Green500: Adventure in Qualifying a Cluster

Green500: Adventure in Qualifying a Cluster Nicolas Richart1[0000−0002−1463−4405] , Julia Paolini2[0009−0004−1283−9943] , Xavier Ouvrard2[0000−0003−4054−2771] , Manuel Cubero-Castan3[0009−0005−3152−6915] , and Gilles Fourestey1[0000−0001−7217−2977] SCITAS, École Polytechnique Fédérale de Lausanne (EPFL), 1015 Lausanne, Switzerland nicolas.richart@epfl.ch 2 EcoCloud, École Polytechnique Fédérale de Lausanne (EPFL), 1015 Lausanne, Switzerland 3 Sustainability, École Polytechnique Fédérale de Lausanne (EPFL), 1015 Lausanne, Switzerland

arXiv:2611.09251v1 [cs.DC] 6 Nov 2026

1

Abstract. Energy efficiency has become a key metric for evaluating High Performance Computing (HPC) systems, as reflected by the Green500 ranking. However, energy efficiency alone does not capture the full environmental impact of an HPC cluster. Using the Kuma cluster at EPFL, we report on two Green500 measurement campaigns at Level 2 and Level 3, and extend this analysis to a full carbon footprint assessment using the MITSI life-cycle based methodology. GPU power capping experiments show that energy efficiency peaks at 450 W, achieving an 11% performance loss offset by a 20% reduction in energy consumption. When embodied emissions are accounted for, the optimal power cap for carbon efficiency shifts to 500 W under Switzerland’s low-carbon electrical grid, revealing a tension between minimizing usage and embodied contributions. We discuss the challenges of generalizing this approach into a broader carbon-based HPC ranking, and argue that moving beyond energy efficiency is both necessary and feasible. Keywords: Green500 · HPC · energy efficiency · carbon efficiency· carbon footprint · GPU power capping · life-cycle assessment

1

State of the Art

High Performance Computing (HPC) systems have evolved through successive architectural paradigms, from vector supercomputers to massively parallel processing systems, commodity clusters, and heterogeneous accelerator-based nodes. Comparing these infrastructures became key, and the introduction of the LINPACK benchmark [8] enabled the TOP500 list [24], first published in 1993, to provide a standardized mechanism to track long-term progress in HPC. TOP500, which ranks the first 500 HPC systems, privileges measured rather than peak performance, highlighting the importance of scalable infrastructures, efficient interconnects and optimized numerical libraries. While HPL—the implementation of LINPACK—allows fine tuning of parameters across a wide range of system sizes [22], it focuses solely on floating-point

2

N. Richart et al.

performance and cannot evaluate other aspects such as bandwidth or memory access. Additionally, it focuses on performance-driven clusters rather than sustainability-driven ones. As early as 2007, several actors were starting to push the idea that another ranking was needed in addition to the TOP500. In [11], the authors were already mentioning how performance-at-any-power-cost in HPC was not sustainable, and put the foundations of the Green500. In the Green500 [25], the HPC clusters that the TOP500 ranks solely on performance are instead ranked by energy efficiency, measured in Gflop/s/W. The first Green500 ranking was published in 2013. Since then the efficiency of HPC systems has doubled every two years [2]. While alternative benchmarks such as HPL-MxP [1, 7], HPL-AI [16], HPCG [13], and STREAM [18] have emerged to address HPL’s limitations, we focus solely here on Green500, using HPL on the Kuma cluster [10]. Focusing on energy efficiency should be seen as a first step toward sustainability. Assessing the full environmental impact of HPC clusters requires a more holistic approach [15], and in particular evaluating the carbon footprint of such installations [17]. In particular, the whole life-cycle of the equipment has to be taken into account, including the embodied emissions (those arising from manufacturing, transportation, and disposal) [12]. We base our evaluation on a new methodology that has been developed at EPFL, a collaboration between the Sustainability department, the Scientific IT and Application Support (SCITAS) platform and the EcoCloud research center.

