Name: (EE7B) Chatbots to Robots
Uploaded: 2025-12-11T21:14:06.768Z
Duration: 23 min 8 s
Description: (EE7B) Chatbots to Robots

Transcript for "(EE7B) Chatbots to Robots": Hi. My name is Drew Barbier, and I'm vice president of the IP business unit here at MIPS. Today, we're gonna talk about from chatbots to robots, a physical AI story. Here at MIPS, we really believe that AI is moving from the cloud to the edge in a really big way. And I'm gonna talk about that in just a second. But first, I wanna talk about how we got to to where we are today with the current AI technology. And I really like to break this up into a a couple technology. And I really like to break this up into a a couple key different AI technologies that, have advanced pretty rapidly over the past, past decade or so. So first, we start with computer vision. I think this is where the the AI, sort of modern journey of AI started. And computer vision really came from convolutional neural networks, a technology that sort of mimics how humans do pattern recognition in our brain. Right? At at the end of the day, humans are very good pattern recognition machines. And we see computer vision deployed, really everywhere today, be it autonomous vehicles, looking at agricultural applications to to identify, crop health and manufacturings, really everywhere. Sort of the next big breakthrough in AI technology came from generative AI. This is where, via prompts, AI is generating digital items like text documents, audio files, video files, and so on. So I don't know if if you're anything like me, but have been having a lot of fun over the past couple of years looking at a lot of these AI generated content, on the Internet. So it's still very compute intensive, and and it's usually not local to your device. Right? It's still all happening in a data center somewhere, but all in the the digital domain. Sort of the big next technological evolution in AI came from reasoning. This is where we are using transformer based models in GPT and Claude, and it goes from, really generating digital content to autonomously acting on that content, understanding, assessing, and and acting. So I don't know if you remember the original digital assistants. If you go back to, Copilot and and Siri and the like a couple years ago and compare that to to where we are now, it it's really, a very stark difference. And, really, if you're not leveraging this kind of AI technology in your daily workflows, you're you're probably behind. So where is it going? So so physical AI, is where we think the next big, evolution of AI comes from. This is where we're taking all of the the benefits, all of the technology that we've developed to this point and taking it out of the digital domain and putting it into the physical domain. This is what we call physical AI. It's where AI leaves the data center and starts interacting with the real world at scale. And we believe, this is gonna drive, the next generation of compute. Because if you think about it, all the the compute, the technology, and the infrastructure, to that we interact with on a day to day basis, It's really massive, and it's all gonna change as this physical AI wave, comes to play out. So so if you're familiar with me, I'm I'm a big believer in in standards based technology. I I I said quite often that, the only way to accelerate adoption is is through standards. And at MIPS, we really take that to heart. So one thing we've done is is adopted risk five and and open and modular ISA. All of our processors are built on, risk five technology. GlobalFoundries, it's right in the name. We take foundries at a at a global scale, but also, in such a way that is accessible, and also power efficient and able to reach, the mass markets and and all the technology that we touch on a day to day basis. And then at MIPS, we have the MIPS Atlas processor IP portfolio that's built on top of these technologies as well as the MIPS, Atlas software portfolio. And I do think the software portfolio is a very important piece of this, because we do at MIPS take a very software first approach, and we also engage our customers in a very software, first approach. So leveraging these standard based technologies is really providing sort of that foundational technology stack for this next generation of physical AI. At MIPS, we're we're also very customer driven. So we, you know, don't develop technology just to develop technology. We engage our customers to understand their problems, their pain points, and try to understand where we fit in and how we can optimize our solutions to to meet those, those customer driven problems. We do this by engaging with leading technology companies all around the world and engaging on this software first approach that that I mentioned, earlier. So we're not developing technology just to develop it. We're not, going after markets that, we don't think we can address. We're going after big problems in the physical AI space, and we're doing this in partnership with our customers. And I mentioned the software first engagement. This is true both from, a MIPS standpoint as well as from our customer standpoint. So we have some interesting and differentiated technology inside of MIPS, that enables us to, take in customer workloads, understand how they how they work, where the, where the bottlenecks are really in the full end to end solution space and where we can go provide targeted optimizations for those solutions in a very deliberate manner to not just get, incremental benefits over, competing solutions, but to really get substantial, benefits, order of magnitude benefits over our customers over our our competition. Sorry. So as we take that forward, this technology, we also make it available to our customers. And what this really enables them to do is to, shift left their engagement, and their development cycles as