pith. sign in

arxiv: 2606.05741 · v1 · pith:HGEW5XDVnew · submitted 2026-06-04 · 💻 cs.AR · cs.ET

Space-CIM: Enabling Compute-In-Memory Accelerators for Thermally-Constrained Space Platforms

Sohan Salahuddin Mugdho , Md. Shahedul Hasan , Cheng Wang This is my paper

Pith reviewed 2026-06-27 23:36 UTC · model grok-4.3

classification 💻 cs.AR cs.ET
keywords compute-in-memoryspace platformsradiative coolingthermal constraintsGPU acceleratorsTOPS/WAI workloads
0
0 comments X

The pith

Compute-in-memory accelerators achieve higher TOPS/W than GPUs when radiator cooling capacity limits space platforms.

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

The paper investigates thermal limits for AI accelerators in space, where systems must rely solely on radiative cooling through large radiators because convection is absent. It develops a co-design approach that ties the maximum allowable system throughput directly to radiator size. Simulations show GPUs with separate dies and HBM create localized hot spots that trigger throttling under tight radiator budgets. CIM accelerators avoid these hot spots through integrated memory and compute, maintaining higher sustained performance across workloads. This matters for orbital data centers that could use abundant solar power but face strict thermal management constraints.

Core claim

CIM accelerators exhibit a much more uniform heat distribution and consistently outperform GPUs in TOPS/W across a wide range of radiator budgets. Systematic evaluation across AI workloads demonstrates that CIM has a magnified advantage for deployment in space under realistic thermal constraints.

What carries the argument

The radiator-in-the-loop co-design methodology that directly links permitted system TOPS with practical radiator cooling capacity under vacuum conditions.

If this is right

  • CIM sustains higher effective throughput without thermal throttling for any fixed radiator area.
  • The performance gap between CIM and GPU widens as available radiator capacity shrinks.
  • CIM reduces the radiator area needed to reach a target compute rate compared with GPU designs.
  • The advantage holds across multiple AI workloads when thermal constraints are enforced.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Orbital AI systems should favor CIM architectures if vacuum thermal behavior matches the modeled conditions.
  • Combining the radiator sizing results with solar array mass and total platform volume would give a fuller picture of end-to-end efficiency.
  • Hardware prototypes of both architectures could be compared in a thermal-vacuum chamber to test the simulation predictions directly.

Load-bearing premise

The thermal simulations accurately model heat flow for separately located GPU die and HBM versus integrated CIM under vacuum radiative cooling with no convection.

What would settle it

Measured heat maps and sustained TOPS from a GPU-HBM system versus an equivalent CIM accelerator tested in a vacuum chamber with controlled radiative cooling would confirm or refute the uniform distribution and performance advantage.

Figures

Figures reproduced from arXiv: 2606.05741 by Cheng Wang, Md. Shahedul Hasan, Sohan Salahuddin Mugdho.

Figure 1
Figure 1. Figure 1: (a) An overview of the key thermal management challenges for space AI hardware compared to terrestrial hardware. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Thermal maps of GPU+HBM (top) and CIM (bottom) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Thermal behavior of GPU+HBM and CIM under [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance scaling of the GPU and representative [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Design space exploration of CIM under a low TRP [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance comparison of the GPU and representative CIM systems across workloads under different radiator [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
read the original abstract

The rapid growth in compute demand from artificial intelligence (AI) has driven a massive surge in data center construction, precipitating an energy and sustainability crisis. Motivated by the abundant solar energy in outer space and the recent sharp reduction in space launch costs, orbital data centers are emerging as a potential pathway for the future scaling of AI compute infrastructure. While the cold background in vacuum seems appealing for cooling, computing systems operating in space without convection ultimately rely on radiative cooling, requiring large-area radiators. Such limitations in thermal management pose a significant challenge for deploying the standard liquid/air-cooled computers in space. In this work, we investigate the impact of the thermal constraints in space on both graphics processing units (GPUs) with high-bandwidth memory (HBM) and the emerging compute-in-memory (CIM) accelerators. We develop a radiator-in-the-loop co-design methodology that directly links the permitted system TOPS (terra-operations per second) with the practical radiator cooling capacity in space. Our thermal simulations reveal that the separately located GPU die and HBMs create severe thermal hotspots under limited radiator capacity, necessitating GPU thermal throttling. In contrast, CIM accelerators exhibit a much more uniform heat distribution and consistently outperform GPUs in TOPS/W across a wide range of radiator budgets. We systematically evaluated the performance of CIM and GPU across various AI workloads and demonstrated that CIM has a magnified advantage for deployment in space under realistic thermal constraints.

