NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Welcome to this webinar. My name is Leela Karumbunathan. I'm a hardware product manager here at NVIDIA. With me, I have my colleagues, Suhas Sheshadri, who's in charge of our software product management on Jetson and Chen Su, who is in charge of our product marketing. Today, we'll be discussing how you can empower your industrial AI applications with NVIDIA Jetson. So what will we be discussing today? We'll start off with an introduction to our latest industrial Jetson AGX Orin module, then we'll go into details about how you can develop industrial AI solutions with our NVIDIA software. And finally, we'll close with some advanced AI use cases powered by Jetson Orin and our Ecosystem. Let's jump into things. So first, I'm going to be talking about the industrial Orin module. Before I go into details about the module, I wanted to give you an overview of our Jetson platform. Jetson is a software-defined platform. We focus on the 3 key pillars that you can see here. The first is we provide modules as well as developer kits that are designed for sensor fusion and aggregation. Our modules come with not only a GPU, but also CPU, as well as various other accelerators. They also come with the memory and power, all combined into 2 different types of form factor, a small SoDIMM form factor as well as a larger 699 pin form factor. They come with various IOs that really enable them to be designed for sensor fusion as well as aggregation. Now not only do we provide the necessary modules that you need for your compute, we also see that a common challenge for embedded and engineers is software development. And so that's where the key to our success is really our software. We make it really easy to develop with jetson with our SDKs, design tools, libraries as well as higher level SDKs that are application-specific, along with special AI model development tools. And then the last key pillar is our ecosystem. We see one company can't serve all of your needs. And that's where we look to use our ecosystem to help you speed up your time to market. That includes ecosystem partners for hardware carrier boards as well as design services, full system providers, sensor partners like cameras as well as other sensors, software management tools, ISV and applications as well as distributors in your local region. We have over 160 plus partners and it's only continuing to grow more. I wanted to highlight the performance that comes with Jetson Orin. The Jetson AGX Orin industrial module leverages the same Orin architecture that was used for submissions to MLPerf. MLPerf is a standard industry body, which has various machine learning benchmarks that you can submit to. We're constantly optimizing our software and we're constantly improving the performance on our hardware platforms. Here, you can see in just one year's time, we've improved the performance by up to 83% as well as the power efficiency, 2 things that are very important for industrial applications. These are a variety of different benchmarks from ResNet-50 to RNN-T, 3D U-Net and RetinaNet and more. And these are just a handful of the benchmarks that can be run on Jetson. The advantage of Jetson is that any kind of model can be run with Jetson. So let's jump into the industrial side of things. What are some of the challenges we see in ruggedized environments. Industrial applications require things like shock, for example, a system might bounce around on the back of a vehicle and experience varying amounts. There are also larger requirements for vibration, for example, a tractor might drive over some bumps and experience more vibration. There are requirements for extended and extreme environment because these systems often might operate in a variety of climates from extreme cold to extreme heat. Picture a tractor on a farm on a really hot day. Operating lifetime as well as product life cycle requirements are often there as well as requirements for memory error correction. So what are some typical industrial applications? It includes a lot of what you can see here. So there's industrial manufacturing where customers need more autonomy in their applications. Agriculture applications like harvesting and construction where it's necessary to enable machinery to endure heavy workloads, navigate challenging bumpy terrains and operate continuously under varying temperatures. Farms require tractors that harvest thousands of acres of crops on various terrains and carry fertilizer and herbicide sprayers. Construction sites contain heavy equipment that must operate in various climates. Logistics and aerospace and defense applications also often require this extended shock, vibration, temperature, along with extended operating lifetime and product life cycles. For example, space applications, radiation-tolerant systems are needed as well as more ruggedization. Unmanned aerial vehicles and drones experience extreme shock impact and vibration while flying in harsh environment, and railway applications, trains generate vibrations when traveling at high speeds. The interaction between a train's wheels and the rails also lead to additional intermittent vibrations and shock. Transportation companies are using Jetson for object detection, accident prevention and optimizing maintenance costs. Oil, gas, inspection and mining applications all have similar requirements to what we just discussed. A mining truck requires