pith. sign in

arxiv: 2605.24461 · v2 · pith:OL6VPJLHnew · submitted 2026-05-23 · 💻 cs.AR · cs.DC· cs.SY· eess.SY

Provisioning to Runtime Optimization of a 100 MW-Scale AI Cluster

Ehsan K. Ardestani , Leonardo Piga , Jovan Stojkovic , Pavan Balaji , Mustafa Ozdal , Mikel Jimenez Fernandez , Mihaela Dimovska , Luka Tadic
show 14 more authors
Hao Shen Devika Vishwanath Richa Mishra Melaku Mihret Valentin Andrei Mauricio Cespedes Julien Prigent James Monahan Tyler Graf Bin Li Charles Marquez Shobhit Kanaujia Kaushik Veeraraghavan Chunqiang Tang
This is my paper

Pith reviewed 2026-06-30 12:28 UTC · model grok-4.3

classification 💻 cs.AR cs.DCcs.SYeess.SY
keywords power managementAI datacenterGPU clusterhyper-scale computingruntime optimizationpower provisioningdatacenter scaling
0
0 comments X

The pith

Power management for a 150 MW AI cluster begins 6-12 months before accelerators arrive and continues through runtime adjustments.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out the sequence of power decisions required to run a hyper-scale AI datacenter. It starts with capacity planning well before new hardware is available, moves to setting adjustments once the machines are installed, and ends with ongoing runtime controls that respond to changing workloads. The authors illustrate each stage with measurements taken from an operating 150 MW facility that holds 83,000 GB200 GPUs. A reader would care because the text identifies electric power supply, rather than accelerator count, as the current binding constraint on further AI scaling.

Core claim

The end-to-end power management process for a hyper-scale AI datacenter consists of early power planning to accommodate next-generation accelerators 6-12 months before their general availability, tuning of power settings after large-scale deployment, and dynamic runtime power management for evolving workloads, illustrated by detailed measurements from a 150 MW datacenter hosting 83K GB200 GPUs.

What carries the argument

The three-stage power management pipeline: pre-deployment provisioning, post-installation tuning, and runtime optimization.

Load-bearing premise

The power management steps and measurements taken on this particular 150 MW cluster of 83K GB200 GPUs apply to other hyper-scale AI datacenters.

What would settle it

Documentation from a second hyper-scale AI cluster showing a materially different sequence of planning, tuning, and runtime steps would indicate that the described process is not general.

Figures

Figures reproduced from arXiv: 2605.24461 by Bin Li, Charles Marquez, Chunqiang Tang, Devika Vishwanath, Ehsan K. Ardestani, Hao Shen, James Monahan, Jovan Stojkovic, Julien Prigent, Kaushik Veeraraghavan, Leonardo Piga, Luka Tadic, Mauricio Cespedes, Melaku Mihret, Mihaela Dimovska, Mikel Jimenez Fernandez, Mustafa Ozdal, Pavan Balaji, Richa Mishra, Shobhit Kanaujia, Tyler Graf, Valentin Andrei.

Figure 1
Figure 1. Figure 1: The Catalina pod GB200 configuration [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Back end network schematics. There are 3 levels of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance improvement of a system with [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: Daily mechanical peak-minute power for the first [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: GB200 FP8 FLOPS sensitivity to power limit. Arith [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: GB200 HBM bandwidth sensitivity to power limit. [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: At a fixed power budget, higher GPU power in [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: High-level diagram of PSU power validation. (a) [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 13
Figure 13. Figure 13: Comparison of maximum DCIM power samples [PITH_FULL_IMAGE:figures/full_fig_p008_13.png] view at source ↗
Figure 11
Figure 11. Figure 11: PSU versus AC oscilloscope power measurements. [PITH_FULL_IMAGE:figures/full_fig_p008_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of maximum DCIM power samples [PITH_FULL_IMAGE:figures/full_fig_p008_12.png] view at source ↗
Figure 15
Figure 15. Figure 15: CDF of planned power headroom across RPPs: (a) [PITH_FULL_IMAGE:figures/full_fig_p008_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Power consumption of a GB200 rack running a [PITH_FULL_IMAGE:figures/full_fig_p009_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Software power smoother draws up to 800W of [PITH_FULL_IMAGE:figures/full_fig_p010_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Software based power smoother on a region-scale [PITH_FULL_IMAGE:figures/full_fig_p010_18.png] view at source ↗
Figure 20
Figure 20. Figure 20: GPU TDP selected by Dimmer for servers running [PITH_FULL_IMAGE:figures/full_fig_p011_20.png] view at source ↗
Figure 19
Figure 19. Figure 19: Adjusting power limit for a host in a training job, [PITH_FULL_IMAGE:figures/full_fig_p011_19.png] view at source ↗
Figure 21
Figure 21. Figure 21: Relative Cluster throughput through Power Man [PITH_FULL_IMAGE:figures/full_fig_p012_21.png] view at source ↗
read the original abstract

