a pretty robust background.

>> Hardware cycle, everyone knows it's at least a two-year cycle. You can't make a chip or a system in less than two years. And at Computex, there was a lot of talk about going down to one year cycle. They're bluffing. You can't do that in that amount of time. But what I'll tell you, six months ago we announced the most powerful personal supercomputer out in Better World. It was 128 cores, two terabytes. That's the first time we were showing. Six months later, we're here with something that's actually twice as powerful. So yeah, I mean, that's the six months cycle.

>> That innovation cycle is amazing, honestly.

>> Well, you had a little inside baseball cheat sheet, cheat codes because you're at NVIDIA. So the software, what you did was clever. You took the NVIDIA stuff, knowing what you know, built it into a system and just optimized the software stack. I mean, that's the secret sauce and that's what everyone's doing. This is what the show's about. Get the software stack and then figure out how to build a system with what you can do.

>> Well, I'm originally Cuban, so we come out of the womb playing baseball.

>> Playing baseball and thinking about high-performance computing. Is that in your case?

>> Well, I said inside ball, right? Whatever chance you get on base, we hit a home run with this because I think we're so proud to have you take some part of our booth. Again, this is our first time we've ever showcased a what's cool. We think what you're doing Cam is totally cool. And again, at re:MARS, the Amazon event, you had that robot. We knew you were cool then. You're even cooler now. Thanks for being part of theCUBE.

>> So I mean, I am assuming this is a yes. Knowing you and your brain, are you going to be able to innovate at the same clip for the next cycle?

>> Yeah, we have a road map because our suppliers also have a road map. We are able to craft a road map. If you look at the things that we're launching, we gave a sneak peek, but there's obviously a cycle that's coming next year. I think for us is getting this into scale production. We went from prototype to engineering validation to design validation. Now we have a design that's ready to do mass production. So we will get this design in mass production, but the chassis itself can leverage any architecture that would come 2025, 2026. So yeah, we'll have more about that next year.

>> Okay. Well, that leads me to my last question because you just teed it up. When we have you hanging out with us at Next Supercomputing, or maybe even sooner than that, quite honestly, what do you hope to be able to say then that you can't yet say now?

>> There's a few things I would say. So I'm very excited about the benefit that our users are getting. And for most of the use cases that we're looking at, radiology, visual pathology, neurology, we're looking at something called radiotherapy, which is volumetric segmentation, mostly used for cancer treatment planning. That's something that we're learning very well now. It's a lot of Monte Carlo simulations. We are at the place where Monte Carlo simulation got solved. Those use cases, unlocking those use cases and adding benefit to those users is what I'm most excited about and I'd love to next time perhaps instead of having the computer, have one of our customers talk about it.

>> Yes, we would love that, Cam, and I just want to say thanks not only for sharing your story and being an incredibly brilliant mind pushing the boundaries here, but also you've been such a joy to share this space with all week, and we look forward to doing a lot more of it. And John, thank you for hanging out for this segment, and thank all of you for tuning in wherever you might be on this beautiful day. We're in Atlanta, Georgia here, day three of Supercomputing 2024. Lots of coverage still to come. My name is Savannah Peterson. You're watching theCUBE, the leading source for enterprise tech news.