2

Green500

2.1

Experimental Set-up

The Green500 power measurement methodology [23] defines three quality levels across different aspects, granularity, machine fraction, included subsystems, and instrumentation sampling rate, accuracy and location. The measured system The experiments described in this paper were conducted on the Kuma cluster [10]. Kuma is composed of 84 compute nodes, each equipped with 4 NVIDIA H100 SXM5 GPUs (94 GB HBM2e per GPU, interconnected via NVLink at 900 GB/s), 64 CPU cores, and 384 GB of DDR5 RAM. Each node is connected to the network fabric through two 200 Gbit/s InfiniBand HDR links. The nodes are distributed across 10 racks; 9 racks of compute nodes and 1 administration rack with the front and administration nodes and the shared storage. There is also an extension with 20 similar nodes equipped with 8 NVIDIA L40s GPUs each instead of the H100. The GPUs are direct-to-chip liquid cooled while the CPUs are air cooled, with the GPU heat transferred to the node’s air cooling loop; active rear-door heat exchangers use fans to help extract warm air from the racks. The recovered heat is fed into EPFL’s central heating system (using on heat pumps) during

Green500: Adventure in Qualifying a Cluster

3

winter; during summer water pumped from Lake Geneva is used to cool the rear-door cooling loop. For the Green500 submission, 80 out of the 84 nodes were used, as 4 nodes were unavailable at the time of the measurements; the L40s nodes were turned off for all the measurements. HPL was configured with a problem size N = 1,943,040, a block size N B = 768, and a process grid of 16 × 20 (P × Q) for the 80 nodes runs. Single-node validation runs used N = 217,088, N B = 1,024, and a 2 × 2 process grid. Level 2 – List of November 2024 For the Level 2 submission, power was measured using the Power Distribution Units (PDUs) installed in each rack. The PDU power readings were collected via SNMP and exported to a Prometheus time-series database using snmp_exporter, with Grafana used for visualization and data export. The PDU readings provide per-rack active power at 1-second intervals, based on sampling at 4 kHz, covering all 10 racks including the compute nodes, the network, the administration and front nodes, and the shared storage. Three GPU power capping configurations were tested: 350 W, 500 W, and 700 W. For each configuration, HPL was run on the full 80-node partition and power was recorded during both the core computation phase (HPL time) and the full run (full SLURM job time). The nodes used NVIDIA driver version 535.154.05. Power cap 700 W 500 W 350 W

Perf. Core Energy Core Avg. Power Energy Efficiency (Pflop/s) (kWh) (kW) (Gflop/s/W) 11.95 31.2 274.4 43.555 11.66 24.8 212.2 54.959 8.85 35.6 167.0 52.995 Table 1. Level 2 results (PDU-based measurements)

Level 3 – List of November 2025 The Level 3 requires energy-based measurements of the entire system at a sample rate of 5 kHz. For this, power meters were installed on the electrical distribution rails on the top of the racks. The compute nodes are spread over 4 power rails (Rail 1 to Rail 4), each line having 3 phases, two of which are connected through Uninterruptible Power Supplies (UPS). Different types of power meters were used due to local availability: two A. Eberle PQ-Box 50 on Rail 1 and Rail 2, one A. Eberle PQ-Box 150 on Rail 3, and one Fluke 1738 on Rail 4. These instruments deliver averaged measures at 1 s intervals and capture both instantaneous power and cumulative energy. The PDU-based monitoring from Level 2 was maintained in parallel, allowing a direct comparison between the two measurement approaches. Before the measurement campaign for the Level 3 submission, the cluster was updated to RedHat 9.4 and the NVIDIA driver to version 560.35.03-1.

4

N. Richart et al.

2.2

Results

Level 2 Results. Table 1 summarizes the Level 2 results. The 500 W configuration, which achieved the best energy efficiency, was used for the Green500 submission, while the 700 W configuration, with the highest absolute performance, was used for the Top500 submission.

