Alibaba Cloud launches FeiTian intelligent computing platform, which can improve AI training efficiency by 11 times

Alibaba Cloud launches FeiTian intelligent computing platform, which can improve AI training efficiency by 11 times

On August 30, Alibaba Cloud announced the official launch of the full-stack intelligent computing solution "Feitian Intelligent Computing Platform" and launched two ultra-large-scale intelligent computing centers, providing powerful intelligent computing services for various scientific research, public services and corporate institutions in both public cloud and private cloud modes. It can increase the utilization rate of computing resources by more than 3 times, AI training efficiency by 11 times, and reasoning efficiency by 6 times.

The FeiTian intelligent computing platform has been widely used within Alibaba, supporting the development of cutting-edge AI and e-commerce intelligent technologies at the DAMO Academy. It has also served institutions and companies such as Xiaopeng Motors, Deepin Technology, SAIC Motor, China Meteorological Administration, and China Southern Power Grid, supporting industries such as autonomous driving, new drug research and development, weather forecasting, and industrial energy to significantly improve AI training efficiency.

It is understood that the platform provides an overall solution of integrated computing power and big data AI integrated platform based on Alibaba Cloud Panjiu infrastructure. It can run on servers with various chip types such as X86, GPU, ARM, etc., realizing "one cloud, multiple cores" and achieving 90% kilocalorie parallel computing efficiency with up to 10 times IO optimization and 5 times communication performance optimization.

In terms of green technology, FeiTian Intelligent Computing reduces carbon emissions per unit computing power in five aspects: technical emission reduction, energy structure optimization, regional layout optimization, supply chain carbon reduction, and resource utilization optimization. In terms of technical emission reduction, energy consumption is reduced through liquid cooling, power supply technology, and intelligent operation and maintenance, with the lowest PUE reaching 1.09.

At the same time, developers can perform data storage, data governance, data analysis, model development, model training and reasoning on the platform. It also provides pre-trained models, as well as model capabilities in the fields of speech, image, natural language processing, decision-making, etc., to facilitate developers to better accelerate the development of AI applications.

Currently, the platform is supporting the construction of two ultra-large-scale intelligent computing centers. Among them, the Zhangbei Intelligent Computing Center has a construction scale of 12 EFLOPS (120 billion floating-point operations per second) AI computing power, which will exceed Google's 9 EFLOPS and Tesla's 1.8 EFLOPS, becoming the world's largest intelligent computing center. The Ulanqab Intelligent Computing Center has a construction scale of 3 EFLOPS (300 billion floating-point operations per second) AI computing power and is located in the Inner Mongolia hub of "Eastern Data and Western Computing".

Cai Yinghua, President of Global Sales at Alibaba Cloud Intelligence, said that intelligent computing is not only about scale, but also needs to be green, efficient and have industrial practice. Computing is a huge and complex system. Without systematic core technical capabilities, piling up hardware will not produce computing power, let alone bring actual industrial value.

It is understood that intelligent computing is different from general computing. It requires massive data to train AI models, and computing power is lost in data migration and synchronization. The minimum computing power output of 1,000 calories is often only about 40%. This leads to high costs for intelligent computing and restricts the development of the industry. Alibaba Cloud has changed the problem of intelligent computing loss through systematic technological innovation, and increased the efficiency of 1,000 calories of parallel computing to more than 90%.

For example, in terms of communication technology, Alibaba Cloud uses a high-performance self-developed Solar-RDMA network to achieve an end-to-end minimum latency of 2 microseconds. Combined with Alibaba Cloud's self-developed non-blocking communication technology, the data exchange speed in the computing process is increased by up to 5 times. At the same time, the application of green technologies such as natural air cooling and liquid cooling reduces the energy consumption of the intelligent computing center, with a PUE as low as 1.09.

At the AI ​​development layer, Alibaba Cloud provides a big data + AI integrated platform to support the entire development and operation process. In particular, in the model training phase, it provides a distributed training framework that can automatically combine and optimize distributed strategies, increasing training efficiency by more than 11 times. In addition, Alibaba Cloud provides users with a one-stop general reasoning optimization tool that can perform operations such as quantization, pruning, sparsification, and distillation on algorithm models, which can increase reasoning efficiency by more than 6 times.

Not long ago, Xiaopeng Motors built the "Fuyao" intelligent computing center in Ulanqab based on Feitian Intelligent Computing, with a computing power of 600PFLOPS, which is the largest intelligent computing center for autonomous driving in China, and has accelerated the training of autonomous driving models by nearly 170 times. Based on Feitian Intelligent Computing, Haomo Auto has achieved a 128-card parallel efficiency of over 96%, reducing the cost of autonomous driving model training by 62%, increasing the training speed by 110%, and significantly shortening the model iteration cycle.

In the field of life sciences, after using the FeiTian intelligent computing platform, Deepin Technology has improved cluster performance optimization by more than 100% , and increased the efficiency of molecular dynamics simulation training by 5 times . In the industrial field, Zhiji Auto has used high-performance computing to improve the efficiency of industrial simulation by 25% , and the efficiency of intelligent driving training by 70%, accelerating the development and launch of new models. Shandong Dezhou Electric Power uses AI to conduct review and prediction with an accuracy rate of 98% , and the time taken is reduced from 1 hour to a few minutes .

In the field of urban governance, Sichuan Chengyi Expressway has reduced the accident rate by 60% through vehicle-road collaborative optimization using digital twins. Chongqing Water Affairs has achieved a 95% accuracy in water conservancy dispatch forecasting through remote sensing data and simulation deduction; China Southern Power Grid and China Meteorological Administration have used intelligent computing capabilities to improve the accuracy and stability of weather forecasts.

In addition, FeiTian Intelligent Computing also supports Alibaba's artificial intelligence practice, supporting Alibaba AI's 1 trillion calls per day and serving 1 billion people worldwide. Among them, the training speed of Pailitao has increased by 200 times, and the full training time of 1 billion pictures has been shortened from 2.5 months to 8 hours. The DAMO Academy's large model M6 only uses 512 GPUs and completes the training of a 10 trillion parameter model in 10 days, with an energy consumption of only 1% of GPT-3 with the same parameter scale.