additional ruggedness for open mine environment. So in summary, all of these applications have a lot of these requirements and that is why I'm happy to introduce our Jetson AGX Orin industrial module. This module is available now, and it's designed to bring server class performance to industrial applications. It brings up to 240 deep learning tops, and that's because the industrial module leverages our Orin architecture. This brings our latest generation Ampere GPU, and the industrial module comes with up to 2,048 CUDA Cores and 64 tensor cores. Designed for sensor fusion, aggregation as well as signal processing. Along with the latest Ampere GPU, you get those third-generation tensor cores that are higher performance and more power efficient. We also bring a new version of the Deep Learning accelerator compared to previous generation architectures, which is a lot more high performance and designed to kind of offload your AI workloads. Along with that, the industrial module also leverages 12 Arm A78 Arm CPU cores along with 64 gigabytes of DRAM with in-line ECC memory. The industrial module really brings server performance to the edge. Not only does the industrial module leverage the Orin architecture, it's also been designed with specific industrial features in mind. That includes features to solve some of the challenges we were just discussing. So the mechanical shock supports up to 50G operational shock as well as operational vibration of up to 5G and extended operating temperature of negative 40c to 85c on the thermal transfer plate, along with an extended operating lifetime of 10 years, and we've added some SoC Corner Bonding & Component Underfill to really make this module designed for some of those harsher environments and be able to withstand some of those extended temperatures. The great thing is the industrial AGX Orin module is 100% pin-and-form factor compatible with the AGX Orin commercial module. So you could design one carrier board to support both modules. The industrial module also supports all of the great security features that are enabled that can often be required for industrial applications, including hardware root of trust, secure boot, hardware accelerated cryptography and support for encrypted storage and memory. The industrial module will be priced at $2,149 for 1000 unit volume. And the great thing is it's really designed for industrial applications. Jumping into a little bit more details about how does the industrial AGX Orin module compares to the commercial 64-gigabyte module. Both leverage the same Orin architecture. There's -- they both contain up to 2,048 CUDA cores and 64 tensor cores, along with a 12 core Arm A78 CPU. The industrial module is designed with additional reliability, extended shock, vibration, humidity, temperature endurance as well as mechanical features compared to the commercial module that make it really optimal for industrial applications. It also comes with an extended product life cycle, which means it will be supported through 2033. Here, I'm showing you how the Industrial AGX Orin module compares to our previous generation AGX Xavier industrial module. We really increased the performance across both AI as well as CPU and DRAM to really enable your next-generation industrial application. That includes on the deep learning side of things with up to 8.2x the performance, on the CUDA side of things for your signal processing applications up to 4x the performance, the DLA, I mentioned there's that new version 2 compared to the previous generation, which is more power efficient. And that brings up to 10x the performance of what was there with AGX Xavier industrial. Also on the CPU side of things, almost 2x the performance as well as with -- on the DRAM we added LPDDR5 to enable 1.5x the bandwidth. Here, I'm showing you a more detailed comparison between our AGX Xavier Industrial module, our AGX Orin Industrial module and our AGX Orin Commercial module. You can see the Orin modules really take advantage of that Ampere GPU architecture along with the new generation deep learning accelerator, bringing that huge jump in performance and really bring server class performance to the edge. We also enhanced the Orin architecture compared to the Xavier Industrial with the latest Arm A78 CPU cores and you have 4 additional cores with 12 cores on both in commercial and industrial modules. Also, we've added 64 gigabytes of LPDDR5 at 3,200 megahertz, along with in-line ECC on the industrial module as well as 64 gigabytes of storage. The modules also support video encode and video decode, along with all of the necessary IO that you'd need for industrial sensor fusion and aggregation. That includes on the camera side of things, up to 16 lanes of MIPI CSI, along with up to 22 lanes of PCIe as well as high-speed USB. Along with that, the Orin architecture brings 10 gigabit Ethernet support along with support for 1 gigabit Ethernet via RGMII that was also supported on AGX Xavier Industrial. The AGX Orin Industrial is form-factor compatible with the AGX Xavier Industrial and it's pin compatible with the commercial module. So you could design a carrier board to support both the industrial and the commercial module. The great thing is our AGX Orin developer kit is designed to support the industrial module. So you can easily plug in and invest your module into the developer kit to jump start your development. The industrial module comes with