Minsheng Securities: Edge AI is an industry trend. It is recommended to pay attention to these targets.

This week’s opinion

1.1 NVIDIA: Embodied Intelligence Path with Omniverse as the Core

NVIDIA chooses the path of embodied intelligence, which is characterized by development from the edge to the cloud. Its main products are VIMA robotic arms and Isaac AMR mobile robot platforms. NVIDIA released a robotic arm based on the VIMA large model at ITF2023. According to NVIDIA, VIMA can understand, reason and interact with the physical world, such as moving and arranging objects according to visual and text prompts. VIMA can also run in Omniverse to simulate physics and interact with the physical world. Make predictions consistent with the laws of physics. Isaac AMR is a platform for simulating, validating, deploying, optimizing, and managing fleets of autonomous mobile robots. It includes edge-to-cloud software services, computing, and a set of reference sensors and robot hardware to accelerate large-scale deployments by connecting to DeepMap's cloud services. The mapping and semantic understanding of the environment reduces the time it takes for robots to map large facilities from weeks to days and achieves centimeter-level accuracy without the need for a senior technical team. By generating rich 3D voxel maps, it can create occupancy maps and semantic maps for many types of adaptive mesh refinement (AMR).

Embodied intelligence refers to training algorithms by imitating the way the human brain works, thereby achieving self-understanding, self-optimization, and ultimately learning and growing in a human-like manner. NVIDIA VIMA supports multi-modal modes such as text, vision, and voice as the robot's task input, and uses systematic generalization tests including simulation benchmark tests, more than 600,000 expert trajectories, and multiple-level evaluation protocols as the target output set; using pre- The trained T5 model encodes multimodal cues and modulates the robot controller on the cues through a cross-attention layer, predicting motor commands conditioned on cues and interaction history as a set of predicted outputs; predicting outputs and targets through the model Output comparison and self-optimization drive VIMA autonomous learning. In the most difficult zero-sample generalization training task, VIMA's task success rate is up to 2.9 times higher than the previous optimal method, while using 10 times less training data, VIMA The performance is still 2.7 times better than the top competing methods.

With real-world data sets, embodied intelligence has core competitiveness, and NVIDIA Omniverse is the industry benchmark. To conduct embodied intelligence training, special data sets are required. For example, for the VIMA robotic arm, the following data is required:

1) Multi-modal task set: 17 meta-tasks, each meta-task can be split into 1000 separate tasks, including various multi-modal tasks such as semantic understanding and visual segmentation

2) Success Stories: NVIDIA has prepared 650K success trajectories

3) Reward benchmark: Establish VIMA-Bench to use probability theory methods to reasonably evaluate task AI prediction results and give AI correct feedback

NVIDIA Omnivers is the world's leading digital twin platform, consisting of five important components: Nucleus, Connect, Kit, Simulation and RTX renderer. These components as well as interoperable third-party digital content creation (DCC) tools and renderers , plus extensions, applications and microservices built by third parties and NVIDIA form a complete Omniverse ecosystem. Through real-world data sets such as USD, digital twins can accurately comply with physical laws, accurately respond to object movements, and be synchronized with displays. These real-world data are an important foundation for NVIDIA's embodied intelligence to be implemented.

NVIDIA's embodied intelligence path for edge AI's bottom-up breakthrough has been opened up. Combined with cloud computing, it can provide rich technical support in all aspects of edge AI from development to deployment. NVIDIA selected Microsoft Azure as the first cloud service provider for Omniverse Cloud. Microsoft Azure will enable enterprises to access the full set of Omniverse software applications and NVIDIA OVX™ computing systems while enjoying the scale analysis features and security of Azure cloud services, providing customers with a set of tools that can be used to design, develop, deploy and manage industrial elements. Universe Application’s full-stack cloud environment and platform functions connect and use related products from NVIDIA’s partner ecosystem, such as NVIDIA certified workstations, edge computing modules, etc.

