 Welcome back to SuperCloud 4. We're here exploring the intersection of AI and the ever-expanding cloud universe. Importantly, as we've said for years, we believe that over the mid to long term that AI inferencing at the edge is going to become a dominant trend and dominant workload that will change the economics of computing permanently. It's going to bring new innovation to enterprises globally, led by low power, low cost and highly performant arm-based hardware and software architectures. And we're pleased to welcome Krishna Rangasai who's the founder and CEO of SEMA AI, a company that focuses on solving for the embedded intelligent edge supporting virtually any computer vision application with a truly revolutionary ML system on chip architecture. Krishna, welcome to theCUBE. Thanks for coming on. Thank you, Dave, I appreciate it. So tell me about your company. Why did you start the company? I'd like to ask founders that question. That's a great question. So I previously was in public companies all my life and the thing that I observed was AI and ML were really beginning to take off in the cloud and in the mobile form factor and we have seen trillion dollar entities really come together. And one thing that I observed was the physical world which is everything in between the cloud and the mobile footprint. Things we touch on a daily basis, robotics, industrial automation, medical, automotive are really being left behind. And I wanted to build a company that really would build a purpose-built platform that could really bring AI and ML and scale to the physical world or in other words, bring the industrialized world which is stuck with technology that's 40, 50 years old architecturally and bring them into the 21st century. Yeah, so obviously you saw the smartphone, the Apple iPhone. The last thing you want to do is try to go head on there but presumably you looked at the trends, the volume trends in the economics and said, wow, what if we could apply this to the physical world? Is that the right way to think about your company? Yeah, it is. And so if you look at the cloud and the mobile form factor, a handful of few companies are really driving innovation. So the hyperscalers are clearly driving things in the cloud, seven or eight of them. The mobile form factor is really fundamentally two or three companies of the world. And today it's really becoming a world of the have and have nots. What is very interesting to me is the physical world that's everything in between that we talk about is 30, 40,000 customers globally. So the challenge is that how do you bring AI and ML and how do you democratize AI and ML to all these companies? So they could all in a way take advantage of AI and ML for their application and platforms. No doubts, very hard to do. But if you really want to have a meaningful impact on AI and ML, you really need to bring a democratized outlook. So it touches all of our lives in many, many more ways beyond just the cloud and the mobile. And when you look at the traditionally Moore's law, let's say, let's be generous and say it's 40% performance improvement annually. If you look at, just take the Apple A series and you combine the performance of the CPU, the GPU, the NPU, the accelerators, and you're up over 100% annually. So you've got this amazing performance, but you've still got the latency challenge. And so maybe you could talk to how you're solving for that. Right, so I think one challenge in throwing everything at the cloud is not every application is built for the latency. A classic example all of us can relate to is you want cars with semi-autonomous or autonomous capability. As the data is being captured, you're driving at 65 miles per hour, you cannot afford for the compute infrastructure to be away from the car, go to the cloud and then come back with the decision. You're really, really moving too fast. Same goes for robotics, same goes for medical. So there's increasingly gonna be a trend in that you wanna get processing localized to where the data is being created for mission-critical or safety-critical applications. And not everybody can suffer the latency loss that I think people can look through. The reason why the cloud has been used for these applications so far is nobody has really built a purpose-built platform to enable high-performance compute at low power and an ease-of-use manner in such a way that the data and the data processing can all be done locally. And that's really the thesis of what we start to do at SEMA. Why, let's talk about ML on a chip. Why did you decide to do a software hardware solution? I mean, there are examples, certainly Steve Jobs bought into Elon Musk. I would even say Larry Ellison in the enterprise, but it brings its own challenges. Why did you decide to go that route? Yeah, no, I mean, so we initially deliberated if a software-only solution could really bring this together. And we quickly came to the conclusion that while it will bring things closer, you really need a combined software hardware solution because you need to be able to do three things. And this is something that we did well at SEMA. You need to be able to create a software front-end that solves for any computer vision application. So you