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 / 2 minor

Summary. The manuscript develops a radiator-in-the-loop co-design methodology that couples permitted system TOPS to practical radiative cooling capacity for space platforms. Thermal simulations are used to show that discrete GPU die + HBM packages produce severe hotspots under limited radiator budgets, forcing thermal throttling, whereas CIM accelerators exhibit uniform heat distribution and deliver consistently higher TOPS/W across a range of radiator sizes; the advantage is reported to be magnified for AI workloads under realistic space thermal constraints.

Significance. If the underlying thermal model is accurate, the work supplies a concrete quantitative link between accelerator microarchitecture, power density, and radiator sizing that could inform accelerator selection for orbital AI infrastructure. The co-design framing itself is a useful contribution even if the numerical gap between CIM and GPU narrows after further model validation.

major comments (2)
  1. [Thermal simulation methodology (likely §4 or equivalent)] The central performance claims rest on the thermal simulation results that predict severe hotspots for separately packaged GPU+HBM versus uniform dissipation in CIM. The manuscript must supply the explicit model parameters (emissivities, view factors to the radiator, component locations and power maps) and any cross-validation against analytical radiative heat-transfer solutions or published vacuum-cooling data; without these, the reported TOPS/W gap cannot be assessed for systematic bias.
  2. [Abstract and §3 (co-design methodology)] The radiator-in-the-loop co-design claims to directly link permitted TOPS to cooling capacity, yet the abstract and visible text provide no governing equations, no error bars on the simulated temperatures or efficiencies, and no workload-specific power traces. These omissions make the quantitative superiority statements impossible to reproduce or bound.
minor comments (2)
  1. [Figures 3–5] Figure captions and axis labels should explicitly state the radiator budget range (e.g., m² or W/K) and the exact AI workloads used for the TOPS/W comparison.
  2. [Throughout] Notation for TOPS/W should be defined once and used consistently; the distinction between peak and sustained efficiency under throttling should be clarified.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help strengthen the reproducibility and clarity of our thermal co-design methodology. We address each major point below and have incorporated revisions to supply the requested details.

read point-by-point responses

  1. Referee: The central performance claims rest on the thermal simulation results that predict severe hotspots for separately packaged GPU+HBM versus uniform dissipation in CIM. The manuscript must supply the explicit model parameters (emissivities, view factors to the radiator, component locations and power maps) and any cross-validation against analytical radiative heat-transfer solutions or published vacuum-cooling data; without these, the reported TOPS/W gap cannot be assessed for systematic bias.

    Authors: We agree that explicit parameters are required to allow independent assessment of the thermal model. In the revised manuscript we have added a dedicated subsection (§4.2) that lists all emissivities, view factors, component locations, and power maps. We also include a direct comparison of the finite-element results against the analytical radiative heat-transfer equation for a simplified two-body system and reference published vacuum-chamber data for similar radiator configurations. These additions enable readers to evaluate potential systematic bias in the reported TOPS/W advantage. revision: yes

  2. Referee: The radiator-in-the-loop co-design claims to directly link permitted TOPS to cooling capacity, yet the abstract and visible text provide no governing equations, no error bars on the simulated temperatures or efficiencies, and no workload-specific power traces. These omissions make the quantitative superiority statements impossible to reproduce or bound.

    Authors: We accept that the governing equations, uncertainty quantification, and workload traces were insufficiently visible. Section 3 has been expanded to present the full set of radiator-in-the-loop equations, including the mapping from radiator area to allowable system power. We now report error bars derived from Monte-Carlo variation of emissivity and view-factor parameters, and we supply per-workload power traces for the evaluated AI models. These changes make the quantitative claims reproducible and bounded. revision: yes

Circularity Check

0 steps flagged

No circularity; claims rest on external thermal simulations and co-design methodology

full rationale

The abstract and provided context describe a radiator-in-the-loop co-design and thermal simulations comparing GPU+HBM hotspots versus CIM uniformity under vacuum radiative cooling. No equations, fitted parameters, self-citations, or ansatzes are shown that reduce any prediction to its own inputs by construction. The derivation chain is self-contained against the external simulation benchmarks rather than self-referential. This matches the reader's non-circularity assessment.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only abstract available; no free parameters, axioms, or invented entities can be extracted.