NVIDIA Isaac is an acceleration platform specially built for robot development and AI, and the AMR mobile robot platform was the first to be launched. The NVIDIA Isaac platform starts with a pre-trained model, augmented with synthetic data generated in Isaac Replicator, and trained with NVIDIA TAO to achieve target performance; leverage NVIDIA Isaac Sim available on-premises and in the cloud to create physically accurate and realistic environments to develop and test all aspects related to running your robot; save time with Nova Ori and bring AI to NVIDIA Jetson-based robots using hardware acceleration SDKs, such as Isaac ROS GEM for ROS-based robots, Isaac ROS GEM for video stream parsing NVIDIA DeepStream SDK, NVIDIA Riva for natural language processing; manage robot fleets to optimize productivity with EGX Fleet Command and Isaac for AMR including Metropolis, CuOpt and DeepMap. Isaac AMR is a platform for simulating, validating, deploying, optimizing and managing fleets of autonomous mobile robots. It includes edge-to-cloud digital twin training, software services, compute and a set of reference sensors and robotic hardware to accelerate the development and deployment of AMRs. Deployment speed, reducing costs and time to market.

Isaac AMR is built on the NVIDIA Nova Orin reference architecture. Nova Orin integrates multiple sensors and system modules including stereo cameras, fisheye cameras, 2D and 3D lidar, supports advanced AI and hardware acceleration algorithms, and provides 275TOPS of real-time edge computing performance. Synchronized and calibrated sensor suites provide sensor diversity and redundancy for real-time 3D sensing and mapping. Cloud-native tools for recording, uploading and playback, making debugging, map creation, training and analysis easier.

1.2 Qualcomm: Hybrid AI path integrating cloud and edge

Qualcomm announced the launch of a cloud-edge hybrid AI solution, committed to in-depth development of edge AI, while also cooperating with Microsoft and other companies to develop cloud AI. According to the Qualcomm hybrid AI white paper, in hybrid AI scenarios, the edge large model is the sensing organ of the cloud large model. For example, the user speaks to the mobile phone, and the automatic speech recognition (ASR) AI model such as Whisper converts the speech into text on the device and sends it to the cloud. , run large models in the cloud and send back text answers. In the advanced version, device AI further protects privacy, takes on more processing, and provides more personalized prompts to the cloud: through device learning and personal data, the device creates a user profile, collaborates with the scheduler, and provides better context-based Prompts; for example, the user asks for a mobile phone reservation to eat with friends at a favorite restaurant. For simple queries, a smaller large model can be run on the device without cloud interaction. If the user needs complex information, the demand is converted into prompts locally and sent to Large model in the cloud and return detailed answers.

According to Qualcomm Hybrid AI white paper, hybrid AI mainly has the following advantages:

1) Economy: Reduce cloud inference costs, efficiently utilize edge computing power, and lower the threshold for AI application development;

2) Low energy consumption: Edge devices can run large models with lower energy consumption. If factors such as processing and data transmission are taken into account, the energy consumption savings will be more significant;

3) Reliability: Edge AI is not affected by network conditions and runs more stably;

4) Privacy: The data is completely inferred locally, greatly reducing the risk of leaks;

5) Personalization: Edge devices can collect users’ real-life behaviors, values, pain points, needs, concerns and other information to form customized services.

Software side: Qualcomm AI development stack has been released. Qualcomm AI development stack supports mainstream AI frameworks such as TensorFlow, PyTorch, ONNX, and Keras, as well as runtimes such as TensorFlow Lite, TensorFlow Lite Micro, and ONNX runtime. In addition, it includes inference software development kits (SDKs) such as the popular Qualcomm® Neural Processing SDK for Android, Linux and Windows. Developer libraries and services support the latest programming languages, virtual platforms, and compilers. At a lower level, Qualcomm AI development stack system software includes a basic real-time operating system (RTOS), system interfaces and drivers. Qualcomm AI development stack provides rich and diverse operating system support on different product lines, including Android, Windows, Linux and QNX, as well as deployment and monitoring infrastructure such as Prometheus, Kubernetes and Docker. Qualcomm AI development stack also includes AI Studio, which supports everything from A complete large model workflow from design to optimization, deployment and analysis, integrating all tools into a graphical user interface and providing visualization tools to simplify the developer experience and enable them to view model development in real time, including AI model efficiency tools Package (AIMET), AIMET Model Repository, Model Analyzer and Neural Architecture Search (NAS).

