The Path of Ascension: From AI Basic Software to Ecological Prosperity

In this year’s government work report during the Two Sessions, it was clearly stated to deepen the research and application of big data, artificial intelligence, and other technologies, to carry out the “Artificial Intelligence +” initiative, and to build a digitally competitive industrial cluster. With the first inclusion of “Artificial Intelligence +” in the government work report, various industries and fields have actively explored and pursued AI technology, reaching a new height of enthusiasm for its development.

At the current stage, AI technology is an important engine for developing new productive forces, and it has become a high consensus among all sectors of society. How to continuously introduce AI technology into various production scenarios has become the most focused topic for the industrial and academic communities.

We must see that the universal development of any technology cannot be separated from its robust growth. For example, the arrival of the “Internet +” era relies on technologies such as IP networks and the construction of social network infrastructure. The implementation of the “Artificial Intelligence +” initiative also depends on the construction of various new AI infrastructures. Among them, the most important aspect on the software level is the AI framework.

The AI framework is the software foundation for large models, the first stop for various industries and fields to engage with artificial intelligence technology and develop large models, and the core force driving the intelligent development of industries. How can the AI framework develop steadily in the face of the new era and new tasks? How can the technical capabilities of the framework be transformed into the driving force for academic applications and the cohesive force for ecological development? These questions urgently need answers today.

Recently, the Ascension Artificial Intelligence Framework Summit 2024 was held at the National Convention Center in Beijing with the theme of “For Wisdom and Ascension, the Source of Thought Creation.” At the summit, the newly upgraded Ascension MindSpore 2.3 version was announced, showcasing the latest progress of Ascension in various aspects such as academic applications and ecological empowerment.

How to establish a pathway from technological leadership to deep academic applications and empower ecological acceleration of development?

In this summit, Ascension has already provided its answer.

The so-called AI pathway should be a road that efficiently extends underlying technologies into upper-level applications and development ecosystems, allowing various sectors of the economy to draw nourishment from foundational technologies. Only when this road is smooth, convenient, and low-cost can AI technology fully blossom and bear fruit, maximizing the value of the “Artificial Intelligence +” initiative.

However, at the current stage, achieving such an AI pathway still faces a series of challenges. There are still bottlenecks and blockages between technology and industrial economy, academic research, and developer ecosystems. For example, the development cost of large models is high and the difficulty is great, making it hard for various industries and fields to fully master them; moreover, some specialized and complex fields still cannot achieve deep technological penetration. Overall, transforming AI technology into the driving force for social and economic development requires upgrades on three levels:

1. Foundation Layer. There is a need to achieve higher training efficiency for large models, lower inference deployment costs, and overall improvement in development usability, achieving a comprehensive upgrade in training and inference capabilities and development efficiency. This especially requires the upgrade and development of deep learning development frameworks, achieving continuous evolution in training, inference, and development tools.

2. Application Layer. AI technology needs to interact deeply with industry scenarios and research fields, achieving a three-dimensional integration of artificial intelligence technology and industry-specific knowledge, enabling the intelligent upgrade of industries, thereby expanding the application boundaries of AI technology and enhancing scenario replication capabilities. Especially in key areas like AI for Science, AI technology needs to closely cooperate with basic discipline researchers for joint exploration.

3. Ecological Layer. Like other software technologies, AI technology cannot scale without an ecosystem; a prosperous ecosystem can achieve the effect of technology being used better and better. The development of the ecosystem requires effective interaction between technological platforms and various sectors of industry, academia, and research, as well as countless developers, continuously empowering them.

It can be seen that for AI technology to become the driving force for social and economic development, a complete pathway must be constructed from underlying technology to academic applications, and then to ecological empowerment. This is precisely the AI pathway that Ascension is actively practicing.

First, a question arises: how to ensure that AI technology continues to innovate and meets the expectations and requirements of various sectors for AI basic software.

In this regard, Ascension is continuously improving. Achieving a three-dimensional and collaborative upgrade of capabilities in training, inference, and model development. This idea is fully reflected in the latest release of Ascension MindSpore 2.3. Through a series of upgrades, Ascension has made large model development training simpler, more stable, and more efficient.

In terms of training, Ascension can support large models to achieve native high-efficiency training. Through original multi-replica, multi-stream interleaving, and other eight parallel technologies, the linearity of the cluster reaches 90%, far exceeding the industry average of less than 60%; through whole graph optimization and down-sinking execution, the computing power utilization rate reaches 55%, significantly surpassing the industry’s average of less than 40%; addressing the common problem of high cluster failure rates and long recovery times, Ascension achieves recovery within 20 minutes through deterministic CKPT technology.

In today’s era of large models, large-scale clustered model training has become the industry’s biggest pain point. Ascension has comprehensively cracked core issues such as cluster linearity, AI computing power utilization efficiency, and training failure recovery through a series of foundational capability upgrades, making large model training no longer filled with challenges.

In terms of deployment, Ascension has achieved one-day completion of Baichuan2-13B inference deployment through training-inference integrated architecture upgrade scripts, distributed strategies, and unified runtime; in large model inference, it has achieved more than double the inference throughput through LLM Serving; by upgrading the model compression tool Gold Rod 2.0, it has compressed hundreds of billions of large models to one-tenth. These capabilities ensure that AI large models are not only trainable but also smoothly applicable, bridging the last mile of AI implementation.