pith-pipeline@v0.9.1-grok · 5798 in / 885 out tokens · 19636 ms · 2026-06-27T23:36:38.143791+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

39 extracted references · 8 canonical work pages

  1. [1]

    Emer, and Vivienne Sze

    Tanner Andrulis, Joel S. Emer, and Vivienne Sze. 2024. CiMLoop: A Flexible, Ac- curate, and Fast Compute-In-Memory Modeling Tool. In2024 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS). 10–23. doi:10.1109/ISPASS61541.2024.00012 Space-CIM: Enabling Compute-In-Memory Accelerators for Thermally-Constrained Space Platforms

    work page doi:10.1109/ispass61541.2024.00012 2024

  2. [2]

    Aayush Ankit, Izzat El Hajj, Sai Rahul Chalamalasetti, Geoffrey Ndu, Martin Foltin, R Stanley Williams, Paolo Faraboschi, Wen-mei W Hwu, John Paul Stra- chan, Kaushik Roy, et al. 2019. PUMA: A programmable ultra-efficient memristor- based accelerator for machine learning inference. InProceedings of the twenty- fourth international conference on architectu...

    2019

  3. [3]

    Umberto Battista, Alberto Landini, Wojciech Golebiowski, Rafal Michalczyk, Adam Czerwinski, Krzysztof Duda, and Agata Sochaczewska. 2017. Design of Net Ejector for Space Debris Capturing. InProceedings of the 7th European Conference on Space Debris (SDC7). ESA Space Debris Office. https://conference. sdo.esoc.esa.int/proceedings/sdc7/paper/279

    2017

  4. [4]

    Travis Beals, Maria Biggs, Jessica V Bloom, Thomas Fischbacher, Konstantin Gromov, Urs Köster, Rishiraj Pravahan, James Manyika, et al. 2025. Towards a future space-based, highly scalable AI infrastructure system design.arXiv preprint arXiv:2511.19468(2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Niladrish Chatterjee, Mike O’Connor, Donghyuk Lee, Daniel R Johnson, Stephen W Keckler, Minsoo Rhu, and William J Dally. 2017. Architecting an energy-efficient dram system for gpus. In2017 IEEE International Symposium on High Performance Computer Architecture (HPCA). IEEE, 73–84

    2017

  6. [6]

    Pranav Singh Chib and Pravendra Singh. 2024. Recent Advancements in End-to- End Autonomous Driving Using Deep Learning: A Survey.IEEE Transactions on Intelligent Vehicles9, 1 (2024), 103–118. doi:10.1109/TIV.2023.3318070

    work page doi:10.1109/tiv.2023.3318070 2024

  7. [7]

    2015.COMSOL Multiphysics ® v5.1

    COMSOL AB. 2015.COMSOL Multiphysics ® v5.1. COMSOL AB, Stockholm, Sweden. https://www.comsol.com

    2015

  8. [8]

    Gholami, Z

    Amir Gholami, Zhewei Yao, Sehoon Kim, Coleman Hooper, Michael W. Mahoney, and Kurt Keutzer. 2024. AI and Memory Wall.IEEE Micro44, 3 (2024), 33–39. doi:10.1109/MM.2024.3373763

    work page doi:10.1109/mm.2024.3373763 2024

  9. [9]

    David G. Gilmore. 2002.Spacecraft Thermal Control Handbook, Volume I: Funda- mental Technologies(2 ed.). The Aerospace Press, El Segundo, CA, USA

    2002

  10. [10]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Ab- hishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schel- ten, Alex Vaughan, et al . 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783(2024)

    Pith/arXiv arXiv 2024

  11. [11]

    Joao Guerreiro, Aleksandar Ilic, Nuno Roma, and Pedro Tomas. 2018. GPGPU power modeling for multi-domain voltage-frequency scaling. In2018 IEEE Inter- national Symposium on High Performance Computer Architecture (HPCA). IEEE, 789–800

    2018

  12. [12]

    Phatak, Cheng Wang, and Supratik Guha

    Wilfried Haensch, Anand Raghunathan, Kaushik Roy, Bhaswar Chakrabarti, Charudatta M. Phatak, Cheng Wang, and Supratik Guha. 2023. Compute in- Memory with Non-Volatile Elements for Neural Networks: A Review from a Co-Design Perspective.Advanced Materials35, 37 (2023), 2204944. doi:10.1002/ adma.202204944

    2023

  13. [13]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition. 770–778

    2016

  14. [14]