Hardware side: Qualcomm Hexagon Processor core. Qualcomm Hexagon Processor adopts a new architecture and a dedicated power supply system. It adds special hardware to improve group convolution and activation function acceleration in AI inference, and doubles the performance of tensor accelerator. It uses a unique method to convert complex The AI ​​model is decomposed into Micro Tiles to accelerate the inference process. Scalar, vector and tensor accelerators can work simultaneously without involving memory each time, thus saving power consumption and time. In addition, seamless multi-IP communication with Hexagon is achieved through physical bridges. This connection can be used to drive those high-bandwidth, low-latency applications, such as cognitive ISPs or boosting low resolutions in gaming scenarios. Qualcomm Hexagon Processor successfully converted multiple deep learning models from FP32 to INT4, reducing energy consumption by up to 60% while increasing performance by 90%.

Qualcomm has already used the stable diffusion model in practice, and its future plans for large-scale artificial intelligence edge deployment are very clear. In February 2023, Qualcomm used the Qualcomm AI Stack to perform full-stack AI optimization and deployed Stable Diffusion on Android smartphones for the first time. Qualcomm released a hybrid artificial intelligence white paper in May 2023, which predicted that by the end of 2023, edge artificial intelligence technology will cover various models with parameters below 1 billion.

1.3 Chuangda: Integrated development of software and hardware for large models and large platforms

On May 18, 2023, Thunderbolt released the Rubik large model, which is the prototype of the first unified operating system for edge AI terminals in China. At the same time, Thunderstar also jointly established an artificial intelligence joint innovation laboratory with Amazon Cloud Technology , and displayed a series of Chuangda Rubik's Cube products such as TurboX modules, smart speaker reference designs, Rubik GeniusCanvas, etc. According to the company's 2022 annual report, since its establishment, the company has been accumulating and accumulating terminal-side, edge-side, and cloud-side technologies, and has now become the world's leading technology manufacturer in the above fields. In addition, the company's robot products have covered almost all currently existing robot scenarios and have been favored by many robot manufacturers around the world. In addition, the company maintains in-depth cooperation with leading companies in the industry chain's technology and products, building an ecological advantage. The company has large models on the software side of edge AI, and on the hardware side, it is deeply involved in the ecology of technology giants such as Qualcomm and Amazon. The development path of integrated software and hardware has great potential.

1.3.1 Horizontal view: Chuangda Rubik builds AI ecosystem around existing business

The core product of the Rubik large model series is the Rubik Language large model. The chairman predicts that it will reach the ChatGPT3.5 level in 2024. Rubik Edge, Rubik Multi-Modal, and the Rubit Robot expected to be launched in 2027 will all serve the fields of smartphones and smart driving, improving the human-computer interaction experience. At the same time, the Rubik series builds an ecosystem around human-computer interaction and existing businesses: RubikStudio, RubikAuto, RubikDevice and Rubik Enterprise. While having a powerful large model, it will also turn the large model into a variety of small and medium-sized models to meet the splitting of various scenarios and knowledge and improve the adaptability with customers.

Similar to Google, Chuangda’s Rubik large model is expected to be the first to be launched as a robot. Among the many intelligent hardware products, Thunderstar's robot products can cover a variety of different application scenarios, and have helped many robot manufacturers around the world achieve mass production of their products. Based on its profound accumulation in the field of robotics, Thunderousas has integrated smart speakers with robots, and through continuous training with the Rubik large model, it has realized an intelligent sales robot that can talk freely and can independently answer customers' questions about the company and products. kind of problem.

