[Music] hello everyone and thanks for joining my talk i'm jerry chen i lead global business development for nvidia's manufacturing and industrial business as you're all fully aware we're now well into an industry transformation driven by ai now it might seem like ai is inescapable in our daily life nowadays with smart devices and internet services than everything we do but we're really only at the beginning of the age of ai economists call ai a general purpose technology but what does that mean simply put a general purpose technology isn't an innovation that serves just one singular purpose a general purpose technology impacts all existing industries what are some examples of technologies that economists consider general purpose technologies well i think there are six big ones domesticated agriculture transformed us from hunter gatherers to allow us to build the first great city-states twelve thousand years ago written language allowed us to communicate over time and space five thousand years ago the internal combustion engine allowed us to exceed the power of animal labor late in the 19th century electricity unshackled workers and equipment from the drive shaft early in the 20th century information technology enabled automation in the last quarter of the 20th century and now ai is able to learn continuously from data enabling superhuman capabilities it's no coincidence that each of these general purpose technologies were sparks that ignited revolutions but noticed that agriculture and written language took thousands of years to transform civilization and yet in the last 100 years civilization transformed not once but four times four industrial revolutions the first industrial revolution was ignited by the internal combustion engine a true general-purpose technology that transformed every economic activity of civilization 50 years later the second industrial revolution was ignited by the deployment of electricity at an industrial scale the third industrial revolution was ignited by the broad adoption of information technology meaning commodity computing systems and as you all know this third revolution the i.t revolution has been the economic engine for taiwan's modernization we've now taken our first steps into the fourth industrial revolution ignited by ai artificial intelligence this might be the most impactful general purpose technology of them all because it has the potential to learn improve and to operate autonomously which means it will get exponentially more powerful over time to put this all into perspective our ai journey started with the big bang of ai when three innovations came together a decade ago machine learning algorithms lots of data and gpu computing that was the dawn of ai computing the first production deployments of ai started with the big hyperscalers like google microsoft baidu and amazon these cloud service providers were the first to build cloud-based ai services into their platforms and as you all know many of these ai services consume a lot of gpus and these cloud investments continue to grow with no signs of slowing down now the next big wave of our ai journey is represented by ai breaking out of the cloud and extending into industrial assets this is what we mean when we say the ai of things or aiot this new industrial wave has very different challenges compared to ai services deployed in managed cloud data centers the industrial edge is represented largely by remote unmanaged industrial assets in factories oil fields or rail networks connectivity is largely non-existent today but emerging technologies like 5g are an enabler to deploy and to manage ai at the industrial edge 5g is also a conduit to collect sensor data to teach edge ai systems how to continuously learn and to improve but not all industrial edge systems have the luxury of relying on connectivity some mission critical applications need intelligent agents to perceive their environments and to operate autonomously we think of these autonomous systems as robots but in reality they can be complex cyber physical systems that continuously learn adapt and improve how they operate unlike the first wave of cloud-based ai systems these next two waves of ai operate in the wild outside of managed data centers but this ai infused industrial edge or ai of things represents enormous sectors of our global economy what sorts of industrial applications will be affected well in short all of them engineers run huge hpc simulations to optimize product design and manufacturing processes ai is becoming infused into surrogate models for design exploration and for inverse design to guide the discovery of new catalysts and industrial materials there's wide adoption for ai for image-based quality inspection in precision manufacturing notably for semi-conductor defect detection and classification and many of you are also deploying ai for predictive maintenance and service operations and to support worker safety and productivity assistance ai is also used increasingly to optimize supply chain and logistics operations we're seeing a lot of collaborative robots and autonomous ground vehicles from material handling finally all of these manufacturing and logistics operations continuously produce enormous amounts of data industrial applications are complex operations that require continuous and near real-time optimization by necessity any intelligent support system needs to fit naturally into the workflow of the human operators the best ai systems are helping workers operate sophisticated machinery through natural language and conversational technologies these are often coupled with augmented vision tools like ar or vr to allow seamless and natural interactions to assist workers simply put every industrial sector will be transformed by ai the potential for value