The electric power supply for AI data centers is now the most significant bottleneck in the race toward Artificial General Intelligence, surpassing even the constraint of AI accelerator availability. To our knowledge, this paper is the first to describe the end-to-end power management process for a hyper-scale AI datacenter; from early power planning to accommodate next-generation accelerators 6--12 months before their general availability, to tuning power settings after large scale deployment, and finally to dynamic, runtime power management for evolving workloads. We present detailed power measurements for a 150 MW datacenter hosting a cluster of 83K GB200 GPUs. We share insights from building this state-of-the-art AI cluster. We hope this work encourages practitioners across the industry to share their own experiences as well.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript claims to be the first to describe the end-to-end power management process for a hyper-scale AI datacenter, covering early power planning 6-12 months before accelerator availability, post-deployment tuning, and dynamic runtime optimization for evolving workloads. It presents detailed power measurements from a 150 MW datacenter hosting 83K GB200 GPUs and shares insights from building this cluster.

Significance. If the reported processes, quantitative measurements, and insights hold and are shown to be representative, the work would address a critical and timely bottleneck in AI infrastructure scaling. The absence of any equations, derivations, or fitted parameters keeps the burden of proof on empirical description rather than theoretical novelty.

major comments (2)
  1. [Abstract] Abstract: the claim that 'detailed power measurements' and 'insights' are presented is not accompanied by any data, methods, error analysis, or validation steps, leaving the central claims resting on unshown evidence.
  2. [Abstract] Abstract: the assertion that the described process is useful for the community and generalizes to other hyper-scale AI datacenters requires an explicit argument or cross-check showing why observed behaviors transfer beyond this specific 150 MW site's power infrastructure, cooling, workload mix, and contractual constraints; none is supplied.
minor comments (1)
  1. [Title] Title states '100 MW-Scale' while the abstract and body reference a 150 MW cluster; this inconsistency should be reconciled.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed review and constructive comments on our manuscript. We provide point-by-point responses to the major comments below. We have revised the manuscript to address the concerns where possible.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that 'detailed power measurements' and 'insights' are presented is not accompanied by any data, methods, error analysis, or validation steps, leaving the central claims resting on unshown evidence.

    Authors: We note that the abstract is intended as a concise overview and does not contain the full empirical details. The manuscript body includes extensive sections with power measurement data from the 150 MW cluster, descriptions of the methods used for data collection, error analysis, and validation procedures. To improve clarity, we will update the abstract to explicitly state that these details are provided in the main text. revision: yes

  2. Referee: [Abstract] Abstract: the assertion that the described process is useful for the community and generalizes to other hyper-scale AI datacenters requires an explicit argument or cross-check showing why observed behaviors transfer beyond this specific 150 MW site's power infrastructure, cooling, workload mix, and contractual constraints; none is supplied.