1.3.2 Vertical view: Chuangda AI application ecosystem is ready to go

The Smart to Intelligent strategy of Thunderstar is launched to realize a new intelligent world from intelligent application-centered to model-driven machine-machine, machine-human interaction. The Rubik large model will be integrated with the company's existing smart car and IoT businesses and meet the needs of various industries through privatized deployment and system tuning. The company is expected to turn robots into reality through continuous optimization of large models and exert its core competitive advantages in the toBtoC field of the future intelligent computing industry.

RUBIK Auto: A car is essentially a robot. Car manufacturers have three main needs for large car models. The first is terminal-side operation. The terminal-side experience, data, and performance are the best and can best protect customer privacy. , but the hardware requirements are higher; the second is private cloud Plugin, which can be flexibly tuned; the third is the coexistence of multiple open models. The company's RUBIK Auto will support customer privatization deployment (it has cooperated with leading overseas car manufacturers to conduct POC research and development based on the company's model), and also supports model quantification and tailoring, and then adapts to various chips to flexibly interface with car manufacturers.

RUBIK Device: Among smart hardware, as long as it involves large computing products, Chuangda’s leading smart hardware originally deployed operating systems can directly add AI. Once AI settles on the edge, it means that the smart hardware becomes a robot. It will form the center of the scene. Whether it is a home scene, a building scene, or a factory scene, every scene can become an intelligent center through edge deployment.

RUBIK Enterprise (Enterprise Edition): One of Chuangda’s obvious advantages is internationalization. The company has R&D centers and teams in 15 countries and regions around the world. The company will deploy through localization and support customers' privatized deployment; the second obvious advantage of the company is the integration of software and hardware. Whether it is inference or training, the company has a very good understanding of the entire underlying platform, because whether it is the AI ​​framework or the Other open source frameworks are essentially part of the operating system. Whether it is data parallelism or model parallelism, the company believes that they are all middleware. These obvious advantages allow us to ultimately optimize model performance and efficiency, optimize model scale, run it on the edge, and empower thousands of industries.

RUBIK Studio: Based on the company’s deep understanding of the operating system, the company differentiates and models each segment of the operating system, accumulating thousands of successfully released packages in the past, hundreds of billions of lines of code accumulation, and long-term development experience Accumulate, turn Android knowledge into a huge knowledge base, and exert huge value. Rubik Studio will be a big tool and environment that changes the entire future. Users can directly and quickly develop PC applications, mobile applications, and websites with relatively closed functions through Rubik Studio.

1.4 Apple: ChatGPT is first connected to IOS, WWDC 2023 is worth looking forward to

On May 18, 2023, OpenAI released the IOS version of the ChatGPT APP (at the same time, it was officially announced that the Android version of the APP is under development). According to Apple’s official website, the application only supports English, has an age rating of “12 years and above”, and supports network synchronization. Chat records, whisper voice input and other functions. The IOS version of the ChatGPT download application is free, but a "ChatGPT Plus" paid item priced at US$19.99 within the APP will be provided to enable the use of GPT4 large models with stronger performance.

Rewritten sentence: ChatGPT APP may become the prototype of mobile phone super APP, and reconstruct the artificial intelligence application ecosystem based on this. According to Jiupai Finance, the IOS version of ChatGPT APP already has the ability to flexibly solve various daily problems. ChatGPT APP can provide instant answers, and users can obtain accurate information without filtering advertisements or multiple results; it provides customized suggestions, and users Turn to them for guidance on cooking, travel planning, or crafting a thoughtful message; provide creative inspiration, generate gift ideas for users, outline a presentation, or write beautiful poetry. In addition, ChatGPT APP can also help users improve work efficiency through professional information, such as idea feedback, note summaries, and technical topics, and provide users with learning opportunities to help them explore new languages, modern history, etc. at their own pace. We believe that under the premise of the successful implementation of the AI ​​application ecosystem of ChatGPT Plugin, ChatGPT APP is expected to become the core of the mobile AI application ecosystem. Through the training of massive interactive data from mobile phone users, in the future users are expected to use ChatGPT APP to call other applications to complete various rigid tasks. , ChatGPT APP will develop into an unprecedented super APP.