well. Right? So it's putting MIPS technology in their hands as soon as possible to enable them to evaluate our solutions and understand how our solutions solve their problems. And then beyond that, enabling them to start developing on that technology as soon as possible. Right? You hear it all the time. Software hardware co design, shift left. With our Atlas software portfolio, we are really enabling our customers to do that. And then in addition to that, enabling them to go enable their customers earlier in the design cycle as well. So when you take a look at the compute challenges of physical AI, we like to break it down or we see it broken down into a couple different buckets. Since, think, act, communicate. So I'm a processor guy. So the first time I saw this, my brain immediately goes to a processor pipeline, where you have fetch, decode, execute, write back. That that to me is what this physical AI compute stack is. So sensing is when you're taking in information from the environment. Right? This could be radars. This could be cameras, lidars, sensors, what have you. Making sense of that data and then passing it on to think. Think is when you're really holistically understanding this data, making decisions, and driving the next stage, which is act. Act is where, that interaction with the real world happens. That's where physical AI really comes to to meet the the the digital meets the analog world, if you will. That's where motors spin. That's where brakes get pushed, and so on and so forth. And then communicate is really all about moving this data in intelligent ways, in efficient ways, because in the physical AI world, data is everywhere. You have the sensor data that we talked about coming in. These could be massive amounts of data. You also have, smaller data, and that might be just as mission critical and as important as some of the larger data. So understanding that and moving this data around in an intelligent and efficient way is extremely crucial for this entire physical AI, stack. Now where where do we see physical AI, in the market? In reality, it's everywhere. Every every application that that you interact with on a day to day basis, eventually can be broken down into a physical AI type application. Now at MIPS, we can't draw, we can't, address every problem as much as we want to, so we tend to focus on a couple of different market segments, that MIPS has a unique expertise and has traditionally played a a very strong role in. That's industrial, that's automotive, that's aerospace, and and enterprise infrastructure. We continue to do that today, and we we think, we we know that physical AI is really gonna be driving these markets forward and driving requirements back into these markets over the next decade. So if we take those, application markets and the SenseThink at communicate processing stack and map that onto the the MIPS products. So it's since we have the p 8,700, this is our first RISC five processor. We announced, general availability of this product, last year. It it's in silicon today. We have a lead customer in in Mobileye who is one of the the tier ones in the automotive, ADAS space, shipping hundreds of millions of units over the past, ten, fifteen years, all with MIPS IP, now doing so on MIPS risk five IP. Think is all about, AI acceleration, AI everywhere. Right? This is where you're synthesizing the data that's coming into the the system, and then making sense of that data and and coming up with actions, driving actions into the rest of the pipeline into how you're gonna address that. So in the case of robots, as an example, understanding, what you're gonna do next with the robot, where you're gonna go, where you're gonna move your your appendages, where you're gonna say, what are you gonna do. Next is ACT m 8,500. M 8,500, we announced at Embedded World this year. It's all about high performance, real time control at the edge. Motor control being a a big function of m a 500, also functional safety in in the automotive space. All these play a part in in the m 8,500 and this physical AI compute challenge in front of us. And then communicate the I 8,500. We announced general availability of this product a month ago. The 8,500 is is very good at processing, efficiently processing and moving data throughout the systems, for communication, and the communication problems that that we see in front of us, for physical AI and helping make sense of moving that data around in a very intelligent and a very efficient way. MEPS Atlas Explorer, is is the productization of our, modeling technology that we've been been developing in house for for some time. This is a virtual platform that enables customers to interact with our IP on a very, very low level to gain an understanding of how their workloads, how their applications map onto MIPS technology, and how they can better optimize, their software, their workloads on top of that. So it gives you very detailed insight into the microarchitecture of our products so that you know when a cache miss happens, you can tie that back to the exact line of code where, that that triggered that cache miss. When your TLB misses and it triggers a TLB refill, why that happened? It gives you, IPC metrics and so on and so forth and enables you to compare different runs from, one software optimization path to another to understand and compare and contrast how that works at a very detailed microarchitecture on the MIPS IP. But in addition to that, with Atlas Explorer, we're gonna be taking this beyond just MIPS processor IP. We're really looking at system level, simulations, digital twins, being able to enable our customers to understand not just how their code runs on a MIPS CPU, but how how their code runs in their SoC and how SoC level choices, affect different optimization points in their software flow. And then even after design