    Authors: The manuscript focuses on a detailed case study of our deployment. While we believe the insights are valuable and the workflow can inform other efforts, we acknowledge the need for a more explicit discussion on generalizability. We will add a new subsection in the discussion that addresses potential variations in power infrastructure, cooling systems, workload characteristics, and contractual aspects, explaining the transferable elements of the approach. revision: yes

Circularity Check

0 steps flagged

No circularity; purely descriptive measurements and process narrative

full rationale

The paper contains no equations, derivations, fitted parameters, predictions, or mathematical models. Its central contribution is a narrative description of power provisioning, tuning, and runtime management on one 150 MW / 83K GB200 cluster, accompanied by measured data points. No load-bearing step reduces to a self-definition, a fitted input renamed as prediction, or a self-citation chain. The claim of being 'first to describe' is a statement of novelty, not a derived result. This is the normal case for an industry experience report and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract contains no mathematical model, free parameters, axioms, or invented entities; the contribution is a process description rather than a formal system.

pith-pipeline@v0.9.1-grok · 5757 in / 1216 out tokens · 42372 ms · 2026-06-30T12:28:07.295773+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 1 canonical work pages

  1. [1]

    The open compute project accelerating deployment of next gen ai clusters

  2. [2]

    Processor state control for your EC2 instance, April 2024

    Amazon Web Services. Processor state control for your EC2 instance, April 2024

    2024

  3. [3]

    AMD and OpenAI Announce Strate- gic Partnership to Deploy 6 Gigawatts of AMD GPUs, 2025

    AMD. AMD and OpenAI Announce Strate- gic Partnership to Deploy 6 Gigawatts of AMD GPUs, 2025. https://ir.amd.com/ news-events/press-releases/detail/1260/ amd-and-openai-announce-strategic-partnership-to-deploy-6-gigawatts-of-amd-gpus

    2025

  4. [4]

    A taxonomy and survey of energy- efficient data centers and cloud computing systems.Ad- vances in computers, 82:47–111, 2011

    Anton Beloglazov, Rajkumar Buyya, Young Choon Lee, and Albert Zomaya. A taxonomy and survey of energy- efficient data centers and cloud computing systems.Ad- vances in computers, 82:47–111, 2011

    2011

  5. [5]

    A survey of design techniques for system- level dynamic power management.IEEE transactions on very large scale integration (VLSI) systems, 8(3):299– 316, 2002

    Luca Benini, Alessandro Bogliolo, and Giovanni De Micheli. A survey of design techniques for system- level dynamic power management.IEEE transactions on very large scale integration (VLSI) systems, 8(3):299– 316, 2002

    2002

  6. [6]

    Bhattacharya, David Culler, Aman Kansal, Sri- ram Govindan, and Sriram Sankar

    Arka A. Bhattacharya, David Culler, Aman Kansal, Sri- ram Govindan, and Sriram Sankar. The need for speed and stability in data center power capping. InProceed- ings of the International Green Computing Conference (IGCC ’12), 2012

    2012

  7. [7]

    Power stabilization for ai training datacenters,

    Esha Choukse, Brijesh Warrier, Scot Heath, Luz Bel- mont, April Zhao, Hassan Ali Khan, Brian Harry, Matthew Kappel, Russell J Hewett, Kushal Datta, et al. Power stabilization for ai training datacenters.arXiv preprint arXiv:2508.14318, 2025

    work page arXiv 2025

  8. [8]

    Scaling llama 3 training with efficient parallelism strategies

    Weiwei Chu, Xinfeng Xie, Jiecao Yu, Jie Wang, Amar Phanishayee, Chunqiang Tang, Yuchen Hao, Jianyu Huang, Mustafa Ozdal, Jun Wang, et al. Scaling llama 3 training with efficient parallelism strategies. InPro- ceedings of the 52nd Annual International Symposium on Computer Architecture, pages 1703–1716, 2025

    2025

  9. [9]

    Data center energy consumption modeling: A survey.IEEE Communications surveys & tutorials, 18(1):732–794, 2015

    Miyuru Dayarathna, Yonggang Wen, and Rui Fan. Data center energy consumption modeling: A survey.IEEE Communications surveys & tutorials, 18(1):732–794, 2015

    2015

  10. [10]

    Electricity explained: Electricity generation, capacity, and sales in the United States, 2024

    EIA. Electricity explained: Electricity generation, capacity, and sales in the United States, 2024. https: //www.eia.gov/energyexplained/electricity/ electricity-in-the-us-generation-capacity-and-sales. php

    2024

  11. [11]

    EIA. U.S. battery capacity increased 66% in 2024,

    2024

  12. [12]

    https://www.eia.gov/todayinenergy/ detail.php?id=64705

  13. [13]

