In 2016, AMD reached an X86 licensing agreement with Tianjin Haiguang Advanced Investment Company in China. The latter spent US$293 million to obtain the most advanced Zen-based X86 processor technology at the time. The domestic Haiguang X86 processors have also been matured in the past two years. The highest also has 32 cores and 64 threads, which will be used in the domestic E-class tens of billions of times in the future.
According to reports from Japanese media, at the German ISC 2019 Supercomputing Conference in June this year, Sugon showed a supercomputing server named Nubla (Nebula), which uses Haiguang X86 processorAnd Haiguang DCU accelerator, the heat dissipation system is immersed liquid cooling, and the scale is huge.
The new generation of supercomputing Sugon E-class developed by Sugon is one of China's three exascale supercomputing. The prototype was developed last year. Japanese media cited a paper published by Professor Qian Depei, an expert in the overall group of China's national "high-performance computing" key research and development plan, and introduced the structure of Sugon's new-generation supercomputer prototype.
From the data point of view, Sugon’s E-level supercomputing takes the X86 route (Haiguang’s spending money to obtain AMD X86 authorization is obviously planned), which is also the mainstream of TOP500 supercomputingChoice, after all, X86 is the most compatible ecosystem.
Dawn’s supercomputer prototype is composed of 512 nodes, each node is equipped with 2 Haiguang X86 processors, 1 Haiguang DCU accelerator (specific information is unknown), 6D Tours interconnect structure, each node The bandwidth can reach 200Gbps, the peak performance is 3.18PFLOPS (1PFLOPS=1 petaflops), the Linpack performance is 2.274PFLOS, and the calculation efficiency is 71.5%.
For the new generation of supercomputing, in addition to the exascale performance, it canEfficiency is becoming more and more important, and the requirements for the cooling system will be higher. The supercomputer prototype shown by Sugon uses immersive liquid cooling. The entire cabinet will be immersed in a special cooling liquid. The material written on the data is Imm058 has a boiling point of only 50°C, which is also the temperature at which most semiconductors work, and can just complete the evaporation-condensation cycle.
Sugon’s supercomputer prototype is also equipped with a display screen, which can display the status of the system in real time, including system pressure, power, liquid temperature, water supply temperature, water supply pressure, liquid supply pressure and so on. However, the original text points out a question, that is, the liquid temperatureThe temperature is actually lower than the water supply temperature (normally it should be higher), I don’t know if it’s because of the display system.
The others are all internal components of liquid cooling. It can be seen that a large part of the future supercomputing space is reserved for the cooling system. The performance of the supercomputing is impossible without the heat dissipation. The total power consumption of the system is also greatly related to the heat dissipation system.