In the TOP500 supercomputing released today, the supercomputing frontier based on AMD processor and encryption card won the championship, which is also the first publicly confirmed 10 billion times of supercomputing in the United States and the world, with a peak performance of more than 11 billion times. The last time amd won the TOP500, it was Titan supercomputing in 2012. At that time, it used a 16 core Haolong processor, but the accelerator card was NVIDIA's k20x. Now the frontier is the highlight moment of AMD processor and mi250x accelerator card. 10 years have not been wasted.
Frontier uses AMD's third-generation epyc processor, but it is not a common Milan series. Instead, it is a 7a53 code named "Trento". It is also a 64 core 128 thread, but supports XGMI bus. It is a product customized for Supercomputing**
1-way 64 core processor with 4-way mi250x accelerator card, connected through infinity fabric, 4-way NIC high-speed network and 4-way PCIe 4.0 nvme SSD hard disk, which together form a node.
The final frontier supercomputing consists of 74 cabinets, 9408 nodes in total, 8.73 million cores (cpu+gpu), 9.2pb DDR4 memory, 37Pb local storage and 716pb central storage
Despite its large scale, frontier has very high energy efficiency. It only uses 40mW of power consumption to achieve 10 billion times of supercomputing. In order to save energy, AMD and HPE have also wasted a lot of effort. The 64 core epyc processor allows only 2GHz, so it does not need high frequency.
At the same time, the traditional air cooling is not enough. This time, frontier uses 100% liquid cooling, which is also one of the key factors for higher energy efficiency