In addition, Ascension has continuously upgraded the MindSpore Transformers large model suite and provided the MindSpore One generative suite, comprehensively enhancing developer efficiency, allowing developers to complete the entire process of large model development within a week. In the key AI for Science field, Ascension has teamed up with top research institutions and partners to create an AI bio-computing suite, including more than 20 SOTA models for protein structure prediction and generation, accelerating scientific innovation in related fields.

From training to inference, and to a series of development suite upgrades, Ascension continues to delve into AI technology, stabilizing the foundational capabilities of AI.

Only when foundational software technology can continuously provide nourishment can AI applications germinate, and the AI ecosystem blossom.

After solidifying the foundational AI technology, it is necessary to deeply integrate technological capabilities with application scenarios. Especially in complex and critical fields, such as AI for Science, it is essential for AI technology and research work to conduct leapfrog exploration, allowing the AI framework to sprout and grow in countless industries and fields.

At the summit, we can see how Ascension integrates with the key field of aerodynamic shape design, opening up a new paradigm for AI for Science. Aerodynamic shape design refers to the aerodynamic shape design of transportation tools such as airplanes, ships, and cars. This field is related to the national economy and people’s livelihood, and the cost of research and verification is enormous. It is very suitable for the addition of large AI models as new research tools. Previously, we often saw the application of large AI models in protein folding and materials analysis research fields; the integration of Ascension with aerodynamic shape design once again broadens the boundaries of AI for Science.

Academician Tang Zhigong, President of the Chinese Society of Aerodynamics, stated: Based on Ascension MindSpore, the generative aerodynamic design large model platform breaks traditional design paradigms, reducing design time from months to minutes, meeting conceptual design requirements. In the future, this platform will expand to multiple aerodynamic fields such as aviation, aerospace, shipping, high-speed rail, energy, and automobiles, leading to a leapfrog development in the design and manufacturing capabilities of large industrial equipment.

It is understood that the generative aerodynamic design large model platform was independently innovatively developed based on the Ascension AI framework. During the model development phase, the Ascension framework and fluid mechanics suite MindSpore Flow provide a comprehensive scientific computing algorithm library and universal model interfaces, enhancing model development efficiency. In the training phase of the model, the MindSpore multi-dimensional distributed parallel interface, supported by the computing power of the Chengdu Intelligent Computing Center, allows for efficient expansion of models and data. In the model deployment phase, the Ascension large model suite can integrate specialized knowledge into the aerodynamic design large model platform.

By linking large language models, aerodynamic shape design models, aerodynamic prediction models, and non-AI tools such as wind and thunder software, the aerodynamic design large model platform can support multiple aerodynamic shape design scenarios, providing foundational scientific research support for a series of key fields.

In various industries and research fields, we can see the new sprouts of AI applications grown from Ascension’s AI technology. Given time, these sprouts will grow into towering trees.

Thus, AI technology will bear fruit and nourish all.

Looking up from the layer of academic applications, the future development of AI technology relies on a vast and prosperous developer ecosystem. This requires active cooperation between the AI framework and various sectors of industry and academia, continuously strengthening talent cultivation and ecological empowerment. In this field, Ascension adopts a strategy of comprehensively empowering academia and ecology, using a multi-pronged approach to integrate Ascension’s technological capabilities with developer needs, assisting in AI talent cultivation.

During the summit, Ding Cheng, chairman of the Ascension MindSpore open-source community, announced four actions to empower academia and ecology.

These include the academic paper fund 2.0 that Ascension is collaborating with the Chinese Association for Artificial Intelligence and Pengcheng Laboratory, which will explore continuously in the academic field with over 50 global AI scholars in the next three years; the application innovation action based on the Orange Pi development board, providing systematic cases, tutorials, and support to help developers quickly get started, rapidly create personalized applications; accelerating the incubation of native large models through incentives, specialized technical support, and joint market promotion, supporting more partners to migrate from Ascension adaptation to native development; and the open-source community internship activities, allowing developers to practice and grow through coding.

In addition to empowering ecological development through transformation actions, Ascension actively promotes open-source innovation, consolidating the open-source ecosystem. In March 2020, the open-source Ascension MindSpore AI framework was officially launched, which has since received positive feedback from a large number of AI developers, with total visits reaching thousands, over 6.87 million downloads and installations, ranking first in the Gitee open-source project, serving over 5,500 enterprises and collaborating with 360 universities.

Since 2023, the number of top conference papers published based on the Ascension framework has exceeded 1,200, ranking first in China and second globally among all AI frameworks, making it the most innovative AI open-source community in the country. Gitee is the largest open-source code hosting platform in China. According to the Gitee Index 2.0, MindSpore performs excellently across various indicators, becoming the No. 1 in the Gitee-AI field classification. Thus, Ascension MindSpore has been recognized as “Gitee’s Best Open Source Contribution Project” and praised by Ma Yue, chairman of Open Source China, as “a treasured gem of Gitee.”

The blooming ecological flower, the open-source flower, confirms that Ascension’s AI pathway has been fully interconnected. From AI foundational technology to applications in various industries, and to the trust of countless developers, Ascension has built a comprehensive and three-dimensional AI development foundation.

In the wave of AI technology, the hope of social economy is that intelligent capabilities can be compatible with efficiency and stability, developing well and quickly.

Ascension’s AI pathway is to solidify the technology, let applications sprout robustly, and allow the ecology to bloom upwards. Based on this, AI can grow into towering trees and become a pillar of society. Only then can the era’s call for intelligence become a reality, and AI technology can be the source of new productive forces.