    Jeremy Hsu. 2025. Data Centers in Space Aren’t as Wild as They Sound.Scien- tific American(9 Dec. 2025). https://www.scientificamerican.com/article/data- centers-in-space/ Online; accessed 2026-03-01

    2025

  15. [15]

    Naoko Iwata, Sogo Nakanoya, Nobuyuki Nakamura, Noboru Takeda, and Fumiya Tsutsui. 2022. Thermal performance evaluation of space radiator for single-phase mechanically pumped fluid loop.Journal of Spacecraft and Rockets59, 1 (2022), 225–235

    2022

  16. [16]

    Seungchul Jung, Hyungwoo Lee, Sungmeen Myung, Hyunsoo Kim, Seung Keun Yoon, Soon-Wan Kwon, Yongmin Ju, Minje Kim, Wooseok Yi, Shinhee Han, et al

  17. [17]

    A crossbar array of magnetoresistive memory devices for in-memory computing.Nature601, 7892 (2022), 211–216

    2022

  18. [18]

    Sitaraman, and Puneet Gupta

    George Karfakis, Myriam Bouzidi, Yunhyeok Im, Alexander Graening, Suresh K. Sitaraman, and Puneet Gupta. 2025. Optimizing Thermal Performance in 2.5D Systems Using Embedded Isolators.IEEE Journal on Emerging and Selected Topics in Circuits and Systems15, 3 (2025), 458–468. doi:10.1109/JETCAS.2025.3595909

    work page doi:10.1109/jetcas.2025.3595909 2025

  19. [19]

    Andrej Karpathy. 2022. nanoGPT: The simplest, fastest repository for train- ing/finetuning medium-sized GPTs. https://github.com/karpathy/nanoGPT. Ac- cessed: 2026-03-18

    2022

  20. [20]

    Riduan Khaddam-Aljameh, Milos Stanisavljevic, Jordi Fornt Mas, Geethan Karunaratne, Matthias Brändli, Feng Liu, Abhairaj Singh, Silvia M Müller, Urs Egger, Anastasios Petropoulos, et al. 2022. HERMES-core—A 1.59-TOPS/mm 2 PCM on 14-nm CMOS in-memory compute core using 300-ps/LSB linearized CCO-based ADCs.IEEE Journal of Solid-State Circuits57, 4 (2022), 1027–1038

    2022

  21. [21]

    Dong Eun Kim, Aayush Ankit, Cheng Wang, and Kaushik Roy. 2023. SAMBA: Sparsity Aware In-Memory Computing Based Machine Learning Accelerator. IEEE Trans. Comput.72, 9 (2023), 2615–2627. doi:10.1109/TC.2023.3257513

    work page doi:10.1109/tc.2023.3257513 2023

  22. [22]

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. 2015. Deep learning.nature 521, 7553 (2015), 436–444

    2015

  23. [23]

    Jingwen Leng, Isaac Hetherington, Ahmed ElTantawy, Syed Gilani, Wilson WL Fung, and Tor M Aamodt. 2013. GPUWattch: enabling energy optimizations in GPGPUs. InProceedings of the 40th Annual International Symposium on Computer Architecture. 487–498. doi:10.1145/2485922.2485964

    work page doi:10.1145/2485922.2485964 2013

  24. [24]

    Ku-Feng Lin, Hiroki Noguchi, Yi-Chun Shih, Perng-Fei Yuh, Yuan-Jen Lee, Tung- Cheng Chang, Sheng-Po Huang, Yu-Fan Lin, Chun-Ying Lee, Yen-Hsiang Huang, et al. 2024. 15.9 A 16nm 16Mb embedded STT-MRAM with a 20ns write time, a 10 12 write endurance and integrated margin-expansion schemes. In2024 IEEE International Solid-State Circuits Conference (ISSCC), V...

    2024

  25. [25]

    Yaoqi Liu, Yinhe Han, Hongxin Li, Shuhao Gu, Jibing Qiu, and Ting Li. 2025. Computing over Space: Status, Challenges, and Opportunities.Engineering54, 11 (2025), 20–25. doi:10.1016/j.eng.2025.06.005

    work page doi:10.1016/j.eng.2025.06.005 2025

  26. [26]

    Rui Lu and Dan Wang. 2025. A Thermal-aware Workload Scheduler for High- performance LLM Inference in Cooling-regulated Datacenters.ACM SIGENERGY Energy Informatics Review5, 2 (2025), 98–104

    2025

  27. [27]

    2024.How data centers and the energy sector can sate AI’s hunger for power

    McKinsey & Company. 2024.How data centers and the energy sector can sate AI’s hunger for power. Technical Report. McKinsey & Com- pany. https://www.mckinsey.com/industries/private-capital/our-insights/how- data-centers-and-the-energy-sector-can-sate-ais-hunger-for-power Data cen- ter load may make up between 30 and 40 percent of all net new demand added u...

