Shutting Down Servers When Power Is Dirty Could Cut Computing Emissions, Simulation Finds

High-performance computing clusters could shave up to 8 percent off their carbon emissions by simply switching themselves off whenever the German electricity mix turns dirty, according to a new simulation study published in EPJ Research Infrastructures. The catch is steep: the trick only works if the hardware idles quietly, and it barely saves any money at all.

The idea behind the study is sector coupling, a concept that is becoming central to Germany’s energy transition. As wind and solar dominate electricity production, the grid experiences sharp swings between oversupply and scarcity. On the German spot market, wholesale electricity prices were negative for a total of 573 hours in 2025, mostly at weekends and on public holidays when renewable generation outstripped demand. A computing cluster that can shift its workload into those windows absorbs surplus renewable energy, stabilises the grid, and in principle lowers both its environmental footprint and its operating costs, since electricity prices track the residual load of the energy system. Science is an ideal testing ground because short-term delays in computing results are often tolerable, provided the overall computing target stays constant over long periods.

The researchers built a simplified simulation of cluster operation using publicly available data from the Fraunhofer Institute for Solar Energy Systems, which records German net electricity generation broken down by source together with spot market prices in 15-minute intervals. Combining generation data with published emission factors for each generation type, they calculated the carbon intensity of German electricity, in kilograms of CO2 per megawatt-hour, for every interval of the calendar year 2024. The simulated cluster was assumed to buy all of its electricity on the spot market. Whenever the carbon intensity of generation exceeded a threshold, the cluster would idle; whenever it fell below, the cluster would run. The threshold was tuned numerically to find the minimum of total emissions.

One crucial constraint shapes the whole analysis: the cluster must deliver the same total computing over the year as a continuously running machine. A cluster that runs only 50 percent of the time must therefore be twice as large, assuming linear scaling, which means more hardware, more embedded emissions from manufacturing and transport, and more idle power draw. The team modelled this by scaling the number of logical cores, and they estimated embedded emissions per core from a life cycle assessment of a Dell R740 rack server, which lists 4092 kilograms of CO2 per server, dominated by SSD production. Assuming 256 logical cores per server and a ten-year life cycle, they arrived at roughly 0.00018 kilograms of CO2 per core-hour. For a default Bonn setup that relies on hard disks instead, the figure was about ten times lower.

Five simplified cluster configurations were tested, representing hardware at the University of Bonn’s Analysis Facility, the DEEP modular supercomputer prototype at Jülich, and ARM-based worker nodes at the GridKa Tier 1 centre in Karlsruhe, each under three workload scenarios including a backfilling mode typical of high-energy physics, in which short single-core jobs keep the cluster fully loaded. The result hinged on one ratio: idle power divided by maximum power. The best performer, a modern Bonn configuration whose idle draw is only 15 percent of its peak, achieved an 8 percent emission reduction at an optimal utilisation of 0.635 in the backfilling scenario. By contrast, a Jülich configuration whose idle draw is 48 percent of peak offered no savings at all; its optimum collapsed to continuous operation. Beyond an idle-to-peak ratio of roughly 0.4, dynamic operation no longer pays off for any setup.

The sensitivity analysis produced the study’s most striking number. If a cluster could be shut down completely rather than idled, so that idle power fell to zero, simulated savings exceeded 50 percent depending on the setup. The authors note that complete shutdown was deliberately excluded from their main scenarios, because rapid restart under favourable conditions and hardware protection argue against it, but the figure illustrates how much headroom idle power consumption leaves on the table. They also tested the uncertainty in embedded emissions by varying the per-core estimate up and down by 50 percent. Even with the higher value, the optimum remained below continuous operation, with an emission reduction of about 5 percent, indicating that hardware manufacturing emissions are not the dominant factor in the optimisation.