a 15 to 75-watt power budget, and we support features that are specific and often required in industrial applications that includes sleep mode along with dynamic voltage frequency scaling. So you can tune the module to your workload. We have a power estimation tool that will also be available for the industrial module. So you can get a sense of which power mode is right for your application, depending on what workload you're running from the GPU, CPU or other IT use along with what IOs are you connecting up. So now with the industrial module, there are 7 different modules in our Jetson Orin family from our entry level to our high performance. This slide showcases, the entire Jetson family. The top of the slide highlights our Orin family and the bottom highlights our previous generation. Our previous generation, typically, we had released modules based on the GPU architecture of that time period. But now with the Orin architecture for the first time ever, we have 7 different modules from our MT level or nano up to the AGX or an industrial module that are all based on the same Orin architecture. They also leverage the same software, and so this makes it really scalable and enables you to create a solution for any kind of application in an edge AI or autonomous machine. We have our industrial module, which is form factor and pin compatible with the commercial AGX Orin series, leads to design a carrier border system to support L3 and deploy in different environments depending on what your requirements are. We also have our smaller form factor, lower power solutions like Orin NX. With these, you can design them in all kinds of applications. You can take into account some of the industry requirements at the system level. We have some documentation on how you can operate some of these at SoDIMM modules in ruggedized environments. So with this, you have a complete set of modules that you can leverage for sensor fusion and aggregation in any of the applications we discussed today. With that, I'll pass it to Suhas to discuss how you can leverage software on these modules.

Hello, everyone. I'm Suhas Sheshadri. I'm the leader software product management on the Jetson Edge AI robotics. Jetson AGX Orin in Industrial is powered by the same software stack that powers all the other digital modules. We have designed the Jetson software stack to hardware accelerate your AI applications end-to-end but also to accelerate time to market for you. Instead of designing and developing the AI applications from scratch, we provide SDK and tools that significantly reduces the time for development. And with that, you can cut down time to market. Every AI application has an AI model at the center to accelerate AI model development, NVIDIA provides 3 things. First is a large collection of pre-trained models for various use cases, which are already trained on a large data set and provides high accuracy. These pre-trained models can cut down your time to market significantly. Various pre-trained models for vision and conversational AI are provided and are hosted in NVIDIA GPU Cloud or the NGC. You can also train these pre-trained models with our own data set using transfer learning. Second, NVIDIA provides developers with synthetic data generation using NVIDIA Omniverse to create a large set of photorealistic data in a short amount of time. Developers can use the synthetic data for training model and improve the accuracy of the models. This eliminates the need for collecting a large set of real-world data, collecting which is not only time-consuming, resource-intensive, but also sometimes impossible to create in real-world settings to cover things like occlusion, different [indiscernible] conditions and such. Third, NVIDIA provides TAO toolkits, which stands for train and optimize. TAO toolkit lets you bring a pre-trained model or your own model, real-world data or synthetic data or a combination of both and to train the model and optimize the model for accuracy and latency. You can also easily perform quantization like converting to FP16 [ open date ]. TAO also added support for AutoML to train high-quality models without having to tune hundreds of parameters. AutoML automatically selects models, hyper parameters based on the model and data set. So you can see how NVIDIA provides you with the tools to create the production-ready model quickly. Next, to help you write AI pipelines quickly and more importantly, write an optimized pipeline, which is hardware accelerated to ensure best performance, NVIDIA provides application frameworks. You can think of application frameworks as LEGO blocks with which you can arrange shape into AI pipeline. These LEGO blocks are already hardware accelerated by NVIDIA and the magic happens underneath as NVIDIA optimizes the framework for 0 memory copy and distributed computing to ensure top performance. Application framework at DeepStream for writing Vision AI pipelines, Isaac for writing robotic applications and Riva for conversational AI is provided. Thousands of customers rely on these application frameworks to power their AI applications. Next, NVIDIA provides Jetpack SDK, which provides out-of-the-box development environment for developing on Jetson. It includes Ubuntu-based desktop environment and includes all the accelerated libraries from NVIDIA, like CUDA, cuDNN and TensorRT, VPI, Jetson multimedia and so on, so that the entire AI pipeline starting from capture to