    2024

  28. [28]

    Mohd Halim Mohd Noor and Ayokunle Olalekan Ige. 2025. A survey on state-of- the-art deep learning applications and challenges.Engineering Applications of Artificial Intelligence159 (2025), 111225

    2025

  29. [29]

    NVIDIA Corporation. 2020. NVIDIA A100 Tensor Core GPU Ar- chitecture. https://resources.nvidia.com/en-us-tensor-core/nvidia-ampere- architecture-whitepaper. Accessed: 2026-03-06

    2020

  30. [30]

    Mike O’Connor, Niladrish Chatterjee, and Others. 2017. Fine-grained DRAM: Energy-efficient DRAM for extreme bandwidth systems. InProceedings of the 50th Annual IEEE/ACM MICRO. 41–54

    2017

  31. [31]

    Anichur Rahman, Tanoy Debnath, Dipanjali Kundu, Md Saikat Islam Khan, Airin Afroj Aishi, Sadia Sazzad, Mohammad Sayduzzaman, and Shahab S Band

  32. [32]

    Machine learning and deep learning-based approach in smart healthcare: Recent advances, applications, challenges and opportunities.AIMS public health 11, 1 (2024), 58

    2024

  33. [33]

    Bharath Ramakrishnan, Cam Turner, Husam Alissa, Dennis Trieu, Felipe Rivera, Luke Melton, Muralikrishna Rao, Sruti Chigullapalli, Tatek Getachew, Vladimir Prodanovic, et al. 2025. Understanding the impact of data center liquid cooling on energy and performance of machine learning and artificial intelligence workloads. Journal of Electronic Packaging147, 2...

    2025

  34. [34]

    Stanley Williams, and Vivek Srikumar

    Ali Shafiee, Anirban Nag, Naveen Muralimanohar, Rajeev Balasubramonian, John Paul Strachan, Miao Hu, R. Stanley Williams, and Vivek Srikumar. 2016. ISAAC: a convolutional neural network accelerator with in-situ analog arithmetic in crossbars. InProceedings of the 43rd International Symposium on Computer Architecture(Seoul, Republic of Korea)(ISCA ’16). IE...

    2016

  35. [35]

    Pia Singh. 2025. Nvidia-backed Starcloud trains first AI model in space, orbital data centers. CNBC News Article. https://www.cnbc.com/2025/12/10/nvidia- backed-starcloud-trains-first-ai-model-in-space-orbital-data-centers.html Ac- cessed 2026-03-01

    2025

  36. [36]

    Keeyoung Son, Joonsang Park, Seongguk Kim, Boogyo Sim, Keunwoo Kim, Seonguk Choi, Hyunsik Kim, and Joungho Kim. 2023. Thermal Analysis of High Bandwidth Memory (HBM)-GPU Module considering Power Consumption. In 2023 IEEE Electrical Design of Advanced Packaging and Systems (EDAPS). 1–3. doi:10.1109/EDAPS58880.2023.10468315

    work page doi:10.1109/edaps58880.2023.10468315 2023

  37. [37]

    Zhenheng Tang, Yuxin Wang, Qiang Wang, and Xiaowen Chu. 2019. The impact of GPU DVFS on the energy and performance of deep learning: An empirical study. InProceedings of the Tenth ACM International Conference on Future Energy Systems. 315–325

    2019

  38. [38]

    Weier Wan, Rajkumar Kubendran, Clemens Schaefer, Sukru Burc Eryilmaz, Wenqiang Zhang, Dabin Wu, Stephen Deiss, Priyanka Raina, He Qian, Bin Gao, et al. 2022. A compute-in-memory chip based on resistive random-access memory. Nature608, 7923 (2022), 504–512

    2022

  39. [39]

    Dan Zhao, Siddharth Samsi, Joseph McDonald, Baolin Li, David Bestor, Michael Jones, Devesh Tiwari, and Vijay Gadepally. 2023. Sustainable supercomputing for AI: GPU power capping at HPC scale. InProceedings of the 2023 ACM symposium on cloud computing. 588–596

    2023