Cooling the world’s fastest supercomputer Fugaku a feat for operators in Japan
RIKEN Research Institute’s Fugaku supercomputer will be shown in a media unveiling on June 16, 2020 in Kobe’s Chuo District. (Mainichi / Ryoichi Mochizuki)
The RIKEN Research Institute’s Fugaku supercomputer, said to be the fastest in the world, generates an enormous amount of heat as its central processing units (CPUs) burn through calculations. While the supercomputer is expected to play a valuable role in cutting-edge research, a look behind its dazzling performance reveals a battle against the heat.
The building that houses Fugaku is in Kobe’s Chuo district, and the steam generated during the cooling process gushes from a cooling tower on the roof – almost like a power station. In fact, according to Riken, there is so much steam that there have even been several cases since last spring where people staying in a nearby hotel mistook it for smoke and alerted the police and fire department.
Toshiyuki Tsukamoto, Deputy Head of Department at the RIKEN Center of Computational Science, explained: “The heat generation density corresponds to that of a nuclear reactor. Depending on the type of calculation, the power consumption varies considerably and the heat value changes in tandem. Only supercomputers require measures to counter such massive changes in the Heat generation. ”
Fugaku, which was built as the successor to the supercomputer “K”, which went offline in August 2019, is scheduled to go into full operation in March of this year. In tests last year, it was used to simulate the spread of droplets when formulating measures against the coronavirus. It has taken first place in four categories in a biannual world ranking of supercomputer performance twice in a row.
In the computer room on the third floor of the building where Fugaku is housed, there are over 160,000 central units in 432 racks that generate heat during operation. The heat density can rise to over 100 kilowatts per square meter. This is like running 100 electric domestic heaters in an area of 1 square meter.
To ensure that the CPUs can work efficiently, they must be kept below 30 degrees Celsius. Without cooling, however, their temperature would rise above 100 ° C within seconds. To prevent this from happening, the supercomputer is equipped with a large water-based liquid cooling unit. The cooling system has primary and secondary branch pipes through which water flows to extract heat from the CPUs.
It is the secondary water system that directly cools the 160,000 or so CPUs. It brings water close to the CPUs at around 15 ° C to dissipate their heat, which in turn increases the water temperature to between 19 and 25 ° C. This water is then sent to heat exchangers. There the water in the secondary system is cooled down again to approx. 15 ° C by the water in the primary system, which is equipped with up to 11 cooling devices. After cooling down, the secondary system water is pumped back to the CPUs.
The reason the primary and secondary water systems are separated is to prevent contaminants from building up and blocking the thin pipes of the secondary system. The water in the secondary system is purified and contains a corrosion-resistant substance, while process water is pumped around the primary system.
The most difficult task, according to Tsukamoto, was dealing with temperature fluctuations. Different types of calculations lead to detailed changes in the operation of the CPUs, which leads to significant changes in power consumption and the amount of heat generated. This can lead to heat shocks in just 1/1000 of a second.
To remedy this, Tsukamoto and other researchers simulated the flow of water in intricate pipeline layouts to develop a system that could cope even in the event of an unexpected event. They installed sensors to monitor the water temperature in the secondary system and made it possible to adjust the water flow in the primary system in order to keep the temperature of the water in the secondary system constant at around 15 ° C.
“When the temperature suddenly rises and falls in a short period of time, the question of how to continuously cool the water becomes important. We haven’t had any cases in the past to relate to, and we thought about how existing technologies put together said Tsukamoto.
Tsukamoto was involved in Fugaku’s operation after working with K. “It’s only natural for it to work properly,” he said. “I do what is natural as if it were only natural. This is a source of pride for those who work behind the scenes and I think it’s also something that I am confident about.”
It is evident that such efforts to advance sophisticated cooling technology have helped bring Fugaku to the forefront of supercomputing.
(Japanese original by Mirai Nagira, Science and Environment News Division)