creation is enormous and the spectrum is staggering from the most modern highly sophisticated sectors like semiconductor design and manufacturing all the way to our civilization's first great general purpose technology in agriculture tremendous potential for trillions of dollars of value creation all through ai let's take a quick look at a few examples [Music] my name is misty nice to meet you so what are the unique challenges for deploying ai for industrial applications well i'd argue there are eight major challenges the first of these is connectivity data is a critical ingredient for training ai agents and without robust high bandwidth connectivity curating that data is impossible that's the first challenge now the second is scalability with enough data and enough talented data scientists it's possible that all ai tasks could be architected and trained from scratch but from a practical perspective that's just not a scalable way for an organization to train their ai models that's the second challenge now of course you'll need to efficiently provision infrastructure for your data scientists to develop and prototype ai models provisioning with maximum availability and maximum infrastructure utilization is the third great challenge and then once you've trained your ai models the fourth challenge is to deploy and manage those models in the field now this may seem like just simple plumbing but imagine supporting sophisticated ai in industrial equipment deployed in remote industrial environments and in addition all this connectivity increases security risk by exposing more cyber attack surfaces simply guarding the boundaries of your infrastructure is woefully insufficient and nobody in taiwan needs to be reminded about well-known cyber security intrusions that cost millions of dollars in lost productivity and remediation even more frightening is the prospect of lost ip due to cyber security intrusions this is the fifth challenge now of course protecting your infrastructure and your ip is secondary to protecting your own people who are your most valuable resource this is especially true during periods of pandemic but also true in dangerous industrial environments where safety equipment and procedures need to be monitored to keep workers safe this is the sixth challenge the seventh challenge is the difficulty of training autonomous ai agents that are sophisticated enough to be adaptive and resilient in messy industrial environments and then finally the eighth challenge is how to design and optimize factory operations in realistic real world physical spaces to embark on a digital transformation journey i'm proud to say that nvidia is investing in solutions to address all of these challenges let's show you a few of them in action high volume manufacturing is challenging even for mainstream consumer products aluminum beverage cans seem pretty low tech but at 60 cans per second quality inspection is a highly challenging problem our partner data monsters created a solution to detect defects on the fly inspecting 100 of every can produced and filled this can save millions of dollars compared to the alternative which is human inspectors looking at just a small sample of cans at 60 cans a second the difference between catching a problem through manual sampling 45 minutes after the problem actually happened versus catching it right away with ai could mean hundreds of thousands of defective product that needs to be thrown away in this video look at the images at the bottom showing an image of the defective can next to a heat map highlighting the defect modern car manufacturing is highly automated for the most part but process errors do occur and if they're not caught right away the car will continue through subsequent manufacturing and assembly steps only to require expensive rework and disruption far down the line our partner v7 labs has developed a platform that enables manufacturers to train inspection systems to recognize defects that are unique to each manufacturing process and to enable their customers to retrain their ai models using their own unique data and of course nvidia uses these ai augmented inspection technologies for our own manufacturing operations here you see images of our pcb assembly lines looking for service mounting failures running on our low-power jets and xavier devices there are plenty of standard aoi systems for pcb assembly inspection but setting up a new production run is a tedious process since every pcb assembly has unique components and configurations traditional inspection systems also suffer from painfully high and expensive false positive rates our ai system learns from every pcb assembly it sees continuously building a catalog of components with every new production run our ai agent also learns more about what defects look like it also automatically learns more and more about how to use different camera perspectives and to adapt to different lighting conditions as i mentioned earlier cyber security has become an existential threat especially for high-tech companies with valuable ip yet modern manufacturing is trending towards integration of even more complex and connected machinery every new manufacturing tool with computing nodes and network connectivity is a potential attack vector what's needed is a cyber security solution that's embedded in the very fabric of the communication network that's exactly the sort of cyber security solution that nvidia's morpheus cyber security application framework is designed to enable working in conjunction with our dpu network devices embedded in the network fabric it starts with a network here we see a representation of a network where dots or servers and lines the edges are packets flowing between those servers except in this network morpheus is deployed this