The demand for computing power has exploded under the wave of AI large models. SenseTime's 'large model + large computing power” empowers the development of multiple industries.

Recently, the "Lingang New Area Intelligent Computing Conference" with the theme of "AI leads the era, computing power drives the future" was held. At the meeting, the New Area Intelligent Computing Industry Alliance was formally established. SenseTime became a member of the alliance as a computing power provider. At the same time, SenseTime was awarded the title of "New Area Intelligent Computing Industry Chain Master" enterprise.

As an active participant in the Lingang computing power ecosystem, SenseTime has built one of the largest intelligent computing platforms in Asia - SenseTime AIDC, which can output a total computing power of 5,000 Petaflops and support 20 100 billion parameters. A large number of very large models are trained simultaneously. SenseCore, a large-scale device based on AIDC and built forward-looking, is committed to creating high-efficiency, low-cost, and large-scale next-generation AI infrastructure and services, empowering a new paradigm of artificial intelligence production, and will become an infrastructure service in the AGI era. leader.

Under the wave of large AI models, the demand for computing power has exploded

The three major elements of artificial intelligence mainly include data, algorithms and computing power. According to data recently released by OpenAI, the computing power used in artificial intelligence training tasks has increased exponentially since 2012, with a growth rate doubling every 3.5 months. So far, people's demand for computing power has increased by more than 300,000 times. The popularity of ChatGPT has triggered new market demands for computing power.

At present, my country's computing power market continues to grow. According to estimates from the Academy of Information and Communications Technology, the total scale of my country's computing equipment computing power will reach 202 EFlops in 2021, with a growth rate of about 50%, which is higher than the global growth rate.

In this context, Shanghai Lingang actively leverages the advantages and ecological traction of the local computing industry and released the "Lingang New Area Action Plan to Accelerate the Construction of a Computing Industry Ecosystem" (hereinafter referred to as the "Plan").

According to reports, the computing power industry in Lingang New Area has made corresponding arrangements in upstream software and hardware, midstream data centers, dispatching platforms, and downstream applications. At present, the total computing power of Lingang exceeds 3EFLOPS (FP32), and the intelligent computing power accounts for than nearly 80%, and the total computing power accounts for nearly 20% of Shanghai.

The "Plan" proposes that by 2025, the Lingang New Area will form a diversified computing power supply system that focuses on intelligent computing power and coordinates basic computing power and super computing power, with a total computing power exceeding 5EFLOPS (FP32) , AI computing power accounts for 80%, and the overall scale of the computing power industry exceeds 10 billion yuan. A public computing power service platform will be built, the computing power trading mechanism will be standardized, regional computing power dispatch will be realized, and a computing power industry cluster with national influence will be created. , and build a number of benchmark scenarios for computing power demonstration applications.

Xu Li, chairman and CEO of SenseTime, said that computing power is the energy source of the new era. To a certain extent, computing power determines the competitiveness of the market. "Computing power is an expression of the entire model's capabilities, which is equal to the parameters of the algorithm or large model multiplied by the amount of data it processes. In the era of large models, the larger the parameters, the greater the amount of data multiplied, and the greater the computing power required. big."

At the same time, the Lingang New Area Intelligent Computing Industry Alliance was officially established. The members of the industry alliance are represented by 25 companies and 3 universities and research institutes. In the future, resource sharing, technical exchanges and project cooperation will be carried out. Promote the application of intelligent computing industry in the new area to empower economic development.

SenseTime was awarded the title of "Chain Master of the Intelligent Computing Industry Chain in the New Area". The SenseTime Intelligent Computing Center located in the Lingang New Area bears the important task of carrying out large-scale artificial intelligence research and development and industrialization in the Yangtze River Delta, and will actively participate in the follow-up To the collaborative integration and clustered development of the Lingang intelligent computing industry chain.

Big model Big computing power integration innovation

The integration of large models and large computing power is causing a major shift in the production paradigm, pushing scientific research and industrial applications towards the era of general artificial intelligence (AGI) driven by intelligent computing. In the early stages of development with rapid technology iteration, the industry urgently needs to build a new generation of infrastructure to lower the application threshold, shorten the research and development cycle, and improve innovation efficiency.

SenseTime Technology laid out its plans ahead of time. It took five years to build SenseCore, a large-scale device of SenseTime. On this basis, it built the "SenseTime Daily New SenseNova" large-scale model system to provide the industry with large-scale model algorithm services and training. And AGI infrastructure that combines software and hardware such as inference optimization and data services.