Because the German grid will keep getting cleaner, the team asked whether dynamic operation would still help in a renewable-rich future. Reducing the carbon intensity of generated electricity by up to 90 percent in the simulation, they found that the optimal utilisation rises and the relative savings shrink, but dynamic operation still yields a reduction even when electricity emissions fall to 20 percent of their 2024 value. A validation exercise tested whether a threshold derived from 2024 data would hold up in other years. Extrapolated to 2023 and 2025 while accounting for the changing share of renewables, the threshold kept the compute target within 2 percent of its goal, and the thresholds individually derived for each validation year deviated by only 1.42 and 0.68 percent.

The financial picture was far less encouraging. Cost optimisation, carried out only for the two Bonn configurations, assumed spot-market purchase of all electricity, a demand charge of 100 euros per kilowatt per year, and acquisition costs per core-hour derived from the actual price of Bonn hardware, counting only CPUs and memory. Across all tested scenarios, relative savings stayed below 1 percent, because the dominant cost term is hardware acquisition, which grows as utilisation falls and more cores are needed to hold the compute target constant. Even when acquisition costs were cut by 50 percent in a further simulation, savings remained below 2 percent. Lower electricity prices erode the benefit further: only a price drop of more than 85 percent would push the optimum back to continuous operation. The authors conclude that dynamic operation can achieve only quite limited cost savings, even under favourable assumptions.

One comparison suggests an alternative lever. For the Karlsruhe ARM setup, the team drew on a prior study that scanned CPU clock frequencies for maximum power efficiency, finding a setting that cuts power consumption by about 40 percent while sacrificing 19 percent of benchmark performance. Compensating that performance loss with additional hardware, continuous operation at the limited frequency reduced emissions by almost 20 percent in the backfilling scenario, outperforming dynamic operation regardless of the idle power ratio, although in the medium workload scenario the extra hardware wiped out the benefit. The authors suggest that combining a frequency limit with dynamic operation could be a viable future option, though simulating that combination was beyond the scope of this study.

The authors are careful to frame the work as a proof of concept rather than a deployment plan. The simulation assumed instantaneous, unlimited switching between idle and running states, homogeneous hardware, and a simple scaling by core count; in practice, jobs must be pausable and resumable, ramp-up and ramp-down times will shift the optima, and heterogeneous clusters with accelerators are harder to model. In the best-performing scenario the cluster had to pause and resume roughly 230 times over the year, with an average interval between switches of about 20.7 hours. The stakes are considerable: data centres already account for 4 percent of Germany’s gross electricity consumption, and forecasts suggest computing could reach 8 percent of total power consumption by 2030. With the carbon intensity of the German electricity mix averaging 344 kilograms of CO2 per megawatt-hour in 2025 and renewables covering 55.1 percent of electricity consumption but only 23.8 percent of final energy use, flexible consumers that soak up surplus wind and solar power are likely to become an increasingly valuable part of the energy system. For research computing, the message of this study is that the environmental case for grid-aware operation is real but conditional, resting on low idle power, tolerant workloads, and a willingness to accept that the financial rewards, for now, are marginal.

Subject of Research: Simulation of dynamically operated computing clusters in Germany to reduce carbon emissions and operating costs through sector coupling

Article Title: Economical and Ecological Impact of Sector Coupling Applied to Computing Clusters

Article References: Bechtle, P., Freyermuth, O., Geffers, M. P., Giffels, M., Hübner, M., Kirfel, F., Kreutz, J., Krieg, S., Matberg, S., & Schnepf, M. (2026). Economical and Ecological Impact of Sector Coupling Applied to Computing Clusters. EPJ Research Infrastructures, 10(1), Article 29. https://doi.org/10.1007/s41781-026-00183-6

Image Credits: AI Generated

DOI: 10.1007/s41781-026-00183-6

Keywords: sector coupling, high-performance computing, carbon emissions, renewable energy, spot market electricity, data centres, energy transition, idle power consumption, embedded emissions, Germany, computing clusters, simulation study