could take any application, any model, any ML framework and traditional legacy CC++ codes. So the any in the front-end solution really is a software solution. But what brings it together, the hardware and software is performance per watt. So unlike the cloud where performance is skim and where power is maybe given a very loose second or third consideration, the large difference in the edge is that power is the most important criteria. So people want high performance compute, but it better be delivered at a form factor of five watts, 10 watts and 20 watts. Very few edge applications can take more than that either from a power consideration or a thermal consideration. The only way you could do that is really building a joint software and hardware combined solution. And there's no other way around it. And so that's why we decided to take that on. The third portion of what we do is really a push button experience where our customers at the edge want the ML experience, but not the learning curve associated. So they really need instant gratification and really an ability to quickly and deploy solutions to the market. And so we call it the any 10x push button. And for anything that's deployed at the edge, I would submit to you need to do these three things really well where it could solve any problem be 10x in performance per watt compared to legacy architectures or legacy solutions. Third, really deliver a push button solution. We came to the conclusion that the only way we could do it or the only way anybody could do it is really in a combined software and hardware solution. And this is why we believe that that ultimately is going to seep into the enterprise because you're right, performance per watt, maybe it's a secondary consideration in the data center today, although increasingly it's front and center with all this AI madness and GPU, that's going to become more and more important. And at the volume that you have the potential of achieving and certainly ARM has achieved from a wafer volume standpoint, the economics are just going to overwhelm traditional general purpose computing. Do you agree with that? Absolutely. And so I think very few people recognize. So if you look at the edge today, though the problem is distributed across 10,000, 20,000 customers, it consumes $40 billion in semiconductor content on an annual basis. So that's a lot of money. And my thesis is that I think as purposeful platforms are built for the edge, the edge is going to be four to five X the size of the market today. So in other words, inherently, while the cloud will continue to grow, the edge, if you really think about it, is a much bigger market than the cloud. So the cumulative volume that really gets shipped there, though distributed across many participants really makes it a very attractive market. Yeah, and again, volume economics who ultimately would determine winners. What are the use cases that you guys are going after? I see in your website, it's smart drones and intelligent robots. I'd love to talk more about autonomous vehicles, autonomous systems, explain that a little bit. Absolutely. And so like we talked about, we are building a purpose-built company to bring the physical world into the 21st century. So early adopters for us have been robotic companies, industrial automation, factory floor automation companies, industrial drones, smart vision systems, medical customers, and increasingly, I think we're now gen two and beyond, beginning to focus on automotive applications as well. So I wonder if we could talk about autonomous vehicles, fully self-driving, where do you stand on that? I mean, you certainly Tesla going out and built its own, designed its own chip, added some value on the NPU, allowed it to sort of reduce its reliance on LiDAR, so that's a use case that we've looked at pretty closely and it's kind of caught our attention. Where do you stand? I mean, every year Elon says, we'll get the full self-driving in a year. I don't know, that doesn't happen, but the world is waiting for that, although it's challenging, right? These are not necessarily learning systems. What's the reality here? Yeah, no, I think there's a lot set, a lot written and no doubts, more than a hundred billion dollars have been spent on this problem so far. I would say this is probably technologically one of the most complicated problems to solve. You are beginning to get bounded environments like cities like San Francisco, where a way more crews really have L4, L5 systems navigating around, but no doubts. I mean, I think while the technology has made a lot of progress, it's inherently such a difficult problem that it's bound to be a difficult situation for somebody to guarantee that there will be no accidents or no casualties or no challenges along the way. And so I really applaud these companies that are pioneering the space for us. I for one think that this technology is still far out. It's probably gonna be 10 years out before we see fully autonomous vehicles, even in city environments, let alone on open terrains and navigate themselves, where one day the car industry would move to autonomous vehicles. And I fundamentally think electrified vehicles is probably a lot more tangible closer target than