    Unified architecture - opc foundation

    OPC Foundation. Unified architecture - opc foundation. https://opcfoundation.org/about/ opc-technologies/opc-ua/, September 2019. Ac- cessed: Dec. 11, 2025

    2019

  14. [14]

    How much power oversubscription is safe and allowed in data cen- ters? InProceedings of the 8th International Conference on Autonomic Computing (ICAC 2011), pages 91–100

    Xing Fu, Jian Guo, Yuanqiang Wang, Yang Zhao, Mian Lin, Fang Liu, Zhiming Wu, and Jie Li. How much power oversubscription is safe and allowed in data cen- ters? InProceedings of the 8th International Conference on Autonomic Computing (ICAC 2011), pages 91–100. ACM, 2011

    2011

  15. [15]

    CPU platforms, April 2024

    Google Compute Platform. CPU platforms, April 2024

    2024

  16. [16]

    Statisti- cal profiling-based techniques for effective power provi- sioning in data centers

    Sriram Govindan, Jeonghwan Choi, Bhuvan Urgaonkar, Anand Sivasubramaniam, and Andrea Baldini. Statisti- cal profiling-based techniques for effective power provi- sioning in data centers. InProceedings of the 4th ACM European Conference on Computer Systems (EuroSys ’09), 2009

    2009

  17. [17]

    The open protocol standard for computerized building systems: Bacnet

    Larry K Haakenstad. The open protocol standard for computerized building systems: Bacnet. InProceedings of the 1999 IEEE International Conference on Control Applications (Cat. No. 99CH36328), volume 2, pages 1585–1590. IEEE, 1999

    1999

  18. [18]

    Haque, Yuxiong He, Sameh Elnikety, Thu D

    Md E. Haque, Yuxiong He, Sameh Elnikety, Thu D. Nguyen, Ricardo Bianchini, and Kathryn S. McKin- ley. Exploiting Heterogeneity for Tail Latency and En- ergy Efficiency. InProceedings of the 50th Annual IEEE/ACM International Symposium on Microarchitec- ture (MICRO ’17), 2017

    2017

  19. [19]

    SmoothOperator: Reducing Power Frag- mentation and Improving Power Utilization in Large- Scale Datacenters

    Chang-Hong Hsu, Qingyuan Deng, Jason Mars, and Lingjia Tang. SmoothOperator: Reducing Power Frag- mentation and Improving Power Utilization in Large- Scale Datacenters. InProceedings of the 23rd Inter- national Conference on Architectural Support for Pro- gramming Languages and Operating Systems (ASPLOS ’18), 2018

    2018

  20. [20]

    WattWiser: Power & Resource-Efficient Schedul- ing for Multi-Model Multi-GPU Inference Servers

    Ali Jahanshahi, Mohammadreza Rezvani, and Daniel Wong. WattWiser: Power & Resource-Efficient Schedul- ing for Multi-Model Multi-GPU Inference Servers. In Proceedings of the 14th International Green and Sus- tainable Computing Conference (IGSC ’23), 2023

    2023

  21. [21]

    SLO-aware GPU DVFS for Energy-efficient LLM Inference Serving

    Andreas Kosmas Kakolyris, Dimosthenis Masouros, Sotirios Xydis, and Dimitrios Soudris. SLO-aware GPU DVFS for Energy-efficient LLM Inference Serving. IEEE Computer Architecture Letters, 2024

    2024

  22. [22]

    AutoScale: Energy Efficiency Optimization for Stochastic Edge In- ference Using Reinforcement Learning

    Young Geun Kim and Carole-Jean Wu. AutoScale: Energy Efficiency Optimization for Stochastic Edge In- ference Using Reinforcement Learning. InProceedings of the 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO ’20), 2020. 13

    2020

  23. [23]

    Power capping of CPU-GPU heterogeneous systems through coordinat- ing DVFS and task mapping

    Toshiya Komoda, Shingo Hayashi, Takashi Nakada, Shi- nobu Miwa, and Hiroshi Nakamura. Power capping of CPU-GPU heterogeneous systems through coordinat- ing DVFS and task mapping. InProceedings of the IEEE 31st International Conference on Computer De- sign (ICCD ’13), 2013

