Source-linked AI summary
Exposing the Cost of Deep Learning Audio Development
Constance Douwes, Paul Magron, Romain Serizel
TL;DR
The full life-cycle environmental impact of deep learning remains understudied, particularly the energy-intensive development phase. Using Grid5000 activity logs, the study estimates energy costs for four audio projects and finds development energy substantially exceeds training the best-performing model alone.
Problem
Full life-cycle energy and carbon assessment of deep-learning systems remains understudied, despite development involving intensive iterative experimentation.
Source: §1 · p. 1Method
The authors estimate development energy from Grid5000 job logs and apply the methodology at laboratory, team, and project scales.
Source: §1 · p. 1, §2.1 · p. 2Results
3 to 256 times greater energy was consumed during development than to train the best-performing model across the four projects.
Source: §3.2 · p. 3Takeaways & Limitations
The findings support monitoring energy consumption across the full life cycle of deep-learning projects to foster more sustainable practices.
Source: §4 · p. 4Retrospective estimates rely on approximations, and carbon estimates include operational electricity emissions but exclude embodied hardware and other environmental impacts.
Source: §3.3 · p. 4
Abstract
from arXiv · showhide
The environmental impact of deep learning has attracted increasing attention over the past decade. Existing studies mainly focus on the energy and carbon emissions of model training and inference, while the whole development phase is often overlooked. Yet, architecture prototyping and intensive experiments are conducted during this stage, which is highly energy-demanding. In this article, we propose a methodology to estimate these costs, based on activity logs from the Grid5000 shared computing platform used by the LORIA laboratory. As a case-study, we focus on audio projects developed in the Multispeech research team. We evaluate the overall energy cost of four projects, and we compare them to those of training the reported models. Our results show that the energy required for the development phase is 3 to 256 times greater than that required to train the best-performing model alone. These results advocate for a more systematic reporting of energy consumption across the entire life cycle of deep learning-based audio projects.
1. INTRODUCTION
The environmental costs of deep learning are increasingly measured, but full life-cycle assessment remains understudied and audio development costs are particularly unexamined. This paper estimates development energy at several scales using Grid5000 job logs and studies audio projects at LORIA.
1. INTRODUCTION
- Full life-cycle environmental impact assessment remains understudied, with existing approaches differing in which stages and energy-estimation methods they cover.Source: §1 · p. 1
- Development spans iterative prototyping, trial-and-error, debugging, evaluation, hyperparameter searches, and ablations, often over months or years.Source: §1 · p. 1
- Prior estimates of development costs vary widely, while analyses of this phase in audio and speech processing are still missing.Two NLP models had development costs estimated at 2,000 and 3,000 times final training; BLOOM and OPT were each estimated at about twice their training cost.Source: §1 · p. 1
- The study estimates development energy at laboratory, team, and project scales using shared-platform job logs, then compares four audio projects’ costs with model training.Source: §1 · p. 1
2. METHODOLOGY
The methodology estimates development energy at multiple scales from computing-platform activity, using complementary energy-estimation approaches and project-level job attribution. It compares four projects’ development costs with training and estimates associated carbon emissions.
2. METHODOLOGY
- The analysis estimates development energy at multiple scales, from laboratory to individual project, using Grid5000 activity records.Source: §2 · p. 2, §2.1 · p. 2
2.1. Energy consumption
- Two complementary job-level methods estimate energy: static hardware specifications and duration, or dynamic node power measurements allocated among concurrent jobs.The static method uses CPU/GPU thermal design power and a fixed memory contribution; BMC measurements are divided according to job GPU allocation.Source: §2.1 · p. 2
2.2. Laboratory and team scale
- Laboratory energy trends use core-hour data from 2014–2025, while team-level GPU consumption uses BMC measurements over September 2023–December 2024.Source: §2.2 · p. 2
2.3.1. Project description
- The four selected projects cover multichannel speech recognition, speech enhancement, audio-visual speech enhancement, and music source separation, with 3–56 reported training runs each.The project descriptions report 16 runs for Cui et al., 6 for Ayilo et al., 3 for Ayilo et al. 2025, and 56 for Magron et al.Source: §2.3.1 · p. 2
2.3.2. From users to project
- Researchers annotated job metadata, especially launch directories and commands, to assign jobs to projects and sum their development energy.Source: §2.3.2 · pp. 2–3
2.3.3. Energy per training
- Training comparisons use energy for either all reported model runs or the best-performing model, estimated by retraining selected models for 2–5 epochs and extrapolating.Redundant retraining of experiments differing only in epoch count was avoided.Source: §2.3.3 · p. 3
2.4. Carbon emissions
- Project emissions are estimated using facility PUE and annual French electricity factors, with a hypothetical US electricity mix; conference-flight emissions are illustrative only.Source: §2.4 · p. 3
3. RESULTS
Across four audio projects, development consumed substantially more energy than training the best-performing model alone. The results also show growing energy use across Grid5000 Nancy clusters and substantial variation in project emissions and energy estimates.
3.1. Laboratory and team energy consumption
- Grid5000 Nancy energy consumption rose from 169 MWh in 2014 to 741 MWh in 2025, while GPUs accounted for more than half of estimated energy in the last four years.GPU activity was about 20% of allocated core-hours; higher GPU-server power contributed to the difference.Source: Figure 1 · p. 3, §3.1 · p. 3
- Multispeech averaged 260 kWh of daily GPU-server electricity use, or roughly 95 MWh yearly—almost one sixth of the Nancy cluster’s annual consumption.Consumption varied across days, with higher values during intensive development periods.Source: §3.1 · p. 3, Figure 2 · p. 4
3.2. Per-project energy consumption comparisons
- Development consumed 3 to 256 times the energy required to train the best-performing model, and 2 to 77 times that required for all reported training runs.Considering only reported-model energy substantially underestimates the whole project cost.Source: §3.2 · p. 3, Table 1 · p. 4
- Ayilo et al. ran more than 3,000 jobs consuming 9.5 MWh, while Magron et al. [25] used 24.2 MWh for replication, 53 times best-model training.The Cui et al. project’s development energy is underestimated because some experiments ran on Jean Zay and were not recorded here.Source: §3.2 · p. 4
- In France, development emissions were below travel emissions but not by much; under the hypothetical US mix, they exceeded travel emissions for most projects.The French-infrastructure estimates ranged from about half of travel emissions for Cui et al. and Magron et al. [25] to about a tenth for Ayilo et al..Source: §3.2 · p. 4
3.3. Discussion
- The retrospective estimates rely on approximations and count operational electricity emissions, not embodied hardware emissions or impacts such as water and resource use.The authors leave those broader environmental impacts to future methodology work.Source: §3.3 · p. 4
4. CONCLUSION
The methodology estimates development-phase energy at laboratory, team, and project levels. The analysis finds rising laboratory energy use and substantial costs for audio projects, with best-model training often representing only a fraction of total project energy.
4. CONCLUSION
- Best-reported-model training often represents only a fraction of the total energy cost of whole-project development.The conclusion encourages monitoring energy consumption throughout the project life cycle.Source: §4 · p. 4
- The methodology estimates development-phase energy at laboratory, research-team, and individual-project levels.Source: §4 · p. 4
- The analysis finds a large general energy increase at the laboratory level and a substantial energy toll for the audio-oriented Multispeech team.Source: §4 · p. 4
- The authors encourage monitoring energy consumption throughout the whole life cycle of DL projects.Source: §4 · p. 4