preprocessing sensor data to inferencing to post processing and beyond is running entirely on the hardware accelerators that are present on Jetson so that your application runs the most optimal way. This is how NVIDIA accelerates time to market for you by turbo boosting your development. We include in-depth documentation for all the tools, frameworks and SDKs and samples to learn from. We also have well-staffed and very responsive NVIDIA forums for all the components of software stack so that developers can get help when needed. Developers developing on Jetson AGX Orin Industrial will enjoy all the values that the Jetson software stack brings. The software stack on Jetson AGX Orin Industrial enables various security features in order to secure your products, which are deployed in the field. Security features include secure boot to encrypt and sign boot start kernel initrd to make sure that the module is booting off of a known and trusted software stack. For the first time, we are also bringing TPM support in firmware. The firmware TPM can be used for device authentication and storing cryptographic keys and certificates. We're also bringing support from measured boot to ensure platform integrity. Support for secure storage, support for encrypting storage and memory are also supported. We provide OP-TEE as the trusted execution environment on Jetson to ensure fully accelerated execution to protect code and data. There is also a security engine, which accelerates cryptographic operations. You can control the different security aspects using hardware fuses like configuring keys for secure boot. We also support rollback protection that you can enable so that no one can rollback your stack to a previous less secure stack. With these security features, you can completely secure your platform and also enable edge to cloud security. Let's take a quick look at the Jetpack road map for this year. We released Jetpack 5.1 and 5 1.1 in Jan and March this year. Jetpack 5.1.2 will be released in first week of August 2023, which will bring support for the Jetson AGX Orin Industrial module. It will also be the last feature release for Jetpack 5 branch. There will be no new features or compute stack updates on Jetpack 5 after Jetpack 5.1.2 release. We will transition Jetpack 5 into sustenance mode and we will release regular releases of Jetpack 5 with security fixes and critical bug fixes only. Starting Q3 2023 around October of this year. We will have our first release from Jetpack 6, which will be the new feature release branch and will include Kernel 5.15 and Ubuntu 22.04. Jetpack 6 will support all Orin modules. We will make regular Jetpack 6 releases with new features and compute stack updates.

Thanks, Suhas. That's excellent. My name is Chen Su, Senior Product Marketing Manager for Embedded AI and Robotics. Today, I'm going to cover AI use cases powered by Jetson Orin and Jetson Ecosystem. Before I jump into the novel AI use cases, it is important to understand some key challenges from development to production. As people often assert, bringing a production solution to the market requires a complete ecosystem. This rings especially true as we are in the early stages of the AI journey. One prevalent challenge faced is the scarcity of prevalidated and compatible industrial-grade cameras, sensor, connectivity modules and ruggedized system to select from. Another challenge people often face is scaling the AI model from general purpose to vertical specific and really optimize them at the best possible performance. As a result, time and the cost of migration taking the AI journey for ruggedized design and system become a huge hurdle for many smaller and medium companies to jump through. From the very beginning, NVIDIA Jetson is on a mission not only delivered the best hardware and software for embedded AI, but also enable a powerful ecosystem, including hardware, camera, sensor, connectivity and system software partners. As a team, we together deliver server class sensor fusion, AI perception and video processing in the most rugged environment. Here, we showcase over 25 system builders capable of creating industrial-grade robotics and automation hardware solution to cater to various industries, including manufacturing, agriculture, railway, construction, energy, aerospace and defense, the list go on and on. These applications often demand extended temperature, shock and vibration specifications and our partners process deep industrial knowledge and expertise to tackle these demanding tasks effectively. Industrial applications frequently require specialized camera, sensor and Wi-Fi, LTE, 5G connectivity. Our partners have years of experiences delivering high-performance solution through off-the-shelf products and custom designs tailored to each use case. Last but not least, mission-critical application often demand real-time operation system to eliminate latency and interrupted control loops. Our partner excel in real-time OS system ensuring optimal results. Let me show you the breadth and depth of the ruggedized systems powered by NVIDIA Jetson. Within the industrial and railway sectors, Jetson partners developed a diverse range of products from industrial MVRs to AI railway computers, driving the innovation in areas like railway collision protection, predictive maintenance and industrial inspections. The key features in the full system includes CAN 2.0 or RS-232/422/485 interfaces, the Power