enables ai inferencing across your entire network including east-west traffic the particular model being used here has been trained to identify sensitive information aws credentials github credentials private keys passwords if observed in the packet these would appear as red lines and we don't see any of that uh-oh what happened an updated configuration was deployed to a critical business app on this server this update accidentally removed encryption and now everything that communicates with that app sends and receives sensitive credentials in the clear this can quickly impact additional servers this translates to continuing exposure on the network the ai model in morpheus is searching through every packet for any of these credentials continually flagging when it encounters such data and rather than using pattern matching this is done with a deep neural network trained to generalize and identify patterns beyond static rule sets notice all of the individual lines it's easy to see how quickly a human could be overwhelmed by the vast amount of data coming in scrolling through the raw data gives a sense of the massive scale and complexity that is involved with morpheus we immediately see the lines that represent leaked sensitive information by hovering over one of those red lines we show complete info about the credential making it easy to triage and remediate but what happens when this remediation is necessary morpheus enables cyber applications to integrate and collect information for automated incident management and action prioritization originating servers destination servers actual exposed credentials and even the raw data is available this speeds recovery and informs which keys were compromised and need to be rotated with morpheus the chaos becomes manageable now let's take a look at simnet a framework for running physics-informed neural networks on gpus here it's used to solve partial differential equations to simulate a coupled fluid flow and heat transport problem to optimize the design of a heat sink on a motherboard the training takes four days on a single dgx system and once the ai model is trained we can evaluate and explore 25 000 different configurations in a matter of seconds this is a groundbreaking example of how ai can be infused with the understanding of coupled physics so that it can be used as a surrogate model to explore the design space to find the optimal configuration we use tools like this internally at nvidia to do design space optimization for our own products here's a demonstration of something i'm really excited about bmw group's smart transport robot and sortbot uses nvidia's isaac platform to train the robots in a simulated environment for motion planning object detection and pose estimation the robots are training in a virtual environment using isaac's sim running on omniverse to provide a simulated training environment with physical rules and realistic visuals this approach called sim to real is designed to train an autonomous ai agent to safely learn in a synthetic environment then the trained ai agent is taken out of the virtual simulation and inserted into a physical robot in the real world that real world environment is practically indistinguishable from the synthetic world where the ai agent was trained the results are remarkable this approach promises to transform how industry trains and deploys robots in all sorts of industrial environments now i've said that these advanced robots all need to be trained in a synthetic environment but modeling a real complex modern manufacturing plant to the level of detail necessary is a daunting task furthermore even a physically accurate model would need to look almost identical to how the robot would see the world with his entire array of different sensors but that's exactly what the bmw group achieved working with nvidia using nvidia's omniverse platform can you tell how far i have to been down there so we'll get you a taller one but there's yeah it's perfect bmw used omniverse to integrate a range of planning data and applications to allow real-time collaboration with open compatibility this achievement represents a huge leap forward in the field of digital virtual planning bmw and nvidia envision a future where a virtual representation of an entire manufacturing production network enables an integrated approach to planning complex manufacturing processes omniverse greatly enhances the precision speed and efficiency of this planning process we look forward to the time when production planners at bmw group will be able to visualize the entire planning life cycle for every plant in their global production network accelerated by scalable gpu infrastructure omniverse will also help enable a wide range of ai capable use cases from autonomous robots to predictive maintenance and data analysis a true digital twin of a factory now i've just shown you many fantastic ai powered technologies all of them show our commitment to creating new markets but none of it would be available unless our technologies are easily available in the market that's why nvidia's business model requires relentlessly broad ecosystem support we're available on every major computer platform in the cloud in data centers and on edge devices we support power arm and of course x86 intel and amd platforms our business model also requires relentless ecosystem building investing resources with all of our partners so we can lift all boats by enabling and accelerating all of our ecosystem partners i want to thank each and every one of you for your support and for your commitment to joining us on our journey into the age of ai we can't wait to transform the future of the industrial edge the industrial ai of things with all of you and i can't wait to see what we can build together thank you