At present, X86 and Power are the main architectures of over-computing CPU computing nodes, so who will be the biggest winners when Nvidia announces that CUDA supports Arm?Wen: Bao Yonggang
Original title: Is CUDA's support for Arm a new way to achieve mega-mega-scale overcomputing, or a good opportunity for Nvidia and Arm?
Supercomputer is usually referred to as overcalculation. It is often regarded as an indicator of a country's technological leadership in the scientific community, because it can provide computational support for the research of AI, aerodynamics, atmospheric science, energy science and other important technologies. Today, the competition between the powerful countries has entered the era of mega-mega-supercomputing, which requires more powerful processors. At present, X86 and Power are the main architectures of over-computing CPU computing nodes, so who will be the biggest winners when Nvidia announces that CUDA supports Arm?
Overcomputing Competition Enters the Million Mega Era
Last June, the U.S. Department of Energy announced the world's fastest supercomputer.
Overcalculation of Top 500 in 2019
Of course, China is also building a mega-mega supercomputer system, which is said to be based on three prototypes already built: Dawn, Tianhe and Shenwei. Japan and Europe are also reluctant to lag behind. Japan hopes to have a mega-mega supercomputer in 2021, while Europeans hope to achieve this goal in 2023. Obviously, the competition for supercomputers has entered the era of Exascale computing (mega-mega computing, also known as E-class supercomputing).
In an imprecise way to explain the millions of megabytes of computing, a millions of megabytes of computing in an instant is equivalent to four years of continuous computing every second by everyone on Earth. Such powerful computing power requires more complex systems. Like ordinary computers, overcalculation is also composed of hardware and software systems, but the hardware part of overcalculation is composed of high-speed computing system, high-speed interconnection communication network system, storage system, maintenance and monitoring system, power supply system, cooling system and structure assembly design.
Among them, the high-speed operation system is responsible for the scheduling of complex logic and the tasks with high parallelism, which can be calculated by isomorphism (pure CPU) or heterogeneous computing (CPU accelerator is composed of computing nodes).
A New Way to Achieve Million Mega-Level Overcalculation
According to the supercomputer list compiled by international organization TOP500, it is not difficult to find that IBM Power, Nvidia Volta/Tesla and Intel Xeon are obviously important components of the supercomputing nodes from the list of the top 500 supercomputers published at the ISC International Overcomputing Conference in 2019.
Overcalculation Green 500 announced at the ISC International Overcalculation Conference in 2019
However, with the further enhancement of computing power, the excessive heat generated will not only lead to more resource consumption, but also the design of cooling system faces greater challenges, so the performance without tile is also very important. TOP500 has also compiled a list of over-calculating Green500, which competes not for performance, but for performance per watt, so even if an over-calculating is at the bottom of the TOP500 list, it gets a good place in Green500.
According to the latest Green 500 ranking, 22 of the 25 most energy-efficient supercomputers in the world are supported by Nvidia.
One of the keys is to adopt a heterogeneous computing method, let the x86 or Power CPU and Nvidia GPU cooperate, and uninstall the heavy processing jobs onto the more energy-saving parallel processing CUDA GPU. However, in the CPU market, the Arm architecture can not be ignored, so can the Arm CPU play an advantage in the over-computing mega-level competition?
It's too early to draw a conclusion, but Invida is not going to miss this possible opportunity. On June 17, 2019, at the ISC International Overcomputing Conference, Nvidia announced that it would provide the Arm ecosystem with a full stack of AI and HPC software by the end of the year. The stack will accelerate over 600 HPC applications and all AI frameworks, including all Nvidia CUDA-X AI and HPC libraries, GPU-accelerated AI frameworks and software development tools, such as PGI compilation supporting OpenACC. Device and performance analyzer.
This means that after Nvidia's stack optimization is completed, Nvidia will provide acceleration for all major CPU architectures, including x86, Power and Arm.
In response to the new announcement, Huang Renxun, founder and chief executive of Nvidia, said:
Nvidia's Good Business
, Arm CPU Nvidia GPU is a new option from the point of view of multimillion-megabyte supercomputing builders, but from the perspective of Nvidia, CUDA's support for Arm is not a simple announcement, but an investment in resources. Giving Nvidia the incentive to make such an investment is the demand and input of countries and giants.
In the United States alone, the total investment in R&D of next-generation supercomputing technology will reach more than$430 million, Secretary of Energy Rick.
Nvidia's efforts in the supercomputing market are not just CUDA's support for Arm,Nvidia 's announcement of the 22nd largest supercomputer in the world at the ISC International Supercomputing Conference in 2019.
By contrast, other TOP500 supercomputer systems with the same performance need to be built by thousands of servers, while DGX SuperPOD occupies less space and is about 400 times smaller than the same system. For deployment, other systems of the same size usually take 6-9 months to complete deployment, and DGX SuperPOD takes only three weeks for engineers to adopt a validated normative approach.
According to Leifeng. com, Nvidia DGX system has served many enterprises that need large-scale computing, such as BMW, Continental, Ford and Zenuity, Facebook, Microsoft and Fuji Film, Japan Institute of Physical Chemistry and Laboratory of the U.S. Department of Energy.
Nvidia hopes that the Nvidia SuperPOD architecture will be used by enterprises that have not yet deployed AI data centers. This can benefit both sides, and Nvidia can learn how to design systems for large-scale AI machines by building such supercomputers.
Obviously, overruns, especially megabytes, are a win-win business for Nvidia.
Arm's Good Opportunities
For Arm, the mega-super market is a good opportunity. When Arm, the leading mobile market, encounters a slowdown in market growth, it also hopes to expand its architecture to new markets to bring growth. In recent years, it has been working hard to promote the development of Arm servers with joint partners, but the situation is not ideal.
But after the super-computing market, especially CUDA's support for Arm, Arm was able to have a good opportunity. Pierre Barnab, Senior Executive Vice President of CUDA, big data and head of Cyber Security
Peter Ungaro, president and chief executive of Cray, a leading global provider of supercomputing, said:
In addition, Ampere Computing, CSC, EPI, HPE, J
Summary of Leifeng Network
At this time, Invida CUDA supports Arm to provide a new choice for countries that are building a new generation of over-calculating and enterprises that have higher demand for computing power. This is not only a good business that Invida can win more, but also a good opportunity for Arm to enter the over-calculating market. There are many supporters who have expressed their support for this new approach. Although we have maintained a positive attitude towards it, the results still need to be obtained through the actual mega-supercomputing system. Innovation will not be 100% successful, but innovation is the greatest driving force for innovation.