On October 20th, the 6th World Internet Conference was held in Wuzhen, Zhejiang, where guests from all over the world conducted a new round of Wuzhen Taoism. The climax of this conference was the announcement of 15 world-leading Internet scientific and technological achievements that shine around the world. These achievements are expected to give strong impetus to human life, scientific and technological progress, economic development and social change.
1. Huawei Technologies Co., Ltd. “Kunpeng 920”;
2. Tsinghua University “Heterogeneous Fusion Tianji Chip for General Artificial Intelligence”;
3. Microsoft’s “Unifying Natural Language Pre-training Models and Machine Reading Comprehension”;
4. “360 All-Seeing Eye – 0day Vulnerability Radar System” by 360 Security Technology Co., Ltd.;
5. Tesla’s “Tesla Fully Self-Driving Chip”;
6. Baidu’s “Flying Paddle”;
7. “POLARDB: Cloud Native Database Based on Storage Computing Separation and Distributed Shared Storage Architecture” by Alibaba Cloud Computing Co., Ltd.;
8. Cambrian company “Siyuan 270”;
9. Shenzhen Tencent Computer System Co., Ltd. “Science and technology for good – Helping modern smart city comprehensive management practice through scientific and technological means”;
10. Megvii Technology Co., Ltd. “Artificial Intelligence Algorithm Platform Brain++”;
11. Dawning Information Industry Co., Ltd. “Silicon Cube Immersion Liquid Cooled Computer”;
12. China Telecom Group Co., Ltd. “IPv6 ultra-large-scale deployment practice and technological innovation”;
13. SAP’s “AI Innovation and Deep Application of Smart Enterprises”;
14. Xilinx’s “Versal Adaptive Computing Acceleration Platform”;
15. Commercial Aircraft Corporation of China, Ltd. “Civil Aircraft Manufacturing 5G Innovative Demonstration Application”.
In the field of Chips, Huawei’s Kunpeng 920, Cambrian’s Siyuan 270, Tsinghua University’s Tianji chip, Tesla’s fully autonomous driving chip, and Xilinx’s Versal adaptive computing acceleration platform have become the most representative for chip development. product direction.
Huawei: Kunpeng 920
In January 2019, Huawei released the industry’s highest-performance Kunpeng 920 processor compatible with the ARM architecture. Through a series of micro-architecture designs such as optimizing the branch prediction algorithm, increasing the number of computing units, and improving the memory subsystem architecture, the single-core performance of the processor is greatly improved. Its innovative breakthroughs in high performance, high throughput, high integration, and high energy efficiency have pushed computing to new heights.
The Kunpeng 920 adopts a 7nm manufacturing process and is licensed based on the ARM architecture. Compared with mainstream processors in the industry, its high throughput is 60% higher, the memory bandwidth is 60% higher, and its energy efficiency ratio is 30% better than the industry benchmark. It can be used as a data center with lower power consumption. Provide stronger performance.
In large-scale data center business practice, every 10,000 servers based on Kunpeng processors can save 10 million kWh of electricity and reduce carbon emissions by 10,000 tons per year. In the cloud service scenario, HUAWEI Cloud Kunpeng cloud service has the natural advantages of being online and easy to obtain resources. It can help customers and partners to implement application migration and migration as quickly as possible, and accelerate business innovation and upgrade.
Not long ago, the Kunpeng 920 processor was officially put into use in Shenyang Aeronautical Industry Pneumatic Institute (hereinafter referred to as “Pneumatic Institute”). The pace is expected to be greatly accelerated, thereby greatly enhancing the independent innovation capability of my country’s existing aviation system.
Tsinghua University: Tianji Chip
On August 1, the cover article of the top academic journal “Nature” published the research results recently released by the Shi Luping team of Tsinghua University – the brain-like computing chip “Tianji”. This achievement has achieved zero breakthroughs in China’s “Nature” papers in the two fields of chips and artificial intelligence.