The slogan of Apple WWDC 2023 is "Code new worlds". As an important entrance to human-computer interaction, Siri is expected to become an important starting point for Apple to enter AI. According to IT House, Xinzhiyuan, and Tencent Cloud, Apple will begin developing applications that allow users to use Siri to create programs for MR in February 2023 at the latest. The technology behind this application construction method comes from Apple’s acquisition of MR in 2017. Fabric Software: Users can use Siri to build AR apps, asking the AI ​​assistant to help build apps that allow virtual animals to move around a room, on top of and around real objects, without having to design the animal from scratch, program its animations, and calculate its movements. Applications for movement in 3D space with obstacles, which include “scanning and importing real-world objects into the headset so that they can be accurately represented in 3D and behave as if they existed in real life.”

1.5 Huawei: IEF Gaussian database fully covers edge AI scenarios

Huawei Intelligent Edge Platform IEF is an edge-cloud collaborative operating system built on cloud-native technology. It can run on a large number of heterogeneous edge devices and integrate rich AI, data analysis, middleware and other applications from The cloud is deployed to the edge to meet users' business demands for edge-cloud collaboration of intelligent applications. IEF can extend Huawei Cloud's AI/big data capabilities to the edge, support intelligent video analysis, text recognition, image recognition, big data stream processing and other capabilities, and provide real-time intelligent edge services nearby; it supports two operating modes: container and function. It meets users' demands for lightweight application management; it natively supports kubernetes and docker ecology, and applications can be quickly started and upgraded quickly; it supports function engines such as Python and NodeJS to quickly respond to edge events; in addition, it has excellent compatibility, security and reliability.

Huawei Gauss Database is an enterprise-level distributed relational database launched by Huawei based on the openGaussDB self-research ecosystem. It has enterprise-level complex transaction mixed load capabilities, supports strong consistency of distributed transactions, can be deployed across AZs in the same city, has zero data loss, and supports 1000 Expansion capabilities, PB-level mass storage. At the same time, it has key capabilities such as high availability, high reliability, high security, elastic scaling, one-click deployment, fast backup and recovery, and monitoring and alarming on the cloud. It can provide enterprises with an enterprise-level database with comprehensive functions, stability, reliability, strong scalability, and superior performance. Serve. On June 7, 2023, Huawei Goss will hold a conference on database acceleration of core financial business upgrades.

Investment advice: The general trend of AI from cloud to edge has been established, and our early judgment has been continuously verified: edge AI is an industry trend, and embodied intelligence is the internal logic (AI self-improvement requires the interaction data between humans and the environment to be concentrated on the terminal) , robots are the ultimate application. We released a report on May 13, 2023, "China Science and Technology Thunder: Large Models from Cloud to Edge, Terminal Interaction Revolution Breeds Historical Opportunities", which clearly stated that Google will vigorously enter the terminal large model market, and terminal AI will become the next must-have for military strategists. land; then released "Google's "Empire Strikes Back": The Inflection Point of AI from Cloud to Edge" on May 14th, detailing Google's AI blueprint from cloud to edge, and clearly stated that AI edge deployment has entered reality; through the release AI experience reports on the IOS version of ChatGPT, Windows AI and other AI experience reports, we released "ChatGPT APP marks a new stage of AI market" on May 21, clearly proposing that the large model is the ultimate operating system in the AI ​​era. ChatGPT super APP is only the first step, generative AI's progress from cloud to end is still accelerating. This week, companies such as Nvidia, Qualcomm, and Thunderstar have released related products to speed up the implementation of edge AI. Next week, Huawei and Apple’s press conferences are expected to focus on terminal AI. The general trend of AI from cloud to edge has become extremely clear. , it is recommended to pay attention to leading companies such as Chuangda, iFlytek, and EZVIZ Network.

Source: Selected Brokerage Research Reports

The above is the detailed content of Minsheng Securities: Edge AI is an industry trend. It is recommended to pay attention to these targets.. For more information, please follow other related articles on the PHP Chinese website!