so, you know, Atlas Explorer is all about shifting left, about, enabling software as early in that design process as possible. But Atlas Explorer is also just as applicable after the fact. When code is in production, and you push changes to your your software stack, you're able to regress, your workloads on Atlas Explorer and really understand, the impacts that a given patch has on the performance, in the system level context. So Atlas Explorer, we're very excited about. You'll be hearing a lot more about Atlas Explorer from MIPS as we move forward. It's a very key and crucial part of our solution space in the for physical AI. Okay. So bringing it all together, chatbots to robots. When we start looking at physical AI, I think robots, is a extremely good example that sort of ties everything together from that entire sense, think, act, communicate paradigm, bringing it all together in a platform, that we can make sense of. So big fan of this. So now, every time the team is bringing forward new ideas, new new products that they wanna take to market on the IP side, my first question to them is where does it go in the robot and how does it make it better? Right? In the robot, there are so many compute challenges, be it from the the control loops, to from the movement of the robots to sensor fusion data, right, where you have cameras and microphones and and different sensors around the robot to com distribute a compute where you're making decisions based on this data, all over the robot, not just in sort of the the brain, if you will, but in all different parts of the robot, like the hands, the appendages, and so on and so forth. And then finally, AI throughout. I I think you're gonna see AI distributed throughout the robot, be it vision processing, audio processing, to how we do different bits of control throughout the the the robot from moving the legs to to the hands and and so on and so forth. So robots, I I think, are a very, very interesting problem and sort of bring to life, this physical AI workspace that MIPS is really doubling down on and and looking to for the the next wave of of compute challenges in our space. So let's take a look at a couple of the the the different, spaces inside the robot. So m 8,500, is really good, on on the act side. Remember, 8,500 is all about hard real time control, all about functional safety. In a robot, there's there's lots of moving pieces. M 8,500 is the perfect, IP to to perform that function. And remember, it it's not just about acting fast or or real time. It's about acting with fidelity. It's all about acting with nuance. Right? You don't want jerky motions. You don't want a a robot holding a hand and and squeezing too tight. You want that feedback loop throughout, and you want that to happen in a very efficient process, in a very fast process. And mini 500 is is, very well positioned to to play that role in the robot going forward. When you look at communication, there's communication all throughout the robot. You have sensors, bringing in all kinds of data. That data needs to flow to the appropriate parts of the robot, and it needs to do so efficiently. I 8,500, is a four way multithreaded processor that really enables you to do that communication processing and to do so in an intelligent fashion, in a real time fashion where you're able to respond to packets as they come in, you're able to to compute, quality of service metrics, understand if this is a very high priority piece of data that needs to go to a certain place, or if there's other higher priority packets that that can slip ahead of it. I e 500 is a a very good applicate a very good processor for these applications. And in smart sensing, this is really, where everything kinda comes together and the robot starts acting autonomously, right, bringing in, the ability to understand its environment and react to its environment. Right, if you have a robot and sitting there in your living room and it hears the tea kettle start whistling, to one, recognize that that noise is in fact the tea kettle, to be able to tell the robot to to stand up and walk towards the stove and and take the tea kettle off the stove, to be able to pour that tea kettle into the cup, with very fine motor control. That all happens with the the smart sensing capabilities of the robot and really across a number of different products, including the s 8,200, which is where those autonomous decisions are really gonna be calculated and performed and and then driven to the rest of the robot to to perform those actions in the the act function of the the robot. So with that, physical AI is built on MIPS. I I talked about how we believe that physical AI is really the the next driving force between compute IP driving force for compute IP and our markets going forward. It's gonna drive things at a scale that that you can imagine. And to do that, we need to, embrace this software first design approach, leveraging standard based technologies, and then really look at targeting our IPs to solve specific problems in this space. Right? As I mentioned, robots is a very big focus, Where and how our solutions solve problems on the robot is the the question that I drive into my team that we're gonna look to be answering over the next ten years. And MIPS has done this before. Right? MIPS has been around for forty years. We're proven in the market. We have hundreds of millions of designs that are shipped. We have, we've done so in demanding environments like automotive, and we're doing this now in partnership with GlobalFoundries, that are extremely focused on delivering these types of solutions at scale, very efficiently, and doing so in a way that makes sense for our customers in a global footprint. With that, I wanna thank you, and I wanna thank EE Times for having MIPS be a part of the AI everywhere virtual conference. Thank you very much.