The industry’s first brain-like chip was TrueNorth, released by IBM in 2014. Then Tsinghua University released the first-generation “Tianji” chip in 2015, achieving zero breakthrough in domestic brain-like chips. In 2017, “Tianji” evolved into the second generation, with faster speed, higher performance and lower power consumption. Compared with the current world-leading IBM TrueNorth, it also has the advantages of more complete functions, flexibility and better expansion. 20% higher, at least 10 times faster, and at least 100 times higher bandwidth.
The latest generation of Tianji chips are manufactured using 28nm process, with a core area of only 3.8 × 3.8 mm, including 156 FCores cores, with about 40,000 neurons and 10 million synapses, which can support machine learning algorithms and brain-like at the same time. circuit.
It not only has high computing power, low power consumption, and supports a variety of different AI algorithms, but also adopts the integrated storage and computing technology, which does not require external DDR cache, which can greatly save space, power consumption and cost.
Tesla: Fully Self-Driving Chip
In April 2019, Tesla released a blockbuster weapon, the FSD chip (Tesla’s fully autonomous driving chip). Before the birth of the FSD chip, Tesla has always been a proponent, and it first cooperated with Mobileye , and then turned to Nvidia. In the middle, there were many ups and downs such as supply cuts and designer resignations, but fortunately, Tesla, who did not give up, finally spent 18 months to implement the FSD chip from an idea to the ground, and Tesla also took its fate in its own hands. Use this self-developed chip to prop up the world for Autopilot.
It consists of two identical independent systems, each with 12 A72 cores, a neural network processor and a GPU. The FSD chip is manufactured using the 14nm process, and the overall computing power of the fully autonomous driving computer equipped with the chip reaches 144TOPS (trillion times per second), which can make Tesla cars reach a new level of safety without affecting cost or mileage. performance and level of automation. At the same time, the Tesla chip is designed with an independent security chip that actively checks all instructions and data in an encrypted manner to monitor the possibility of hackers attacking self-driving cars. In March of this year, fully autonomous driving chips were officially loaded into Model S, Model X and Model 3.
Cambrian: Siyuan 270
Siyuan 270 was launched in June this year. Siyuan is the Chinese brand of the product, which means “the basic unit of thinking”.
The Siyuan 270 chip is manufactured using the TSMC 16nm process and integrates some innovative technologies of the Cambrian in the field of processor architecture, which can increase the theoretical peak performance by 4 times, and the theoretical peak performance of processing non-sparse deep learning models is improved to that of the previous generation MLU100. 4 times, reaching 128TOPS. In addition, the chip has also made some breakthroughs in the field of fixed-point training. The Siyuan 270 training board can provide excellent artificial intelligence training performance through 8-bit or 16-bit fixed-point computing.
Not only that, the Siyuan 270 chip is compatible with INT4 and INT16 operations, floating-point operations and mixed-precision operations. The chip uses its self-developed MLUv02 instruction set to support diverse artificial intelligence applications such as vision, speech, natural language processing, and traditional machine learning, and integrates ample video and image encoding and decoding hardware units for vision applications.
In terms of system software and tool chain, the chip not only supports the Cambrian Neuware software tool chain, but also supports various mainstream programming frameworks in the industry. Cambrian officials said that in order to facilitate developers to further explore the super computing power of Siyuan 270 and open up more application scenarios, it will open a dedicated programming language to developers and the community in the near future.
Xilinx: Versal Adaptive Compute Acceleration Platform
Xilinx’s Versal ACAP is a heterogeneous computing platform that fully supports software programming that combines a scalar engine, an adaptive engine, and an intelligent engine to achieve significant performance improvements, 20 times faster than the current fastest FPGAs Today’s fastest CPUs are 100 times faster and the platform targets data center, wired networking, 5G wireless and automotive driver assistance applications.
“The early shipment of the first Versal ACAP devices from TSMC to early access customers is another historic milestone and engineering achievement for Xilinx,” said Victor Peng, President and CEO of Xilinx. Xilinx’s years of software and hardware investment and the accumulation of all knowledge of product architecture over the past 35 years. Versal ACAP is a major technological breakthrough that will open a new era of heterogeneous computing acceleration for all applications and developers.”