30

200

20

100

10

0

Energy [kW h]

Power [kW]

300

0 0

100

200

300

400

500

Time from start of job [s] Fig. 1. Power measured (thin lines) from PDUs and corresponding energies (thick lines) for 3 values of GPU power capping, 350 1 W (green), 500 W (blue), 700 W (orange). Shaded regions correspond to the core part of HPL.

Since the Rmax values were not as high as expected, a campaign of node characterization was launched. The main results obtained from this campaign show to what extent single nodes sustain HPL runs. Node Variability and Thermal Throttling Prior to the Level 3 campaign, HPL was run 40 times on each node, once allowing the GPU to cooldown between runs and once without. Several nodes exhibited significant performance degradation over successive runs without cooldown (Fig. 2, left). While healthy nodes sustained approximately 189 Tflop/s consistently, the affected nodes showed a significant performance drop after 2 to 3 runs. This is indicative of thermal throttling under sustained GPU load. The fact that not all nodes exhibit the same behavior was a sign of a potential hardware issue. Investigation with the hardware vendor identified defective thermal pads on the H100 GPUs of the affected nodes, which prevented adequate heat dissipation. After replacement of the thermal pads by the vendor, the single-node tests were repeated (Fig. 2, right). The post-correction results confirmed that the affected nodes recovered to nominal performance levels, except for one node that developed issues only during the second campaign. Level 3 Results. Table 2 presents the Level 3 results submitted for the Top500 and Green500 lists. The average power values are higher than those from the

Performance [Tflop/s]

Green500: Adventure in Qualifying a Cluster kh005 kh011 kh020

kh033 kh036 kh051

kh056 kh060 kh068

5

kh078 kh080 kh081

190 180 170 160 0

20

40

60

80

0

10

20

30

40

Fig. 2. Consecutive runs of HPL, before (left) and after (right) vendor support intervention on the GPUs. The red thicker line 1corresponds to thermal throttling, the blue lines correspond to the nodes that behave properly Table 2. Level 3 results (August 2025, 80 nodes). PM = power meters. Power cap 700 W 450 W

Perf. Core Energy Core Avg. Power Energy Efficiency (Pflop/s) (kWh) (kW) (Gflop/s/W) 12.31 32.3 291.9 42.169 10.99 25.6 207.4 52.979

Level 2 PDUs for three reasons: more subsystems are connected to the power rails (cooling fans and L40s extension network); integration errors differ between PDU sampling at 1 Hz and power-meter energy integrated from data measured at above 5 kHz; and PDU noise from Grafana aggregation, visible in Fig. 1 and confirmed by comparison with the power meter data in Fig. 3. It should be noted that the power meters had no network access to an NTP server and required manual clock synchronization via a laptop. Despite this, clock offsets between meters were corrected in post-processing by correlating the power line frequency, which is identical across all rails. Residual offsets were small: both PQ-50 meters were in sync, the PQ-150 was 2 s behind and the Fluke 1738 was 1.88 s behind.

Energy Efficiency Sweet Spot The Level 2 campaign tested only three power capping values. During the Level 3 campaign, this was extended to all 50 W increments from 200 W to 700 W, with additional 25 W steps between 400 W and 575 W to refine the sweep. Fig. 4 shows the resulting energy efficiency curve, together with Level 2-equivalent PDU values and a direct Prometheus measurement for comparison. The optimal energy efficiency is obtained for a capping of 450 W with a loss of 11% of performance but a reduction of 20% in energy consumption; this configuration was selected for the Level 3 Green500 submission. These results

300

30

200

20

100

10

0

0 0

100

200

300

400

Energy [kWh]

N. Richart et al.

Power [kW]

6

500

Time [s] Fig. 3. Energy measured from power meters and corresponding power for 2 values of 1 W (blue). Shaded regions correspond to GPU power capping, 450 W (orange) and 700 the core part of HPL.

confirm what was already observed for example in [14], that energy savings can be achieved at low performance losses through GPU power capping.