    2013

  24. [24]

    Tullsen, and Tajana Simunic Rosing

    Vasileios Kontorinis, Liuyi Eric Zhang, Baris Ak- sanli, Jack Sampson, Houman Homayoun, Eddie Pettis, Dean M. Tullsen, and Tajana Simunic Rosing. Manag- ing Distributed Ups Energy for Effective Power Capping in Data Centers. InProceedings of the 39th Annual In- ternational Symposium on Computer Architecture (ISCA ’12), 2012

    2012

  25. [25]

    Misra, Seyyed Ahmad Javadi, Bianca Schroeder, Marcus Fontoura, and Ricardo Bianchini

    Alok Gautam Kumbhare, Reza Azimi, Ioannis Manousakis, Anand Bonde, Felipe Frujeri, Nithish Mahalingam, Pulkit A. Misra, Seyyed Ahmad Javadi, Bianca Schroeder, Marcus Fontoura, and Ricardo Bianchini. Prediction-Based Power Oversubscription in Cloud Platforms. InProceedings of the USENIX Annual Technical Conference (USENIX ATC ’21), 2021

    2021

  26. [26]

    Thunderbolt: Throughput-Optimized, Quality-of-Service-Aware Power Capping at Scale

    Shaohong Li, Xi Wang, Xiao Zhang, Vasileios Kontori- nis, Sreekumar Kodakara, David Lo, and Parthasarathy Ranganathan. Thunderbolt: Throughput-Optimized, Quality-of-Service-Aware Power Capping at Scale. In Proceedings of the 14th USENIX Symposium on Oper- ating Systems Design and Implementation (OSDI ’20), 2020

    2020

  27. [27]

    Towards energy pro- portionality for large-scale latency-critical workloads

    David Lo, Liqun Cheng, Rama Govindaraju, Luiz André Barroso, and Christos Kozyrakis. Towards energy pro- portionality for large-scale latency-critical workloads. InProceedings of the ACM/IEEE 41st International Symposium on Computer Architecture (ISCA ’14), 2014

    2014

  28. [28]

    Virtual Machine series, April 2024

    Microsoft Azure. Virtual Machine series, April 2024

    2024

  29. [29]

    BatchSizer: Power-Performance Trade-off for DNN Inference

    Seyed Morteza Nabavinejad, Sherief Reda, and Ma- soumeh Ebrahimi. BatchSizer: Power-Performance Trade-off for DNN Inference. InProceedings of the 26th Asia and South Pacific Design Automation Confer- ence (ASP-DAC ’21), 2021

    2021

  30. [30]

    Coordinated Batching and DVFS for DNN Inference on GPU Accelerators.IEEE Trans- actions on Parallel and Distributed Systems, 33(10), 2022

    Seyed Morteza Nabavinejad, Sherief Reda, and Ma- soumeh Ebrahimi. Coordinated Batching and DVFS for DNN Inference on GPU Accelerators.IEEE Trans- actions on Parallel and Distributed Systems, 33(10), 2022

    2022

  31. [31]

    Twig: Multi-Agent Task Manage- ment for Colocated Latency-Critical Cloud Services

    Rajiv Nishtala, Vinicius Petrucci, Paul Carpenter, and Magnus Sjalander. Twig: Multi-Agent Task Manage- ment for Colocated Latency-Critical Cloud Services. In Proceedings of the IEEE International Symposium on High Performance Computer Architecture (HPCA ’20), 2020

    2020

  32. [32]

    OpenAI and NVIDIA An- nounce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems, 2025

    NVIDIA. OpenAI and NVIDIA An- nounce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems, 2025. https://nvidianews.nvidia.com/news/ openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems

    2025

  33. [33]

    OpenAI and Broadcom announce strategic col- laboration to deploy 10 gigawatts of OpenAI-designed AI accelerators, 2025

    OpenAI. OpenAI and Broadcom announce strategic col- laboration to deploy 10 gigawatts of OpenAI-designed AI accelerators, 2025. https://openai.com/index/ openai-and-broadcom-announce-strategic-collaboration/

    2025

  34. [34]

    OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites, 2025

    OpenAI. OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites, 2025. https:// openai.com/index/five-new-stargate-sites/

    2025

  35. [35]

    Expanding datacenter capacity with dvfs boosting: A safe and scalable deployment ex- perience

    Leonardo Piga, Iyswarya Narayanan, Aditya Sundar- rajan, Matt Skach, Qingyuan Deng, Biswadip Maity, Manoj Chakkaravarthy, Alison Huang, Abhishek Dhan- otia, and Parth Malani. Expanding datacenter capacity with dvfs boosting: A safe and scalable deployment ex- perience. InProceedings of the 29th ACM International Conference on Architectural Support for Pro...

    2024

  36. [36]

    Ranganathan, P

    P. Ranganathan, P. Leech, D. Irwin, and J. Chase. Ensemble-level Power Management for Dense Blade Servers. InProceedings of the 33rd Annual Interna- tional Symposium on Computer Architecture (ISCA ’06), 2006

    2006

  37. [37]

    Data Center Power Oversubscription with a Medium V olt- age Power Plane and Priority-Aware Capping

    Varun Sakalkar, Vasileios Kontorinis, David Landhuis, Shaohong Li, Darren De Ronde, Thomas Blooming, Anand Ramesh, James Kennedy, Christopher Malone, Jimmy Clidaras, and Parthasarathy Ranganathan. Data Center Power Oversubscription with a Medium V olt- age Power Plane and Priority-Aware Capping. InPro- ceedings of the 25th International Conference on Ar- ...

    2020

  38. [38]

    From words to watts: Benchmarking the energy costs of large language model inference

    Siddharth Samsi, Dan Zhao, Joseph McDonald, Baolin Li, Adam Michaleas, Michael Jones, William Bergeron, Jeremy Kepner, Devesh Tiwari, and Vijay Gadepally. From words to watts: Benchmarking the energy costs of large language model inference. In2023 IEEE High Performance Extreme Computing Conference (HPEC), pages 1–9. IEEE, 2023

    2023

  39. [39]

    EcoFaaS: Rethinking the Design of Serverless Environments for Energy Effi- ciency

    Jovan Stojkovic, Nikoleta Iliakopoulou, Tianyin Xu, Hu- bertus Franke, and Josep Torrellas. EcoFaaS: Rethinking the Design of Serverless Environments for Energy Effi- ciency. InProceedings of the 51st Annual International Symposium on Computer Architecture (ISCA ’24), 2024. 14

    2024

  40. [40]

    SmartOClock: Workload- and Risk-Aware Overclocking in the Cloud

    Jovan Stojkovic, Pulkit Misra, Inigo Goiri, Sam Whit- lock, Esha Choukse, Mayukh Das, Chetan Bansal, Ja- son Lee, Zoey Sun, Haoran Qiu, Reed Zimmermann, Savyasachi Samal, Brijesh Warrier, Ashish Raniwala, and Ricardo Bianchini. SmartOClock: Workload- and Risk-Aware Overclocking in the Cloud. InProceed- ings of the 51st Annual International Symposium on Co...

    2024

  41. [41]

    DynamoLLM: Designing LLM Inference Clusters for Performance and Energy Efficiency

    Jovan Stojkovic, Chaojie Zhang, Íñigo Goiri, Josep Tor- rellas, and Esha Choukse. DynamoLLM: Designing LLM Inference Clusters for Performance and Energy Efficiency. InProceedings of the IEEE International Symposium on High Performance Computer Architec- ture (HPCA’25), 2025

    2025

  42. [42]

    Rush, David Brooks, and Gu-Yeon Wei

    Thierry Tambe, Coleman Hooper, Lillian Pentecost, Tianyu Jia, En-Yu Yang, Marco Donato, Victor Sanh, Paul Whatmough, Alexander M. Rush, David Brooks, and Gu-Yeon Wei. EdgeBERT: Sentence-Level Energy Optimizations for Latency-Aware Multi-Task NLP In- ference. InProceedings of the 54th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO ’21), 2021

    2021

  43. [43]

    The Impact of GPU DVFS on the En- ergy and Performance of Deep Learning: An Empirical Study

