We mentioned in the introduction of the previous article: "AI artificial intelligence, a concept that has existed for decades, has become a core variable in the development of human society in the next few decades. In the face of change, everyone and every enterprise needs to think about how to leap forward with the trend of the times." Recently, the MeiG Intelligent R&D team successfully ran Stable Diffusion, a large Vincentian graph model, on the self-developed high-computing power module, and this is just the first step in turning our thinking into practical actions. The era of AIGC has just begun!
▌Not limited to the cloud, the technical path of edge computing + terminal-side AI has matured
Since 2022, AIGC, as a transformative technology, is gradually subverting the original way of work, life, creation and entertainment of human beings. Various generative large models represented by ChatGPT, Midjourney, Stable Diffusion, Wenxin Yiyan, etc. can complete various tasks including searching, drawing, creating, table editing, code writing, etc. within seconds, and various capabilities are constantly evolving. Data shows that there are at least a thousand generative AI applications available, all of which show that AIGC is ushering in an unprecedented period of explosion.
Numerous generative AI models with billions of parameters place extremely high demands on computing architecture and computing power. The previous situation was that whether it was AI training for optimizing parameters or AI reasoning for executing models, they had always been limited by large and complex models and insufficient deployment of end-side computing power, and were bound to the cloud.
Preliminary estimates show that the generative AI market has huge potential and will widely cover all participants in the ecological chain. This huge market opportunity is bound to drive the simultaneous evolution of different types of computing architectures and sources of computing power, so as to meet the growing computing power demand of large-scale generative AI.
The recent actions of global top players also show that the trend of computing power construction and computing power competition has just begun, and everyone is starting. Qualcomm announced not long ago that it has completed the world's first terminal-side demonstration of Stable Diffusion running on an Android phone. This is the first instance of a large-scale language-generated AI model supported by the second-generation Snapdragon 8 mobile platform running on the terminal side.
Subsequently, Qualcomm's official statement on the transformation of edge computing companies, as well as its vigorous promotion of hybrid AI, all indicate that the technology and product path of edge computing + terminal side AI is rapidly becoming a reality.
▌The ability of high-computing AI modules has been further evolved, and the Stable Diffusion large model has been successfully run
As the first enterprise in the industry to propose the concept of intelligent modules, MeiG Intelligent continues to pay attention to the profound changes brought about by the development of computing power to the module industry. It is also the first to propose that: with the explosive growth of computing power demand and applications in various industries, high computing power AI modules will become a subset of smart module development. "No computing power is not smart" is becoming an important development trend in the module 3.0 era. High computing power AI module products will be independently formed product teams and resources will be continuously invested.
Recently, the MeiG Intelligent high-computing power AI module product team gave full play to the ability of integrated software and hardware collaborative development, and successfully ran the Vincent graph large model Stable Diffusion on the self-developed computing power module based on the Qualcomm Snapdragon 800 platform.
Stable Diffusion is an excellent generative AI large model from text to image, which can create realistic images in tens of seconds based on any text input. Stable Diffusion has over a billion parameters and has so far been largely limited to running in the cloud.
▲Meige Intelligent High Computing Power AI Module SNM970 EVB Development Board
The AI module team of Meig high computing power has carried out the research work on the operation of large models on the end side of the computing power module in a forward-looking manner. The high-computing module that runs the Stable Diffusion large model uses the Qualcomm Snapdragon 800 series high-computing SoC processor for the hardware, and the Qualcomm AI Model Enhancement Toolkit (AIMET) for the software to achieve model conversion. Without retraining, it overcomes the loss of accuracy caused by the compression of model parameters from FP32 to INT8, so that the large model can always run efficiently on the computing module.
The MeiG Intelligent R&D team further optimized the hardware and software to ensure the running accuracy of the large model while effectively reducing memory bandwidth consumption and giving full play to the maximum performance of the AI hardware of the high-computing module. Finally, the Stable Diffusion large model was successfully run on the side of the computing power module.
▲Example of the picture generated by the large model, the prompt word is: a dog wearing a spacesuit, running in space
▌Create and lead the development of high-computing AI modules, and drive the deployment of generative AI on the device side with product power
MeiG Intelligent continues to develop the high computing power AI module product line, and has experienced three generations of product evolution so far. Relying on the powerful integrated research and development capabilities of software and hardware, MeiG Intelligent continues to promote the continuous upgrading of AI hardware of high-computing AI modules. The design and development capabilities of the MeiG R&D team in the fields of AI application scenario development, AI performance optimization, and AI low-power program development are at the leading level in the industry, and related computing power products are widely used in various core scenarios.
In the field of cloud computing servers , the high-computing AI module launched by MeiG Smart is based on powerful computing power, low-power chips and MeiG Smart’s unique customized solutions. It is perfectly applied to the leading arm-based array server cluster. Under the same computing tasks, the power consumption is increased by more than 2 times.
In the field of intelligent robots , high-computing AI modules can easily match the high computing power requirements of intelligent robots and next-generation embodied intelligent robots. At the same time, they are equipped with rich peripheral interfaces. In the form of AI modules, speech recognition, machine vision, navigation, positioning, obstacle avoidance and multi-modal technologies are implanted into the terminal side through various sensors, truly realizing the AI awakening of robots.
In the fields of industrial vision, AI retail and other fields , the standardized AI kit launched by MeiG Intelligent can support the revolutionary INT4 Al precision format in terms of AI acceleration, and achieve a 60% improvement in energy efficiency compared with the previous generation in terms of continuous AI reasoning. At the same time, it can perform video structural analysis on multiple cameras, quickly realize the functions of intelligent perception, data analysis, security management, etc. of people, things, and things, and quickly realize AI embedding in many traditional industries.
As the earliest enterprise in the industry to deploy computing power module products, MeiG Intelligent high computing power AI module products have achieved full coverage from 1.2T to nearly 48T products, fully meeting the computing power needs of different industries. The R&D team has accumulated a lot in areas such as the docking of large computing power platforms and algorithms in different industries, low power consumption control of AI operation, and 5G+AI integrated development.
With the successful operation of the AI large model on the self-developed computing power module product for the first time, the capability boundary of the computing power module has been greatly expanded, and the forward-looking experience of large model deployment and application on the device side has been accumulated, which will truly drive the large-scale deployment of generative AI on the device side and edge side, and make the technical picture of edge computing + terminal side AI more and more clear!
The number of Internet of Things connections has grown exponentially in the past ten years, and the "Internet of Everything" that was once written in the book has become a reality. The ubiquitous connection has laid a solid foundation for the digitalization and intelligence of human society. When the connection and computing power are integrated, the computing power of hundreds of millions of IoT nodes will converge to build the largest computing power network, interconnected with cloud computing power, and jointly build the computing power base of AIGC and other cross-age technologies.