really fully autonomous vehicles. And having said that L2 plus and L3 systems, which are pseudo semi-autonomous vehicles, that I think is gonna be a massive volume market. And it's gonna be in the tens of millions of units. Almost every vehicle is probably gonna take advantage of a semi-autonomous mode. And I believe that technology is here today and it's pragmatically something we can deploy in efficient cost, structure and volume. And I believe that starting 2025 or 2027, in that time horizon, you're gonna see almost every major vehicle really pick a L2 plus horizon. So as a company, we are focused on solving the L2 plus problem. We have a lot of interest in our technology from folks in the L4 and L5. And I'll leave you with one thing that really is very exciting as to why many automotive companies are beginning to wanna talk to us and engage with us is, if you wanna get L2 plus L3 systems today, the peer technologies are anywhere from 300 to 600 watts, depending on the configuration, depending on the use model, we can show to all of our customers a clear line of sight where they can get the compute needed for L2 plus L3, but solve the problem in 50 watts. And so this is really important with electrified cars. And we truly believe that we can move all the needle as a company and really enabling the highest performance in the industry, but also the lowest power. Yeah, and you can deliver immediate value at much lower risk. And it's a nice bridge to the future. I happen to agree. I think it's going to be a decade or even more. These are complicated situations, but I appreciate the answer. I want to come back to this notion of a shift from general purpose workloads to accelerated or intelligent computing. I want you to talk about the pace of innovation because we're trying to understand how it evolves. Like for example, I mean, you were at Xilinx, you know, the semiconductor business. How has the time to go from design to tape out changed? How is that compressing? And how does that compare to sort of traditional computing? Yeah, so a lot's happening in chips. And so I'm finally happy that we went through a winter as it comes to semiconductors. Nobody was investing anything in the 2000 and the 2010 horizon. And no doubt, something AIS per a spur of activity, as you know, more than 100 plus companies have really ventured into the chip business again. The VC community is beginning to be interested in engaging in new architectures. I would say we had reached a steady state of traditional compute servicing all the market needs and the big work getting bigger, a lot of consolidation in the M&A in the industry. And by around the 2010, 2015 horizon, Moore's loss impact of slowing really started impacting what computer architectures could deliver on a performance power or a cost equation. And so the industry was beginning to hit a wall and in full credit to Nvidia, they were one of the first pioneers to really latch on to the AI trend. And they latched on to really that being a growth vector for them as a company and the rest of history and being the first company to break a trillion dollar market cap horizon. So really kudos to them for pioneering and seeing the trend early. Be fair to say, if you don't touch AI and ML as a company in your roadmap today, you're probably gonna be a sitting duck in the next five to 10 years. So every single semiconductor company, public or private, fundamentally is weaving their computer architecture into an AI and ML form factor. That's super exciting. And along the way, what used to take three, three and a half, four years to go to production from a concept to really a tape up and into getting into active production. There's a lot of good innovation happening right now. And people even beginning to use AI and ML to really build out chips faster, verify them faster, get them into production faster. And so in reality, I think this is now a spring for the semiconductor industry in not only compute architectures moving from classic computer AI ML, but also chip development, getting a lot of innovation that it's never really gotten the attention. So both are very interesting trends to monitor. Okay, so you're saying we're going from three to four years to what? Is it 18 months or is it actually with arms latest announcement, less? I think IP companies are innovating, EDA companies are innovating and no doubts manufacturing houses like TSMC are innovating. And I would say they're actively shaving off a year away from what used to be a four and a half, five year cycle into a three, three and a half, four year cycle. Ah, okay, right. So you brought up TSM. I mean, obviously there's a number of fabs being built in onshore in the United States. The CHIPS Act is helping to accelerate that, although it appears that it's not going as great as everybody would like, the costs are significantly higher, according to TSM. How much of a concern is that? Do you think that that can be overcome and it certainly won't be overcome in the near term, but maybe in the mid to longterm? I think, I mean, so we're being public. TSM is a great, great partner for us and we continue to enjoy a very strong relationship with them. And though we are a startup company, they've