    Zhenheng Tang, Yuxin Wang, Qiang Wang, and Xi- aowen Chu. The Impact of GPU DVFS on the En- ergy and Performance of Deep Learning: An Empirical Study. InProceedings of the Tenth ACM International Conference on Future Energy Systems (e-Energy ’19), 2019

    2019

  44. [44]

    Introduction to the modbus protocol

    George Thomas. Introduction to the modbus protocol. The Extension, 9(4):1–4, 2008

    2008

  45. [45]

    ALERT: Accurate learning for energy and timeliness

    Chengcheng Wan, Muhammad Santriaji, Eri Rogers, Henry Hoffmann, Michael Maire, and Shan Lu. ALERT: Accurate learning for energy and timeliness. InPro- ceedings of the USENIX Annual Technical Conference (USENIX ATC ’20), 2020

    2020

  46. [46]

    Dynamic GPU Energy Optimiza- tion for Machine Learning Training Workloads.IEEE Transactions on Parallel and Distributed Systems, 2022

    Farui Wang, Weizhe Zhang, Shichao Lai, Meng Hao, and Zheng Wang. Dynamic GPU Energy Optimiza- tion for Machine Learning Training Workloads.IEEE Transactions on Parallel and Distributed Systems, 2022

    2022

  47. [47]

    Dynamo: Facebook’s Data Center- Wide Power Management System

    Qiang Wu, Qingyuan Deng, Lakshmi Ganesh, Chang- Hong Hsu, Yun Jin, Sanjeev Kumar, Bin Li, Justin Meza, and Yee Jiun Song. Dynamo: Facebook’s Data Center- Wide Power Management System. InProceedings of the 43rd Annual International Symposium on Computer Architecture (ISCA ’16), 2016

    2016

  48. [48]

    Zeus: Understanding and optimizing GPU energy con- sumption of DNN training

    Jie You, Jae-Won Chung, and Mosharaf Chowdhury. Zeus: Understanding and optimizing GPU energy con- sumption of DNN training. InProceedings of the 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI ’23), 2023

    2023

  49. [49]

    Know Your Enemy To Save Cloud Energy: Energy- Performance Characterization of Machine Learning Serving

    Junyeol Yu, Jongseok Kim, and Euiseong Seo. Know Your Enemy To Save Cloud Energy: Energy- Performance Characterization of Machine Learning Serving. In2023 IEEE International Symposium on High-Performance Computer Architecture (HPCA), 2023

    2023

  50. [50]

    Misra, Rod As- sis, Kyle Woolcock, Nithish Mahalingam, Brijesh War- rier, David Gauthier, Lalu Kunnath, Steve Solomon, Os- valdo Morales, Marcus Fontoura, and Ricardo Bianchini

    Chaojie Zhang, Alok Gautam Kumbhare, Ioannis Manousakis, Deli Zhang, Pulkit A. Misra, Rod As- sis, Kyle Woolcock, Nithish Mahalingam, Brijesh War- rier, David Gauthier, Lalu Kunnath, Steve Solomon, Os- valdo Morales, Marcus Fontoura, and Ricardo Bianchini. Flex: High-Availability Datacenters with Zero Reserved Power. InProceedings of the 48th Annual Inter...

    2021

  51. [51]

    Improving GPU Energy Efficiency through an Application-transparent Frequency Scaling Policy with Performance Assurance

    Yijia Zhang, Qiang Wang, Zhe Lin, Pengxiang Xu, and Bingqiang Wang. Improving GPU Energy Efficiency through an Application-transparent Frequency Scaling Policy with Performance Assurance. InProceedings of the Nineteenth European Conference on Computer Systems (EuroSys ’24), 2024

    2024

  52. [52]

    Indicator-Directed Dynamic Power Management for It- erative Workloads on GPU-Accelerated Systems

    Pengfei Zou, Ang Li, Kevin Barker, and Rong Ge. Indicator-Directed Dynamic Power Management for It- erative Workloads on GPU-Accelerated Systems. In Proceedings of the 20th IEEE/ACM International Sym- posium on Cluster, Cloud and Internet Computing (CC- GRID ’20), 2020. 15

    2020