According to reports, SenseCore, a large device of SenseTime, uses the SenseTime Artificial Intelligence Computing Center (referred to as "SenseTime Intelligent Computing Center or SenseTime AIDC") as the computing power base. It contains 27,000 GPUs and can output a total computing power of 5,000 Petaflops. , with industry-leading computing power output capabilities, ultra-large model training and large-scale reasoning capabilities, it is currently one of the largest intelligent computing platforms in Asia.

The current computing power of SenseCore, a large device of SenseTime, can support the simultaneous training of 20 ultra-large models with hundreds of billions of parameters, and provides a one-stop large model infrastructure service system covering data, training tools, inference deployment, and performance optimization.

SenseTime’s large device has excellent parallel computing capabilities and can conduct single-task training with a maximum 3200 card scale cluster, and can achieve uninterrupted stable training for more than seven days. It not only supports SenseTime’s own large model training projects, It also trained models customized by other companies.

In addition, SenseTime large-scale devices integrate the core capabilities of AI, supercomputing and big data. Through high-performance computing, high-performance storage and caching, and high-performance networks optimized for AI, it achieves separation of storage and computing, large-scale elasticity, Features such as fault-tolerant scheduling support the training of large models with trillions of parameters on thousands of cards and PB-level storage.

SenseCore AI platform products also provide modular, full-chain data, training and reasoning capabilities. It can realize tens of billions of data management and retrieval, manual annotation services, and accelerate the efficiency of AI large model development. One-click quantification, one-click deployment, and one-click application provide tools for rapid online verification of large models and accelerate innovation.

In addition, Big Device also provides customers and ecological partners with full-chain MaaS big model-as-a-service, accelerating the innovation and application efficiency of large models.

Among them, the automated data annotation service can increase the efficiency of intelligent annotation by a hundred times; the large model inference deployment service can increase the efficiency of large model inference by 600%; the large model parallel training service supports single cluster 3200 cards and 500 billion dense parameter model training; Model incremental training service can reduce incremental fine-tuning costs by 90%.

Sensetime AI large model empowers multi-industry development

Enabled by large devices, SenseTime has achieved rapid development in the field of large models.

According to Xu Li, the "Scholar 2.5" multi-modal large model that was open sourced in March this year has taken the lead in more than 20 authoritative data sets in the three mainstream visual tasks of detection, segmentation and classification. This provides opportunities for autonomous driving, Provide efficient and accurate perception and understanding support for general scenario tasks such as robots.

For artificial intelligence basic science (AI For Science), among meteorological and climate forecast tasks, global medium-term weather forecast is one of the most important prediction tasks. The large global medium-term weather forecast AI model "Fengwu" launched in April this year achieved effective forecasting of core atmospheric variables at high resolution for more than 10 days for the first time, and surpassed the GraphCast model in 80% of the evaluation indicators. Thanks to high-resolution global atmospheric data modeling, "Fengwu" can also simulate extreme weather such as typhoons and accurately predict typhoon trajectories.

UniAD, the industry's first end-to-end autonomous driving solution with integrated perception and decision-making, has surpassed the SOTA method in a number of key data sets and indicators, improving lane line prediction accuracy by 30%. The error in predicting motion displacement is reduced by nearly 40%, and the planning error is reduced by nearly 30%.

In addition, the SenseEarth 3.0 remote sensing large model launched by SenseTime last month not only has the most comprehensive interpretation categories in the industry, but also has achieved technological breakthroughs in many indicators such as interpretation efficiency, generalization ability, and interpretation accuracy. .

Xu Li said, "In the AGI era, the ability of a model can be measured by computing power. We use SenseCore, a large device of SenseTime, to build the infrastructure of the AGI era and update it daily in terms of model iteration speed and problem-solving capabilities. , and constantly unlock more possibilities of AGI.”

It is reported that as of May this year, SenseTime has served more than 40 core customers, including more than 10 large model customers, covering cutting-edge fields such as intelligent driving, biopharmaceuticals, chip design, smart business, and university scientific research. And has achieved large model delivery in more than 20 landing scenarios.

Yang Fan, co-founder of SenseTime and president of the large device business group, said that the performance development of large models seen today is the improvement of technical value brought about by the continuous increase in the scale of the three elements of artificial intelligence. A perfect combination of basic R&D capabilities and systematic engineering capabilities. These three elements are often jointly tuned. Algorithm optimization, data sorting and selection, and computing power platforms are often interconnected. It is difficult to turn them into separate links and do them alone. This is why we need to build an intelligent computing power industry chain, because only if more companies on the chain promote exchanges, cooperation, thinking, and more in-depth cooperation can we do better in the new critical wave of major technologies. technological advancement and support.

The above is the detailed content of The demand for computing power has exploded under the wave of AI large models. SenseTime's 'large model + large computing power” empowers the development of multiple industries.. For more information, please follow other related articles on the PHP Chinese website!