3

Moving a step forward: Carbon footprint

The energy efficiency metric requested by the Green500 ranking offers a complementary ranking to the performance-based TOP500 ranking. However, looking solely at energy optimization obscures the overall environmental impact generated by HPC clusters. This section estimates the carbon footprint of Kuma, acknowledging that a full assessment would require a multi-criteria life-cycle analysis beyond CO2 alone. This study follows MITSI (Methodology for IT Services environmental Impact assessment) [19]. This methodology follows the Life Cycle Analysis framework, estimating the carbon footprint on four stages: manufacturing (including abiotic resources extraction), distribution, usage and end-of-life. We encompass the manufacturing, distribution and end-of-life phases in what is commonly named embodied emissions. 3.1

Usage emissions

The electrical carbon footprint of data centers varies according to the time of day, season, and geographic scope. For this study, we use a yearly national average, which smooths out these variations. Such data are available from platforms such as Electricity Maps [9]; for Switzerland, we use the values provided by the Swiss association of electricity providers (AES) [4]: 90 g CO2 eq/kWh for 2024 and 57 g CO2 eq/kWh for 2025.

Green500: Adventure in Qualifying a Cluster

Energy eciency [Gflop/s/W]

Power meters (Energy)

7

PDUs (Integrated power Grafana)

PDUs (Energy Prometeus)

60

50

40

30 200

300

400

500

Power cap [W]

600

700

Fig. 4. Energy efficiency for different GPU power capping.

1

As recommended by MITSI, the datacenter infrastructure electrical consumption (cooling, lights, UPS overheads, etc.) is added to the usage emissions, estimated using the Power Usage Effectiveness (PUE). The EPFL datacenter PUE is estimated at 1.3, meaning 0.3 kWh of datacenter infrastructure overheads per 1 kWh of IT load. 3.2

Embodied emissions

The embodied emissions cover the Kuma configuration used for the Green500 submission (Section 2.1): the 80 H100 nodes, administration servers, InfiniBand network, storage, and spare parts. To get emission factors for each element composing the Kuma cluster, the Resilio database [5] was used. This database offers one of the most up-to-date environmental database for ICT equipment and services, and has the advantage to be freely available for academic users. Following MITSI with Resilio data, we estimate a total footprint for the embodied emissions at 387.2 t CO2 eq. To add the embodied emissions to usage emissions, this total footprint was allocated based on the total lifetime of the cluster, which has been committed to 6 years with a 100% usage. This gives us an embodied cost per second of 2.1 g CO2 eq/s. For Kuma, the embodied emission per GPU-hour is 23.0 g CO2 eq/hGPU. This is significant: the commonly used GPU-hour metric encompasses only the usage phase, which, depending on the electrical carbon footprint, can vary between approximately 40 g CO2 eq/hGPU and 500 g CO2 eq/hGPU, meaning embodied emissions represent between 10% and 60% of usage emissions. The embodied emissions of the datacenter building and cooling infrastructure were not included, as these are a minor contribution to total datacenter footprint [3, 6].

8

N. Richart et al.

3.3

Results

Carbon Eciency Embodied Contribution

Usage Contribution

2

Carbon Eciency [Pflop/g CO eq]

The methodology was applied to the Level 3 measurements to assess the carbon emissions of the Kuma cluster, both for embodied and usage emissions. The results are shown in Fig. 5.

0.16

6 0.14

2

Carbon rate [g CO eq/s]

8

4 0.12

2 0.10

0 200

300

400

500

Power cap [W]

600

700

Fig. 5. Carbon efficiency for different GPU power capping for the Level 3 measures.

1

As expected, the embodied contribution is constant, as it is modeled as a fixed carbon rate independent of the power cap. On the other hand, the usage contribution scales roughly linearly with the power cap, as higher GPU power directly increases the average power draw. This contribution is therefore low at 200 W and high at 700 W. Ceff =

P erf P erf = Cembodied + Cusage Cembodied + (P̄ · CI · P U E)

(1)

