Latest News

On June 3rd and 4th, 2025, all SustainML partners convened in person at eProsima’s office in Tres Cantos (Madrid) for a two-day workshop.

 Attendees included representatives from each consortium member:

• eProsima: Jaime Martín Losa and Raúl Sánchez-Mateos Lizano
• DFKI: Paul Lukowicz, Bo Zhou, and Daniel Geissler
• Inria: Janin Koch
• RPTU (Paderborn University): Vladimir Rybalkin and Mohamed Moursi
• KU (University of Copenhagen): Bob Pepin
• IBM: Niccolò Avogaro

The primary aims of this face-to-face meeting were to exchange status updates from each work package team (WP1–WP7), foster closer collaboration among partners, and chart a clear roadmap for the remaining year of the project. 

Objectives

• Present current progress across all partners and work packages
• Identify key inter-WP collaborations and dependencies
• Define action items and milestones for next year of the project

June 3rd:

• WP 1: Energy-Oriented ML Task Modeling (DFKI)
  • Presenters: Bo Zhou & Daniel Geissler
    • Shared the latest “Knowledge2Model” RDF repository structure.
    • Reviewed progress on Tasks T1.2 (collaboration with Inria) and T1.4/T1.5 (collaboration with IBM & KU).
    • Agreed on integration steps for DFKI’s ML-problem knowledge graph into the centralized SustainML metadata service.

• WP 2: Hardware Architectures for Low-Power ML (RPTU)
  • Presenter: Vladimir Rybalkin (RPTU)
    • Presented new benchmark results on RPTU’s PIM‐based prototypes.
    • Discussed joint experiments to compare energy profiles between CPU, GPU, and UPMEM modules.
      Decided to share RPTU’s low-power measurement scripts with the CarbonTracker integration team.

• WP 3: Energy-Optimized ML Toolkit & Methods (KU & IBM)
  • Presenters: Bob Pepin (KU) & Niccolò Avogaro (IBM)
    • KU demonstrated the updated CarbonTracker configuration, including “predicted” vs. “actual” CO₂ estimates.
    • IBM explained how they fuse in-situ power logs (Fast DDS hooks) with model inference timing.
    • Partners agreed to standardize the logging format so that “live” energy/CO₂ data can feed into the SustainML toolkit.

• WP 4: Interaction & User Studies (Inria)
  • Presenter: Janin Koch (Inria)
    • Showed initial QtQuick GUI mockups and user scenarios.
    • Collected feedback on layout, terminology, and data‐visualization needs.
    • Identified key usability metrics (e.g., “time to submit a task” and “clarity of performance vs. energy graphs”).

• WP 5: Framework Architecture & Implementation (eProsima)
  • Presenter: Raúl Sánchez-Mateos (eProsima)
    • Walked through the Docker-based backend orchestrator and node deployment procedure.
    • Demonstrated how each “SustainML node” (WP1 → WP6 modules) can be started via docker-compose or the new bash launcher.
    • Reviewed CI/CD status, including the initial black-box & communication tests for C++ and Python modules.

• WP 6: Application-Oriented Validation (IBM)
  • Presenter: Niccolò Avogaro (IBM)
    • Ran through two demo scenarios:
    • Reiterating a non-final ML task (to compare “first pass” vs. “optimized pass” predictions).
    • Multi-output workflow (returning multiple candidate models+metrics from a single request).
    • Collected partner feedback on desired “next validation cases”—for example, comparing CPU vs. GPU latency+energy on the same model.

• WP 7: Dissemination, Exploitation & Management (eProsima + Inria + KU)
  • Leads: eProsima (Raúl Sánchez‐Mateos), Inria (Janin Koch), KU (Bob Pepin)
    • Reviewed the SustainML dashboard and identified outdated sections (e.g., missing “how to install SustainML” pages).
    • Agreed on upcoming blog posts, website announcements, and social-media outreach:
    • A “project overview” article for HumaneAI Net
    • A case study focused on “measuring real-time CO₂ in inference” for ADRA-e
    • A joint press release around our June 27 v0.1.0 backend + front-end launch
    • Tasked each partner with drafting at least two promotion assets (e.g., slides, short videos) by the end of July.

June 4th:

Representatives from eProsima, RPTU, DFKI, and IBM spent the entire day in a hands‐on technical workshop. Their goal was to deepen and accelerate the interconnections between each Work Package—refining the RDF knowledge graphs, improving backend communication and configuration request for the nodes, as well as, ML’s data-set incorporation as inputs on the application—to ensure that all components work seamlessly together as a cohesive SustainML framework.

MORE INFORMATION ABOUT SUSTAINML:

If you have any questions, please get in touch with This email address is being protected from spambots. You need JavaScript enabled to view it..