Over the past 5 years, Big Data and new technology applications have appeared one after another, so the most powerful supercomputers have started to increase the use of professional chips.
In April 2020, Mellanox, Israel’s leading data center networking technology company, was finally acquired by Nvidia for nearly $ 7 billion, opening two major industry events after more than six months. “Silent silence”. Shortly after, Eyal Waldman, founder / CEO of this company, announced his resignation and left a statement: “When this deal is signed, I know I’m leaving. I spent many years founding it. company and make all the decisions, not because of second place.
Just a few days ago, Mellanox was integrated into NVIDIA’s new network architecture product – the NVIDIA Mellanox 400G InfiniBand. Since then, it will be the first to join the most elite force in the data center market – the supercomputer market as a “GPU companion”.
Nvidia deployed the DGX system, including the A10080GB, at the Supercomputer Center in Cambridge, UK
NVIDIA Mellanox 400G InfiniBand product architecture, to put it simply, is a component used to connect other servers to the NVIDIA ecosystem. At the recent Global Supercomputing Conference, NVIDIA set a new performance record for the most powerful GPU product line – the A100 enterprise accelerator.
The new A100 GPU will increase high-bandwidth memory to 80 GB, which is double that of the previous generation. This means that more than 2TB of memory bandwidth per second will allow data to flow faster between memory and GPU, to “withstand” the pressure researchers are putting on building models and the larger AI dataset. . “In order to continue to exceed the upper limit of AI and high performance computing (HPC) research findings, scientists need to build larger and more complex models, so they need kit capacity. The memory is matching and the bandwidth is higher than before, ”said Bryan Catanzaro, vice president of NVIDIA.
In fact, in addition to changing the architecture of chip products and “connecting” hundreds of thousands of chips in their most efficient form, this microprocessor can also solve the enormous computing needs of scientists. This is the objective of the acquisition of Mellanox by NVIDIA. The role of the NVIDIA Mellanox 400G InfiniBand is to “connect” tens of thousands of processors, GPUs, and other chips into a supercomputer. While maximizing performance, the data transfer efficiency of each chip should not be lost too much.
Supercomputer market, “delicious piece of cake” cannot be ignored
In the list of Top500 Global Supercomputers published in June 2020, Oak Ridge National Laboratory’s supercomputer, ranked second in the United States, is equipped with two IBM Power9 processors and six NVIDIA Tesla V100s. There are 4,356 such nodes at a total cost of $ 200 million. In March 2019, the US Department of Energy’s Argonne lab publicly announced it would spend $ 500 million to build a new generation of Aurora supercomputers. This supercomputer, due for delivery in 2021, does not pursue the blind goal of computing speed, but requires new design ideas for the application of new technologies such as artificial intelligence.
So who are the main beneficiaries of this huge government contract? Apart from speculation, Intel, the largest processor manufacturer in the United States, and Cray, a famous integrator of supercomputer systems.
In fact, competition on supercomputers is still seen as a place of demonstration of technological power between countries. Although this is only one aspect, these supercomputers actually play an extremely important role in many military and scientific missions, such as weapon design and decoding and variable simulation. However, it is impossible to ignore the increasing use of supercomputer systems in the field of artificial intelligence research, the acceleration of GPU chips the figs have gradually become a staple for the construction of supercomputers. As a result, NVIDIA has also reaped significant benefits in many high-performance computing projects around the world, including the most powerful national laboratories in the United States – Oak Ridge Laboratory and Argonne Laboratory, which are the first of NVIDIA’s most successful products. more efficient.
There are many unprecedented chemical and hardware experiments that can be performed in the cloud, so it is especially important to deploy high performance servers. In this contest, China and the United States are of course the best candidates, and they are also the two technological powers ready to spend money on supercomputer systems.
In June 2020, the billion-dollar Fugaku supercomputer system at the Kobe Riken data center in Japan suddenly jumped to the top spot among Top500 supercomputers, but among these, China has 226 machines and 114 in the United States. United. Therefore, if the unit price of each supercomputer is hundreds of millions of dollars, it is a very coveted market. Governments have placed record orders time and time again, which is becoming more and more attractive to business enterprises.
In addition, there is no doubt that since supercomputers perform the most difficult technical research and development tasks, their existence is the source of the power of future technologies to enter the industrial market. and consumption.
The supercomputer war between nations has been going on for decades. Initially, most supercomputers used Intel and AMD microprocessors, derived from the latest generation of PC chips. However, over the past 5 years, Big Data and new technology applications have appeared one after another, so the most powerful supercomputers have started to increase the use of professional chips. But what is interesting is that the Japanese Fugaku has beaten the “power source” of the most powerful computers in China and America. As it turned out, it was thanks to Fujitsu’s 48-core A64FX SoC, and that it should be the first ARM-powered supercomputer system on the global supercomputer list.