We now analyze the carbon efficiency Ceff as defined in Eq.1, where P erf is the performance in flop/s, Cembodied and Cusage the embodied and usage carbon rates, in g CO2 eq/s, P̄ the average power per run in W, CI the carbon intensity in g CO2 eq/kWh. Taking as electricity carbon footprint 57 g CO2 eq/kWh—2025 Switzerland value—and as embodied emissions 2.1 g CO2 eq/s—since our cluster lifetime is committed to 6 years—a similar behavior is observed between the carbon efficiency and the energy efficiency when power capping, but with a maximum of carbon efficiency observed at a power capping of 500 W instead of 450 W for energy efficiency. Indeed, in such condition (low carbon electrical footprint and 6 years cluster lifetime), the embodied emissions have an important contribution: to reduce this contribution, the cluster lifetime should be extended as much as possible, whereas to reduce usage contribution and maximizing energy efficiency, it should run as efficiently as possible.

Green500: Adventure in Qualifying a Cluster

2

Eciency [Pflop/g CO eq]

Baseline Scenario 1

9

Scenario 2 Scenario 3

0.15

0.10

0.05

200

300

400

500

Power cap [W]

600

700

Fig. 6. Carbon efficiency for the 4 scenarios (Section 3.3).

1

We now consider three alternative scenarios varying the energy mix and cluster lifetime. Scenario 1 reflects our 2024 measurement conditions (90 g CO2 eq/kWh, 6 years). Scenario 2 places the cluster in a region with a higher-carbon electrical mix (400 g CO2 eq/kWh, 6 years). Scenario 3 reduces the cluster lifetime to 4 years, raising the embodied cost per second to 3.1 g CO2 eq/s (57 g CO2 eq/kWh, 4 years). The four scenarios (baseline, 1, 2, 3) are shown in Fig. 6. Whether we consider Swiss electrical mix on 2024 or 2025, or we reduce the cluster lifetime to 4 years (adding then a higher importance to embodied emissions), the power capping that maximizes the carbon efficiency stays at 500 W. It is only when considering a high energy mix around 400 g CO2 eq/kWh that usage contribution has much more impact than the cluster lifetime. There, the carbon efficiency is much closer to the energy efficiency.

4

Discussions

4.1

On the importance of taking into account average vs. measured power estimation

MITSI estimates usage emissions based on the average power consumption of the cluster over a full year of operation. This yearly average smooths out the variability between workloads and idle periods, and is well-suited for an annual carbon footprint assessment of the data center as a whole. However, it does not reflect the actual energy consumed during a specific computation. In this paper, the carbon efficiency presented in Section 3 was computed from the actual power measured during the HPL runs, rather than from the yearly average. This distinction matters: the HPL benchmark stresses the cluster

10

N. Richart et al.

close to its peak power draw, making it unrepresentative of typical usage. As a consequence, the carbon efficiency figures reported here should be understood as the efficiency of the cluster during the benchmark, not during normal operation. 4.2

Toward a Carbon500 ranking

As discussed in previous sections, MITSI assesses the carbon footprint of IT services, during their whole lifespan, from manufacturing and distribution to usage and end-of-life. Applied to HPC clusters from Top500, a new ranking could be derived, named for instance "Carbon500", but it raises several challenges. A major contribution in such a "Carbon500" ranking would be the electrical carbon footprint, which is largely based on the location of the considered HPC cluster. Even if energy efficiency of data center does play a role in usage emissions, the carbon efficiency is mainly driven by the electrical carbon footprint. Different strategies are possible to assess the electrical carbon footprint, which could blur the comparison between different HPC clusters. Moreover, the annual electrical carbon footprint can be known only the following year, meaning that the cluster carbon footprints could only be computed up to a year after the measurements are achieved. Regarding the measurement of the embodied part, the data source is still too fragmented. Many parameters are only declarative and cannot be checked. This is particularly true for the data center’s PUE, and the server type (number of GPUs, network parameters, etc.) for instance. This would complicate the validation of this ranking. Nevertheless, focusing the ranking on carbon impact enables HPC clusters to be compared on their environmental impact, even if only from a carbon perspective. It has two positive impacts: it helps estimate the environmental impact of computation, especially from the scientific community, and it enables choosing the most environmentally friendly clusters on which large scientific projects are run. It then helps reduce their environmental impact and promote low-carbon research.

