Reports from the Heart of the Machine
Author: Zenan
3740 quadrillion floating-point operations per second.
When SenseTime went public at the end of last year, the ARTIFICIAL computing power center disclosed in its prospectus attracted people's attention: the company invested 5.6 billion yuan in the AI computing infrastructure built in Shanghai Lingang, and quickly announced the roof of the building in only 168 days last February, setting a new record for Lingang construction.
On the morning of January 24, SenseTime announced in Shanghai that its artificial intelligence computing center AIDC has started operations.
SenseTime Artificial Intelligence Computing Center (AIDC) is an open, large-scale, low-carbon, energy-efficient advanced computing infrastructure built by SenseTime, which started construction in July 2020.
As an important part of SenseCore's AI computing infrastructure, AIDC Phase I was designed to have a hash rate of 3740 Petaflops per second (1 Petaflops is a quadrillion floating-point operation per second), making it one of the largest supercomputing centers in Asia. AIDC covers an area of 130,000 square meters and has 5,000 cabinets in the first phase.
In the news of supercomputing, we often come across petaflops as a unit, which represents one quadrillion of a single-precision floating-point operation per second (10 to the 15th power). At present, the largest known computing power center of artificial intelligence in China is the "Pengcheng Cloud Brain II" of Pengcheng Laboratory, which outputs 1000 Petaflops computing power. SenseTime's HASH can process 23,600 years of video per day.
Different from the data infrastructure of traditional cloud services, SenseTime has chosen the direction of Artificial Intelligence Data Center on the road to helping the industry to make intelligent transformation.
Yang Fan, co-founder and vice president of SenseTime, said: "Compared with the traditional network infrastructure IDC, AIDC is like a brand franchise store, with users facing industrial intelligence, urban intelligence, AI for science, and services closer to the physical industry. AIDC is the foundation for the upgrading of the artificial intelligence industry."
On the basis of the network and hardware environment provided by IDC, SenseTime's AIDC focuses on artificial intelligence technology, which can provide more comprehensive service capabilities for the entire industry and help the physical industry to achieve intelligent upgrades.
Yang Fan, co-founder and vice president of SenseTime.
For chip providers, working with AIDC saves hardware adaptation, while for application vendors, using SenseTime's system can quickly land new technologies and save a lot of time.
In addition to the scope of application, SenseTime's infrastructure is also very technical: from the perspective of investment / computing power ratio, AIDC's efficiency is several times higher than the previous similar commercial AI hash rate. SenseTime said aidC will also help domestic hardware to be used on a larger scale during the construction process.
As a unicorn company with leading AI technology in China, SenseTime has been continuously iterating its own AI system since 2014. In the past seven years, the tens of thousands of industrial-grade algorithms proposed by SenseTime have promoted the continuous implementation of artificial intelligence technology in various industries and applications. SenseCore AI device proposed by SenseTime is a general artificial intelligence infrastructure that supports the company's products and services in smart business, smart city, smart life, smart cars and other fields.
Sense Core is the foundation of the AI business, and AIDC is the base of a large device. Since its inception, SenseTime has been deeply involved in the construction of computing power infrastructure, and has established 23 AI supercomputing clusters in major regional markets.
SenseTime's plan to build AIDC began in April 2018, when the company began laying out a prototype development project. Without any case experience and experimental site to refer to, SenseTime even sacrificed office space to build the core computer room of the prototype in the headquarters office building, and invested nearly 700 million yuan for project pre-research.
In 2019, SenseTime's ultra-large-scale intelligent computing center plan in Shanghai was put on the agenda, and in March 2020, the company signed a contract with Lingang, started construction in July, and passed the acceptance last year. After AIDC starts operations, it will further expand senseTime's AI supercomputing scale, so that SenseTime's final force exceeds 4910 Petaflops per second.
The construction of such a large-scale AI computing center in a short period of time is the first case in China. The successful launch of AIDC will further enhance SenseTime's AI-as-a-Service services and SenseTime's leading position in China. SenseTime hopes to promote the production of AI models and technology development across the industry to be more efficient and inclusive, so as to continuously expand the boundaries of AI industrialization.
SenseTime said AIDC can provide large-scale elastic computing power to achieve complete training of 1,000 billion-level parametric models. As a general-purpose AI infrastructure, SenseCore's AI device will rely on three parts: computing infrastructure, deep learning platform, and model layer to achieve batch algorithm model production, deployment, and iterative upgrade.
In addition to commercial use, SenseTime AIDC has also reached cooperation with some institutions on national basic scientific research projects. SenseTime said that the current AIDC computing power uses mainstream GPU devices, and it is expected that there will be an increase in localization rate in the future.
Use Python to quickly build an NVIDIA RIVA-based intelligent Q&A bot
NVIDIA Riva is an SDK that uses GPU acceleration to rapidly deploy high-performance conversational AI services for rapid development of speech AI applications. Riva is designed to provide easy, fast access to conversational AI capabilities, out of the box, and to quickly build high-level conversational AI services with a few simple commands and API operations.
January 26, 2022, 19:30-21:00, the latest issue of online sharing mainly introduces:
Introduction to Conversational AI and NVIDIA Riva
Build a speech recognition module with NVIDIA Riva
Build intelligent Q&A modules with NVIDIA Riva
Build speech synthesis modules with NVIDIA Riva