Ali cloud announced the launch of Feitian intelligent computing platform - the world's largest intelligent computing center
On August 30, 2022, Aliyun announced the official launch of its full-stack intelligent computing solution "Feitian Intelligent Computing Platform" and the launch of two ultra-large scale intelligent computing centers to provide powerful intelligent computing services for scientific research, public services and enterprise organizations, which can increase the utilization of computing resources by more than 3 times, improve the efficiency of AI training by 11 times, and improve the efficiency of reasoning by 6 times. The inference efficiency will be increased by 6 times.
--Ying-Hua Cai, President of Global Sales, AliCloud Intelligence
01Supporting the construction of super intelligent computing center
The Flying Smart Computing Platform originates from Alibaba's extensive internal practices and from extensive industrial intelligence practices. Currently, the platform is supporting the construction of two super-scale intelligent computing centers.
The Zhangbei Smart Computing Center is being built with 12 EFLOPS (120 billion floating point operations per second) AI computing power. In the cloud, Zhangbei Smart Computing Center supports AI pre-training large models, AI Earth, digital people and other cutting-edge AI applications, and will become the world's largest smart computing center.
The Wulanchabu Wise Computing Center, with a construction scale of 3 EFLOPS (30 billion floating point operations per second) AI computing power, is located in the hub node of Inner Mongolia of the "East Digital West Computing" and adopts green technologies such as natural air cooling and AI temperature regulation to achieve green and low-carbon.
02Fully open Feitian intelligent computing platform
Feitian Intelligent Computing Platform is the latest full-stack intelligent computing solution released by AliCloud, providing both public cloud and proprietary cloud models to provide powerful intelligent computing services for various research and intelligent enterprise organizations.
Through advanced technology architecture, Feitian Intelligent Computing Platform achieves 90% parallel computing efficiency of thousands of cards, which can increase the utilization of computing power resources by more than 3 times, AI training efficiency by 11 times, and inference efficiency by 6 times.
1) Breakthrough computing efficiency limit
With the advanced technology architecture, the core indicator of computing efficiency, "1,000-card parallel computing efficiency", is increased from 40% to 90%, which can improve the utilization of computing resources by more than 3 times, AI training efficiency by 11 times, and inference efficiency by 6 times.
2) Support "One Cloud, Many Cores"
Feitian Intelligent Computing Platform provides an integrated solution of converged arithmetic and big data AI platform based on AliCloud Panjiu infrastructure, which can run on servers with multiple chip types such as X86, GPU and ARM, realizing "one cloud with multiple cores", supporting mixed deployment and unified scheduling of multiple processors, and allowing application optimization. Part of the performance is improved by more than 100%.
3）Efficient intelligent service
Developers can carry out data storage, data governance, data analysis, model development, model training and inference on the Feitian intelligent computing platform. It also provides pre-trained models, as well as model capabilities in the fields of speech, image, natural language processing, and decision making, to facilitate developers to better accelerate the development of AI applications.
4) Green and low-carbon
Feitian Smart Computing reduces carbon emissions per unit of computing power in five areas: technology emission reduction, energy structure optimization, regional layout optimization, supply chain carbon reduction, and resource utilization optimization. In terms of technical emission reduction, the PUE can be as low as 1.09 by reducing energy consumption through liquid cooling, power supply technology and intelligent operation and maintenance.
Ali Cloud Immersion Liquid Cooling Cluster
03Systematic technological innovation
Intelligent computing is different from general-purpose computing, requiring massive amounts of data for training AI models, and arithmetic power is lost in data migration, synchronization and other links, and the minimum arithmetic output for scales above 1,000 cards is often only about 40%. This leads to the high cost of intelligent computing power, which restricts the development of the industry.
Ali cloud through the system of technological innovation, change the loss of intelligent computing problems, the thousand card parallel computing efficiency to more than 90%.
In network technology: AliCloud adopts high-performance RDMA network to achieve the lowest end-to-end 2-micron latency.
In communication technology: Ali cloud self-researched non-blocking communication technology, so that the data exchange speed in the computing process increased by more than 5 times.
In storage technology: Self-developed IO gas pedal allows data access to achieve up to 10x performance improvement.
In the big data AI development layer: AliCloud provides distributed training framework and automatic combination and tuning of distributed strategies through APIs, which improves training efficiency by more than 11 times.
One-stop AI computing development service: AliCloud provides users with one-stop universal inference optimization tools to quantize, prune, sparse, and distill algorithmic models, improving inference efficiency by more than 6x.
AliCloud Fetion Smart Computing Center
04Build the intelligent computing ecological industry circle
Today, the Feitian wise computing platform has been widely used within Ali, supporting the development of cutting-edge AI and electric intelligence technologies in Dharma Institute, and serving organizations and enterprises such as Xiaopeng Automobile, Shensi Technology, SAIC Group, China Meteorological Administration, Southern Power Grid, etc., supporting autonomous driving, new drug research and development, weather forecasting, industrial energy and other industries to significantly improve AI training efficiency.
1) In the field of autonomous driving
Xiaopeng Auto: Based on Feitian intelligent computing in Ulanqab, the construction of intelligent computing center "Fuyao", with an arithmetic scale of 600 PFLOPS, is the largest intelligent computing center for autonomous driving in China, speeding up the training of autonomous driving models by nearly 170 times.
Million-dollar car: based on Fei Tian Zhi Cai, it realizes 128 card parallelism efficiency of over 96%, which reduces the training cost of autonomous driving model by 62% and increases the training speed by 110%, so that the model iteration cycle is significantly shortened.
2) In life science field
After adopting FMSI platform, the cluster performance optimization is improved by more than 100%, and the training efficiency of molecular dynamics simulation is improved by more than 5 times.
3）In the industrial field
Jiji Auto: Using high performance computing to improve the efficiency of industrial simulation by 25% and intelligent driving training efficiency by 70%, accelerating the development of new models to market.
Shandong Texas Electric Power: The accuracy rate of review and prediction by AI reached 98%, and the time spent was shortened from one hour to a few minutes.
4) In the field of urban governance
Sichuan Chengyi Expressway: Vehicle-road collaboration optimization through digital twin reduced accident rate by 60%.
Chongqing Water: water dispatch prediction accuracy of up to 95% through remote sensing data and simulation projection.
South China Power Grid and China Meteorological Administration: using smart computing capabilities to improve the accuracy and stability of weather forecasting.
Take a look at
AliCloud Fetion Smart Computing Platform