5

Conclusion

This study reports two Green500 submissions (Level 2 in 2024, Level 3 in 2025) and assesses the full carbon footprint of the Kuma cluster, covering both embodied and usage emissions. The Green500 energy efficiency optimum is found at a GPU power cap of 450 W. Results differ slightly depending on whether we consider the energy efficiency or the carbon emission efficiency, where for the latter an optimum is found closer to 500 W for HPC clusters running in a low-carbon electrical grid such as Switzerland. Other benchmarks, closer to real AI loads, should also be tested as they become a common run in HPC infrastructures. Green500 is already an improvement with respect to Top500, assessing the HPC cluster in regards of energy efficiency. But to encompass environmental impacts, this is not enough. Another ranking, focused on carbon emissions, that

Green500: Adventure in Qualifying a Cluster

11

we propose to name Carbon500, will also integrate the fluctuations of electrical mix, highly dependent on HPC locations and years. This would be a way to enhance motivation to highlight low-carbon-emission data centers—where scientific computation could then be located—and strengthen ambitious policies to reduce the environmental footprint of electricity production. Finally, carbon assessment alone is not enough to capture the full environmental impact of HPC; a multi-criteria life-cycle analysis encompassing abiotic resource depletion, water and land use, freshwater eco-toxicity, and fossil resource usage is needed to avoid erroneous conclusions [20]. Acknowledgments. MITSI was made thanks to the contribution of the Swiss Centre of Excellence on Net-Zero Emissions (SCENE [21]). Anthropic Claude was used as an AI assistant to help write this article. Disclosure of Interests. The authors have no competing interests to declare that are relevant to the content of this article.

References 1. Hpl-mxp mixed-precision benchmark, https://hpl-mxp.org/ 2. Adhinarayanan, V., Feng, W.c.: Looking back to look forward: 15 years of the green500. Computer 58(1), 76–86 (2025). https://doi.org/10.1109/MC.2023. 3333316 3. Alissa, H., Nick, T., Raniwala, A., et al.: Using life cycle assessment to drive innovation for sustainable cool clouds. Nature 641(8062), 331–338 (2025). https: //doi.org/10.1038/s41586-025-08832-3 4. Association des entreprises électriques suisses AES: Teneur en CO2 du mix électrique suisse d’ici 2050. Tech. rep. (2026), https://www.strom.ch/fr/document/ teneur-en-co2-du-mix-electrique-suisse-dici-2050-2026-0 5. Aubet, L., Violet, B.: Environmental impacts of it equipment - a comparative analysis - resilio database and ‘product carbon footprint’ sheets. Tech. rep. (2024), https://resilio-solutions.com/files/news/publications/ comparative-analysis/Resilio_Aguaro_Whitepaper_RDB_vs_PCF_20241025_ web_en.pdf 6. Bux, C., Rana, R.L., Lombardi, M., Giungato, P., Tricase, C.: A critical analysis of global warming potential of data centers in the digital era. The International Journal of Life Cycle Assessment 30(11), 2390–2402 (2025). https://doi.org/10. 1007/s11367-024-02419-2 7. Dongarra, J.J., Luszczek, P.: Hpl-mxp benchmark: Mixed-precision algorithms, iterative refinement, and scalable data generation. The International Journal of High Performance Computing Applications p. 10943420251382476 (2025). https: //doi.org/10.1177/10943420251382476 8. Dongarra, J.J., Moler, C.B., Bunch, J.R., Stewart, G.W.: LINPACK users’ guide. SIAM (1979) 9. Electricity Maps Aps: Electricity maps portal, https://app.electricitymaps. com/ 10. EPFL SCITAS: Epfl scitas kuma, https://www.epfl.ch/research/facilities/ scitas/hardware/kuma/

