High-performance computing faces a technological, economic and environmental dilemma

As more powerful computers increase their capacity to perform billions of operations per second, they also demand greater amounts of electricity comparable to those of small cities, which places high-performance computing in a technological, economic and environmental dilemma, in a context marked by the climate crisis.

consulted by The DayIsidoro Gitler Goldwain, from the Center for Research and Advanced Studies of the National Polytechnic Institute (Cinvestav), explained that “today, the performance of the most powerful computing systems exceeds exascale and the number of petascale systems continues to grow. Unfortunately, this growth is also accompanied by a constant increase in energy costs, which in turn implies a significant carbon footprint.”

As an example, he mentions that the energy expenditure of the supercomputer Captain (1.7 exaflops of computing power) number 1 on the list of the 500 most powerful supercomputers in the world is 29,685 kW (29.7 MW). “This consumption is equivalent to the energy expenditure of a population with approximately 25 thousand homes,” highlighted the academic.

Two decades ago energy efficiency became a concern in high-performance computing. “The cost of operating supercomputers exceeded that of acquiring and maintaining them. This highlighted the importance of improving the energy efficiency of future supercomputers.”

The TOP500, the list of the 500 most powerful supercomputers in the world, ordered by their computing power, has the ranking Green500, which not only measures power, but how much electrical energy a supercomputer consumes to produce that performance. This rating measures how many calculation operations they can perform for each watt of electrical energy.

In the November measurement, the most efficient supercomputer on this list of 500 is Kairos with approximately 73 Gigafloops/Watt. “Flops is a measure of floating point operations per second. Watts (watts) is a measure of power consumption. The relationship between the two gives the performance that a processing unit obtains per watt (watt).”

Gitler said that achieving greater energy efficiency “remains one of the most important challenges in the development of advanced computing and the primary constraint for the design of new computing technologies.”

Beyond the consumption for their operation, there is the challenge of cooling: in addition to the energy required for data processing, supercomputers demand intensive systems to dissipate the heat generated by the graphics processors (GPU), which concentrate high thermal loads.

In Mexico, this debate is beginning to be reflected in strategic projects such as Coatlicuethe new national supercomputer, whose construction site is still under analysis. The decision, experts say, must consider both the availability of water for its cooling system and an adequate electrical infrastructure.

The facility, they add, will have to be located in a region without water stress and with moderate temperatures, since operating in extreme climates would increase energy and water consumption, a dilemma that places the planning of scientific computing at the intersection between scientific policy, natural resources and environmental sustainability.

Green500

The Green500 ranks supercomputers not by their total power, but by their energy efficiency, measured in gigaflops per watt of electricity consumed.

1. Kairos (France) University of Toulouse–CNRS. Efficiency: 73.3 GFlops/W. Power: 3.05 petaflops. Consumption: 46 kW.

2. Romeo-2025 (France). Romeo HPC Center, Champagne-Ardenne. Efficiency: 70.9 GFlops/W. Power: 9.86 petaflops. Consumption: 160 kW.

3. Levante (GPU extension) Germany. German Climate Computing Center (DKRZ). Efficiency: 69.4 GFlops/W. Power: 6.75 petaflops

Consumus: 110 kW.

4. Isambard-AI (phase 1) United Kingdom. University of Bristol. Efficiency: 68.8 GFlops/W. Power: 7.42 petaflops. Consumption: 117 kW.

5. Otus (GPU only) Germany. University of Paderborn. Efficiency: 68.2 GFlops/W. Power: 4.66 petaflops.