On October 20th, the 6th World Internet Conference was held in Wuzhen, Zhejiang, where guests from all over the world conducted a new round of Wuzhen Taoism. The climax of this conference was the announcement of 15 world-leading Internet scientific and technological achievements that shine around the world. These achievements are expected to give strong impetus to human life, scientific and technological progress, economic development and social change.
1. Huawei Technologies Co., Ltd. “Kunpeng 920”;
2. Tsinghua University “Heterogeneous Fusion Tianji Chip for General Artificial Intelligence”;
3. Microsoft’s “Unifying Natural Language Pre-training Models and Machine Reading Comprehension”;
4. “360 All-Seeing Eye – 0day Vulnerability Radar System” by 360 Security Technology Co., Ltd.;
5. Tesla’s “Tesla Fully Self-Driving Chip”;
6. Baidu’s “Flying Paddle”;
7. “POLARDB: Cloud Native Database Based on Storage Computing Separation and Distributed Shared Storage Architecture” by Alibaba Cloud Computing Co., Ltd.;
8. Cambrian company “Siyuan 270”;
9. Shenzhen Tencent Computer System Co., Ltd. “Science and technology for good – Helping modern smart city comprehensive management practice through scientific and technological means”;
10. Megvii Technology Co., Ltd. “Artificial Intelligence Algorithm Platform Brain++”;
11. Dawning Information Industry Co., Ltd. “Silicon Cube Immersion Liquid Cooled Computer”;
12. China Telecom Group Co., Ltd. “IPv6 ultra-large-scale deployment practice and technological innovation”;
13. SAP’s “AI Innovation and Deep Application of Smart Enterprises”;
14. Xilinx’s “Versal Adaptive Computing Acceleration Platform”;
15. Commercial Aircraft Corporation of China, Ltd. “Civil Aircraft Manufacturing 5G Innovative Demonstration Application”.
In the field of chips, Huawei’s Kunpeng 920, Cambrian’s Siyuan 270, Tsinghua University’s Tianji chip, Tesla’s fully autonomous driving chip, and Xilinx’s Versal adaptive computing acceleration platform have become the most representative for chip development. product direction.
Huawei: Kunpeng 920
In January 2019, Huawei released the industry’s highest-performance Kunpeng 920 processor compatible with the ARM architecture. Through a series of micro-architecture designs such as optimizing the branch prediction algorithm, increasing the number of computing units, and improving the memory subsystem architecture, the single-core performance of the processor is greatly improved. Its innovative breakthroughs in high performance, high throughput, high integration, and high energy efficiency have pushed computing to new heights.
The Kunpeng 920 adopts a 7nm manufacturing process and is licensed based on the ARM architecture. Compared with mainstream processors in the industry, its high throughput is 60% higher, the memory bandwidth is 60% higher, and its energy efficiency ratio is 30% better than the industry benchmark. It can be used as a data center with lower power consumption. Provide stronger performance.
In large-scale data center business practice, every 10,000 servers based on Kunpeng processors can save 10 million kWh of electricity and reduce carbon emissions by 10,000 tons per year. In the cloud service scenario, HUAWEI Cloud Kunpeng cloud service has the natural advantages of being online and easy to obtain resources. It can help customers and partners to implement application migration and migration as quickly as possible, and accelerate business innovation and upgrade.
Not long ago, the Kunpeng 920 processor was officially put into use in Shenyang Aeronautical Industry Pneumatic Institute (hereinafter referred to as “Pneumatic Institute”). The pace is expected to be greatly accelerated, thereby greatly enhancing the independent innovation capability of my country’s existing aviation system.
Tsinghua University: Tianji Chip
On August 1, the cover article of the top academic journal “Nature” published the research results recently released by the Shi Luping team of Tsinghua University – the brain-like computing chip “Tianji”. This achievement has achieved zero breakthroughs in China’s “Nature” papers in the two fields of chips and artificial intelligence.