gone out of their way to really give us a big company consideration and really give us the attention that we deserve and the attention that we need. So we're really very appreciative of what they're doing. Geopolitically, no doubt the world is really, I think trying to really figure out what to do with manufacturing going forward. But for the scale of who we are and the complexity of what we're doing and the attention and the partnership that we have at TSMC, those have not been concerns for us. And I think either in design consideration or in supply chain issues, we could have asked for a better partner than TSMC. Yeah, amazing company. Let's come back to sort of how you think about privacy and security and how you ensure that for your architecture and your customers. And great, great question. And so the one bigger on us if you move into the edge market like we are is every node is exposed. Every node is a point of vulnerability. So as you particularly move AI and ML algorithms and applications towards the edge, the industry benefits in that I think you could do localized privacy, localized security, but you really need to pay a lot of attention to architecture as privacy and security. And this is not an easy thing to go do. We have really a software hardware solution built around privacy and security. We do not only partner with Synopsys, that's our EDA partner from an IP perspective on privacy and security. We also partner with ARM from a ARM trust zone and how their software envelope really creates an additional software layer of security for us. We have a, sorry to interrupt, Christian, we have a graphic on this. I think bring up the second one, Alex, if you would. This is a functional diagram. Is this the relevant one? We have another one in your software architecture, but let's leave this up while you talk to this issue. That's great. And I'll talk around this and I'll also talk about our architecture. So if you look at the boot security, system management, debug tracing, so those are elements from a silicon perspective we have built up into a security and a privacy element of it. And a combination of the ARM trust zone, which is a software layer, really creates one of the most secure privacy oriented system that I think we could possibly build. And we are really, I think, very lucky that we are able to get two large companies like Synopsys and ARM that are partnered with us and being able to create this infrastructure because another day, if you're not building something with this private and secure, it doesn't matter how amazing what you do is. And so protecting the privacy and security is stable stakes for us. And so we've been very lucky that I think partnered with them and we have given the attention that it deserves. We are engaged with many aerospace and defense customers. We're also engaged with a lot of commercial customers. So they both derive the benefit of the infrastructure that we have. And the one other thing that I would point out that we have done is, Dave, you've been a student of the space for a long time. The 40 billion that's really consumed on an annual basis are system on a chip architectures or SOC architectures. One thing when I think we started the company is we saw a lot of people rushing to ML more as a capability than really anything else. And so what they've really ended up doing is build standalone ML accelerators. And this has been their response to how the industry should adopt Edge. And if you've noticed for the last five years, standalone ML accelerator companies have not taken off because it really doesn't solve the system problem. To address system on a chip classic compute, you need to bring machine learning in a system on a chip form factor so that people with risk mitigation get legacy support from day one. And they could then gradually move portions to the problem from what you see on the gray box which is ARM processor complex, the light green box which is the computer vision processing pipeline. They could gradually move those classic computer problems into the purple box, which is a proprietary ML accelerator that delivers a 50 tera ops capability at five watts. So we are very uniquely positioned as a company and we call this product an ML SOC, a machine learning system on a chip that we are able to deliver this capability in a system on a chip form factor and enable people to partition the problem and risk mitigate their face rollout into an ML world. And you have a software architecture that orchestrates this, Alex, if you bring up the third chart, I mean, we've seen what NVIDIA has done with its new architecture and its software based system, but maybe you could take us through the importance of your software component. Happy to, and I would say, I mean, I've been in the chip business now for 33 years. A successful chip company is dictated by how successful their software is. And so this is particularly true in AI and ML because software, if you really think through why I will pick on NVIDIA have done well is really because of the software. And so GPU architectures are not relatively new, but software has really been the defining element to them and company after company, public or private, we see it really beginning