over Ethernet with regular port or M12 connectors. USB 3.1, GMSL, large DC input range and also railway and industrial certifications. In Ag tech and mining, Jetson ruggedized systems empower a wide array of applications, including sensor-based soil monitoring, site surveying, autonomous agriculture/mining vehicles, and teleoperation of heavy machinery. These systems offer exceptional AI processing power and video encode and decode capabilities. Key features you can see over there, like GMSL, auto-grade camera support, CAN 2.0 and CAN FD, Gigabit Ethernet, IP65 water protection, shock and vibration resistance and ASIL-D safety level with additional Safety MCU. For satellite and space shuttle, defense and other mission-critical applications, partners provide expertise to create solutions to meet diverse requirements in graphic processing and AI inferencing. You can see key features over there like Gigabit Ethernet, USB 3.1, HDMI, CAN 2.0, UART, GPIO, IP67, ingress protection and, of course, qualification testing for extreme ministry requirements in thermal, shock, vibration, altitude and humidity tests. Powered by hundred of those powerful systems, AI and advanced edge computing made possible almost everywhere. Now we'll go through some of those use cases to demonstrate those examples. I will start with the construction use case. Sodex Innovations creates a groundbreaking digital recording solution for construction site, simplifying the task of determining and documenting excavated materials. Unlike existing methods, which can be costly and require separate workflows, this technology can be retrofitted to any excavator, making it highly versatile. It utilizes advanced laser and camera technology to automatically identify and capture images of excavated materials, along with 38 million georeferenced measuring points. This data is then transmitted to the cloud, where it is converted into comprehensive 3D models. The benefits of this solution include accurate recording and invoicing of construction work, streamline processes for construction companies, enhanced transparency for end customer throughout the project. And of course, reduced site visits and time savings that leads to cost reduction. Jetson partner Syslogic works with Sodex to deliver the ruggedized design and extensive IOs for the solutions so that we can focus on developing software, sensor technology and edge to cloud workflow. WAYTOUS is a leading autonomous solution provider for mining industries. We developed a robust high-performance domain controllers connecting with 4 LiDARs, 4 radars and 1 camera to enhance the reliability and safety of mining trucks. The controller is equipped with dual NVIDIA Jetson AGX Xavier industrial modules for autonomous driving, allowing mining trucks to execute complex operations without human supervision. On the software side, WAYTOUS deployed YOLO model for object detection, and leveraging 3D Point Cloud for obstacle detection. Effectively, the solution improves the worker safety and reduces probability of casualty incidents in the high-risk environment. NVIDIA Jetson industrial modules deliver excellent performance, efficiency and programmability for them and the dual Xavier design ensures that we have enough AI compute power to achieve L4 autonomous mining capability. Vulcan is a next-generation weeding implement created by Farmwise. It packs cutting-edge computer vision into a lightweight open cultivator frame. Designed around NVIDIA Jetson Orin, the brain of Vulcan can process high-quality images and provide reliable weed control solutions for vegetables. While moving at the high speed, Vulcan can see each plant between 5 to 20x thanks to the high frame per second rate of cameras designed by Jetson partners famous in vision components. This leads to a higher accuracy in predictions and extreme blade actuation precision. The end result is to see farmers labor and reduce the use of herbicide. CIDI works with Jetson partners real time, created a 3D Surround View Intelligent System, that is designed to aid special engineering vehicles in navigating challenging terrains and on structure environment by providing comprehensive 360-degree visual information. It eliminates limited blind spots and utilizes sensor fusion for accurate perception of the surroundings. Powered by Jetson AGX Xavier industrial modules to process and stitch 16 camera views. This system enables drivers to see 3D view in real time and operate the vehicle safely. The end result is to significantly reduce the risk of accidents, such as rolling and collision. We have seen so many use cases deployed on earth how about sending a Jetson to the space? That's even more challenging due to extreme environment but is definitely possible. We've seen Jetson implementations by space agency like ESA and NASA and many commercial companies as well. One common feedback we've heard from those customers is that Jetson provides much lower price compared to traditional CPU solutions and reduce time to market from years to months. One example is Orbital Sidekick, who operate satellite with Hyperspectral images payloads. Hyperspectral data is critical for chemical fingerprint analytics, such as mapping, soil and many other hydrocarbons. With AGX Xavier industrial module combined with special sensor design, Orbital Sidekick can process all the sensor data in the space using primary analytic algorithms such as matched future molecular