12

N. Richart et al.

11. Feng, W.c., Cameron, K.: The green500 list: Encouraging sustainable supercomputing. Computer 40(12), 50–55 (2007). https://doi.org/10.1109/MC.2007.445 12. Gupta, U., Kim, Y.G., Lee, S., Tse, J., Lee, H.H.S., Wei, G.Y., Brooks, D., Wu, C.J.: Chasing carbon: The elusive environmental footprint of computing. In: 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA). pp. 854–867. IEEE (2021). https://doi.org/10.1109/HPCA51647.2021. 00076 13. Heroux, M.A., Dongarra, J.J., Luszczek, P.: Hpcg benchmark technical specification. Tech. rep., Sandia National Lab.(SNL-NM), Albuquerque, NM (United States) (2013) 14. Karimi, A.M., Maiterth, M., Shin, W., Sattar, N.S., Lu, H., Wang, F.: Exploring the frontiers of energy efficiency using power management at system scale. In: Proceedings of the SC ’24 Workshops of the International Conference on High Performance Computing, Network, Storage, and Analysis. pp. 1835–1844. SC-W ’24, IEEE Press (2025). https://doi.org/10.1109/SCW63240.2024.00230 15. Kline, D., Parshook, N., Ge, X., Brunvand, E., Melhem, R., Chrysanthis, P.K., Jones, A.K.: Holistically evaluating the environmental impacts in modern computing systems. In: 2016 Seventh International Green and Sustainable Computing Conference (IGSC). pp. 1–8. IEEE (2016). https://doi.org/10.1109/IGCC.2016. 7892605 16. Kudo, S., Nitadori, K., Ina, T., Imamura, T.: Implementation and numerical techniques for one eflop/s hpl-ai benchmark on fugaku. In: 2020 IEEE/ACM 11th Workshop on Latest Advances in Scalable Algorithms for Large-Scale Systems (ScalA). pp. 69–76. IEEE (2020). https://doi.org/10.1109/ScalA51936.2020. 00014 17. Li, B., Basu Roy, R., Wang, D., Samsi, S., Gadepally, V., Tiwari, D.: Toward sustainable hpc: Carbon footprint estimation and environmental implications of hpc systems. In: Proceedings of the international conference for high performance computing, networking, storage and analysis. pp. 1–15 (2023). https://doi.org/ 10.1145/3581784.3607035 18. McCalpin, J.D.: Memory bandwidth and machine balance in current high performance computers. IEEE Computer Society Technical Committee on Computer Architecture (TCCA) Newsletter pp. 19–25 (Dec 1995) 19. Paolini, J., Cubero-Castan, M., Ouvrard, X., Peon Quirós, M., Richart, N.: MITSI - methodology for it services environmental impact assessment. Tech. rep. (2026), https://infoscience.epfl.ch/handle/20.500.14299/261022, version 1.0 - to be published soon 20. Resilio: Environmental impacts of the ict sector in switzerland. Tech. rep. (2025), version 1.0 - September 2025 21. SCENE Project: Scene project, https://www.scene-project.ch/ 22. Strohmaier, E., Meuer, H.W., Dongarra, J.J., Simon, H.D.: The top500 list and progress in high-performance computing. Computer 48(11), 42–49 (2015). https: //doi.org/10.1109/MC.2015.338 23. The Green500, Top500, The Green Grid, EEHPC WG: Energy efficient high performance computing power measurement methodology. Tech. rep. (2014), https:// www.top500.org/static/media/uploads/methodology-2.0rc1.pdf, version 2.0 RC 1.0 24. TOP500.org: Top500 (1993), https://top500.org/lists/top500/ 25. TOP500.org: Green500 (2013), https://top500.org/lists/green500/

Related documents

• HiPEAC Vision 2025: High Performance, Edge and Cloud Computing#1361094
• KubePACS: Kubernetes Cluster Using Performant, Highly Available, and Cost Efficient Spot Instances#138897
• Leveraging Serverless Computing for Flexibility and Scalability in Applications#106919
• HetCCL: Enabling Collective Communication For Mixed-Vendor Heterogeneous Clusters#246493
• Holistic Buoyancy-Driven Cloud Native Resource Management#1401338