to do a good job in new architectures, but where they seem to fall behind or fall apart is really under software architecture. We think of ourselves more as software company building our own silicon than anything else. And so we are very software centric in our approach. What you see here on the screen is really, we are the first company to be able to take any ML framework, any resolution, any sensor, and you could keep adding the any we refer to ourselves as the Ellis Island of everything in computer vision. You could bring in your poor, you're tired, you're hungry, and once they come in, they're American. So that simplicity is what we've taken to ML models because the industry is quite nascent. Every single use case is radically different than the next one. So we really had to build a very, very flexible software front end. And we do three things well, as you see in this picture. One is we do machine learning model development. We provide libraries that go with it. We do application and pipeline development and again, libraries that go with that and then build and deploy and device manage that. And we provide all of this in a single hosted Docker container. So people could really do model development, pipeline development and device management all in a single software package. And we're only four years old, but we're already recognized to be one of the easiest to use software in the ML industry. And that'll continue to be a story we'll fortify going forward. I want to close by just sort of underscoring the market opportunity here, Alex. If you bring up the power law, this is something we've shared before, but I think it most specifically applies here. We haven't talked a ton about the edge piece. This is a power law that we developed with myself, John Furrier, Rob Streche. It's done the vertical axis is the size of the model. And we know that the Vertex AI, the Google, the open AI and Amazon and the like, Anthropics, et cetera, they're going to have these big models and the horizontal axis is the model specificity. So maybe smaller models, but very specific to industry. And this orange line is sort of a historical view of the music industry where you had a hard right angle, very few companies dominated labels, record labels dominated the music industry. You had this very long tail. We see the JAN AI and AI is different that red is getting pulled, that red torso pulled from the open source community. And then, but the real opportunity here is at the edge, that real time inferencing at the edge. We've sort of shared some examples here of SEMA AI we put in there as well, but it's Tesla, it's Apple. And I want to come back to some of the data. You said $40 billion in just in semiconductor content. And that edge piece could four to five X that, that's just the semiconductor content. Imagine the total available market up the value chain. This is enormous and it's massive volume. And this is why we think it is ultimately going to seep into the enterprise. It already is. You're certainly seeing it at AWS and all the cloud companies, but it's going to have a dramatic effect on the economics of computing in the next 20 to 30 years. Your final thoughts, Krishna. No, I fully agree and I love the chart that you showed and I wish I'd seen it before, but I fully agree. And I think that one day while the cloud will continue to grow, the edge will absolutely continue to be the market driver. And particularly from an innovation perspective in the next 20 years, you're going to see a tremendous amount of solutions built out of the edge. And the thing that's really going to define it from my perspective, from a market adoption is really going to be over ease of use and software. So like we talked about, I think ease of use is the large predictor and this market is going to be an amazing market. How amazing is going to be predicted by how easy the solutions are to build. What we also announced recently in addition to our software package that I think you talked about, Dave, is that we have announced a platform called an edgematic platform, a pallet edgematic platform. This pushes from software development and ML and we have really released the industry's first no-code ML environment. So you really don't need to have any degree in ML, you don't need to really understand the architectural details of who we are. You could really fundamentally construct complicated computer vision pipelines, drag and drop the menu, post the design, compile the design online and quickly immediately in minutes get results in sort of months. And so we want to really democratize edge AI. And to the chart that you showed now, we truly are big believers that one day the edge market is really going to be the one that outpaces the cloud and that exciting journey of a 4x, 5x multiplier in the next 10 to 15 years is the exciting journey ahead of us. Well, thank you so much, Krishna, for coming on theCUBE for SuperCloud 4. Wish you the best of luck. Thanks for the opportunity. You're very welcome. And thank you for watching SuperCloud 4. We'll be right back with more live and on-demand content from our studios and Palo Alto, John Furrier, Rob Streche and myself with more great AI conversations. Keep it right there.