absorption models and even neuro network model, detecting problems such as hydrocarbon leaks, intrusions, new mineral exploration, hazardous waste monitoring and even precision in agriculture. Jetson GPU acceleration has positioned it as a front runner in democratizing AI technology for aerospace applications. Another great space application is wildfire detection and monitoring. As we all heard about, wildfire has become the most common natural disasters in regions like California, Australia and, lately, Canada, causing millions of dollars in infrastructure damage and ton of CO2 emissions. OroraTech is a global Intelligence-as-a-Service company, leveraging thermal intelligence for sustainable earth. Their product, Wildfire Solution, powered by high-resolution thermal data from their proprietary space-based sensor system is trusted worldwide for revolutionizing wildfire detection. Due to the compute capability, the Wildfire Solution processes the data well in orbit so that the data sent to earth is optimized, only included useful imagery instead of sending back the raw information. This provides real-time alerts and insights to keep communities safe and protect our most valuable asset. Well, we hope that today's webinar has provided you with valuable insights and knowledge on the topic at hand. As we conclude this session, we encourage you to continue exploring the possibilities and opportunities discussed today. In the end, I would like to leave a few technical resources for you. First, you can learn more about the Jetson AGX Orin industrial from the download center. We have a few technical collaterals over there, like technical data sheets, so on and so forth. Second, you can keep exploring all the Jetson offerings for the Jetson ruggedized systems in a special designed web page. We have laid out different modules and [indiscernible]

Okay. I'll try to look for questions and handle them. Okay. I'll try to look for questions and handle them. There are a few questions came in. Thanks for your patience. We'll start to address them. The first one that I see is, are there any performance metrics such as vibration, shock available for the other Orin such as Orin Nano Orin NX? And for this question, we have done many reliability testing on the other modules, as you mentioned. And you can reference the molecule data sheet on the download centers that have a reliability table. Also, many of our partners have designed system with extended shock and abrasion for the other Jetson Orin module. I would suggest you to go to NVIDIA Jetson ecosystem web page to search for those features system? And the next question is the nonindustrial version of AGX Orin includes DRAM Memory ECC? So to answer that question, the -- currently, the ECC is only available for the Jetson AGX Orin industrial module, not the commercial module. Okay. Let me see what would be the next question. The next question is, as you have talked a lot about some space application, what measures does Jetson have to deal with gamma X-rays to protect hardware? That's a really good question. Many of -- from my knowledge, many of these are like gamma X-rays protection feature are handled at the system level by our partners and also customers. They do have a special design. We have seen multiple customers deploying Jetson with enclosures to protect those -- protect their performance in the low earth orbit environments. There is a question about what about the hardware in harsh environments, for example, mining? That was a good question. I have shared two or one example during my presentation that people are actually using Jetson AGX Orin for mining application. Actually, I think I have 2 examples. One is the 3D rich construction example. And the other one is autonomous mining example. And normally, for the hardware, if you're looking for -- from the question, it seems like you're looking for a system level hardware that is capable of doing that. I would suggest -- again, I would suggest to go to the adjustment ecosystem partner page. We had a new filter system where you can actually filter by industry and you can search for mining or sometimes heavy industrial. Those are the keywords that will help you find the ecosystem partners that are capable of designing the devices that can handle a harsh environment. And the second question does Jetson support OpenCV? And the answer is, yes, we do OpenCV, but also we also have our own SDK to handle the Vision AI application. And you can -- if you want to get hardware acceleration, leveraging some accelerator in Jetson, for example, we have one kind of EDA and there's a software compute all of video processing interface. You can use the actual program. It is actually going to give you probably better performance for your vision-based applications. All right. I think that's all the questions I have seen so far. I will thanks so much for attending the call. I will pass down to Jen-Hsun to close today's presentation.

Super. Chen, thank you so much, and thanks, of course, to Leela and Suhas for their expertise as well. And thank you, everyone, for joining us today for this presentation. You can go back and you can watch a recording of it that will be available about an hour from now using the same link that you used to join us today. and you'll also get an e-mail in the next day that will have a copy of the presentation for download. So thank you all again for joining, and have a great day.