The industry’s first brain-like chip was TrueNorth, released by IBM in 2014. Then Tsinghua University released the first-generation “Tianji” chip in 2015, achieving zero breakthrough in domestic brain-like chips. In 2017, “Tianji” evolved into the second generation, with faster speed, higher performance and lower power consumption. Compared with the current world-leading IBM TrueNorth, it also has the advantages of more complete functions, flexibility and better expansion. 20% higher, at least 10 times faster, and at least 100 times higher bandwidth.
The latest generation of Tianji chips are manufactured using 28nm process, with a core area of only 3.8 × 3.8 mm, including 156 FCores cores, with about 40,000 neurons and 10 million synapses, which can support machine learning algorithms and brain-like at the same time. circuit.
It not only has high computing power, low power consumption, and supports a variety of different AI algorithms, but also adopts the integrated storage and computing technology, which does not require external DDR cache, which can greatly save space, power consumption and cost.
Tesla: Fully Self-Driving Chip
In April 2019, Tesla released a blockbuster weapon, the FSD chip (Tesla’s fully autonomous driving chip). Before the birth of the FSD chip, Tesla has always been a proponent, and it first cooperated with Mobileye , and then turned to Nvidia. In the middle, there were many ups and downs such as supply cuts and designer resignations, but fortunately, Tesla, who did not give up, finally spent 18 months to implement the FSD chip from an idea to the ground, and Tesla also took its fate in its own hands. Use this self-developed chip to prop up the world for Autopilot.
It consists of two identical independent systems, each with 12 A72 cores, a neural network processor and a GPU. The FSD chip is manufactured using the 14nm process, and the overall computing power of the fully autonomous driving computer equipped with the chip reaches 144TOPS (trillion times per second), which can make Tesla cars reach a new level of safety without affecting cost or mileage. performance and level of automation. At the same time, the Tesla chip is designed with an independent security chip that actively checks all instructions and data in an encrypted manner to monitor the possibility of hackers attacking self-driving cars. In March of this year, fully autonomous driving chips were officially loaded into Model S, Model X and Model 3.
Cambrian: Siyuan 270
Siyuan 270 was launched in June this year. Siyuan is the Chinese brand of the product, which means “the basic unit of thinking”.
The Siyuan 270 chip is manufactured using the TSMC 16nm process and integrates some innovative technologies of the Cambrian in the field of processor architecture, which can increase the theoretical peak performance by 4 times, and the theoretical peak performance of processing non-sparse deep learning models is improved to that of the previous generation MLU100. 4 times, reaching 128TOPS. In addition, the chip has also made some breakthroughs in the field of fixed-point training. The Siyuan 270 training board can provide excellent artificial intelligence training performance through 8-bit or 16-bit fixed-point computing.
Not only that, the Siyuan 270 chip is compatible with INT4 and INT16 operations, floating-point operations and mixed-precision operations. The chip uses its self-developed MLUv02 instruction set to support diverse artificial intelligence applications such as vision, speech, natural language processing, and traditional machine learning, and integrates ample video and image encoding and decoding hardware units for vision applications.
In terms of system software and tool chain, the chip not only supports the Cambrian Neuware software tool chain, but also supports various mainstream programming frameworks in the industry. Cambrian officials said that in order to facilitate developers to further explore the super computing power of Siyuan 270 and open up more application scenarios, it will open a dedicated programming language to developers and the community in the near future.
Xilinx: Versal Adaptive Compute Acceleration Platform
Xilinx’s Versal ACAP is a heterogeneous computing platform that fully supports software programming that combines a scalar engine, an adaptive engine, and an intelligent engine to achieve significant performance improvements, 20 times faster than the current fastest FPGAs Today’s fastest CPUs are 100 times faster and the platform targets data center, wired networking, 5G wireless and automotive driver assistance applications.
“The early shipment of the first Versal ACAP devices from TSMC to early access customers is another historic milestone and engineering achievement for Xilinx,” said Victor Peng, President and CEO of Xilinx. Xilinx’s years of software and hardware investment and the accumulation of all knowledge of product architecture over the past 35 years. Versal ACAP is a major technological breakthrough that will open a new era of heterogeneous computing acceleration for all applications and developers.”
The Links: DMF50081ZNB-FW PM50CSE120